public inbox for [email protected]
help / color / mirror / Atom feedRe: POC: GROUP BY optimization
38+ messages / 10 participants
[nested] [flat]
* Re: POC: GROUP BY optimization
@ 2022-01-21 11:09 Andrey Lepikhov <[email protected]>
0 siblings, 1 reply; 38+ messages in thread
From: Andrey Lepikhov @ 2022-01-21 11:09 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; Teodor Sigaev <[email protected]>; +Cc: Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On 7/22/21 3:58 AM, Tomas Vondra wrote:
> 4) I'm not sure it's actually a good idea to pass tuplesPerPrevGroup to
> estimate_num_groups_incremental. In principle yes, if we use "group
> size" from the previous step, then the returned value is the number of
> new groups after adding the "new" pathkey.
> But even if we ignore the issues with amplification mentioned in (3),
> there's an issue with non-linear behavior in estimate_num_groups,
> because at some point it's calculating
>
> D(N,n,p) = n * (1 - ((N-p)/N)^(N/n))
>
> where N - total rows, p - sample size, n - number of distinct values.
> And if we have (N1,n1) and (N2,n2) then the ratio of calculated
> estimated (which is pretty much what calculating group size does)
>
> D(N2,n2,p2) / D(N1,n1,p1)
>
> which will differ depending on p1 and p2. And if we're tweaking the
> tuplesPerPrevGroup all the time, that's really annoying, as it may make
> the groups smaller or larger, which is unpredictable and annoying, and I
> wonder if it might go against the idea of penalizing tuplesPerPrevGroup
> to some extent.
> We could simply use the input "tuples" value here, and then divide the
> current and previous estimate to calculate the number of new groups.
tuplesPerPrevGroup is only a top boundary for the estimation. I think,
we should use result of previous estimation as a limit for the next
during incremental estimation. Maybe we simply limit the
tuplesPerPrevGroup value to ensure of monotonic non-growth of this
value? - see in attachment patch to previous fixes.
--
regards,
Andrey Lepikhov
Postgres Professional
diff --git a/src/backend/optimizer/path/costsize.c b/src/backend/optimizer/path/costsize.c
index 70af9c91d5..4e26cd275d 100644
--- a/src/backend/optimizer/path/costsize.c
+++ b/src/backend/optimizer/path/costsize.c
@@ -2025,8 +2025,10 @@ compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
/*
* Real-world distribution isn't uniform but now we don't have a way to
* determine that, so, add multiplier to get closer to worst case.
+ * Guard this value against irrational decisions.
*/
- tuplesPerPrevGroup = ceil(1.5 * tuplesPerPrevGroup / nGroups);
+ tuplesPerPrevGroup = Min(tuplesPerPrevGroup,
+ ceil(1.5 * tuplesPerPrevGroup / nGroups));
/*
* We could skip all following columns for cost estimation, because we
Attachments:
[text/plain] guard_tuplesPerPrevGroup.txt (775B, ../../[email protected]/2-guard_tuplesPerPrevGroup.txt)
download | inline diff:
diff --git a/src/backend/optimizer/path/costsize.c b/src/backend/optimizer/path/costsize.c
index 70af9c91d5..4e26cd275d 100644
--- a/src/backend/optimizer/path/costsize.c
+++ b/src/backend/optimizer/path/costsize.c
@@ -2025,8 +2025,10 @@ compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
/*
* Real-world distribution isn't uniform but now we don't have a way to
* determine that, so, add multiplier to get closer to worst case.
+ * Guard this value against irrational decisions.
*/
- tuplesPerPrevGroup = ceil(1.5 * tuplesPerPrevGroup / nGroups);
+ tuplesPerPrevGroup = Min(tuplesPerPrevGroup,
+ ceil(1.5 * tuplesPerPrevGroup / nGroups));
/*
* We could skip all following columns for cost estimation, because we
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-01-21 20:34 Tomas Vondra <[email protected]>
parent: Andrey Lepikhov <[email protected]>
0 siblings, 2 replies; 38+ messages in thread
From: Tomas Vondra @ 2022-01-21 20:34 UTC (permalink / raw)
To: Andrey Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; +Cc: Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On 1/21/22 12:09, Andrey Lepikhov wrote:
> On 7/22/21 3:58 AM, Tomas Vondra wrote:
>> 4) I'm not sure it's actually a good idea to pass tuplesPerPrevGroup to
>> estimate_num_groups_incremental. In principle yes, if we use "group
>> size" from the previous step, then the returned value is the number of
>> new groups after adding the "new" pathkey.
>> But even if we ignore the issues with amplification mentioned in (3),
>> there's an issue with non-linear behavior in estimate_num_groups,
>> because at some point it's calculating
>>
>> D(N,n,p) = n * (1 - ((N-p)/N)^(N/n))
>>
>> where N - total rows, p - sample size, n - number of distinct values.
>> And if we have (N1,n1) and (N2,n2) then the ratio of calculated
>> estimated (which is pretty much what calculating group size does)
>>
>> D(N2,n2,p2) / D(N1,n1,p1)
>>
>> which will differ depending on p1 and p2. And if we're tweaking the
>> tuplesPerPrevGroup all the time, that's really annoying, as it may make
>> the groups smaller or larger, which is unpredictable and annoying, and I
>> wonder if it might go against the idea of penalizing tuplesPerPrevGroup
>> to some extent.
>> We could simply use the input "tuples" value here, and then divide the
>> current and previous estimate to calculate the number of new groups.
>
> tuplesPerPrevGroup is only a top boundary for the estimation. I think,
> we should use result of previous estimation as a limit for the next
> during incremental estimation. Maybe we simply limit the
> tuplesPerPrevGroup value to ensure of monotonic non-growth of this
> value? - see in attachment patch to previous fixes.
>
Yes, I think that's a reasonable defense - it makes no sense to exceed
the group size in the preceding step, which could happen with small
number of groups or some unexpected ndistinct estimate.
The other thing we could do is reduce the coefficient gradually - so
it'd be 1.5 for the first pathkey, then 1.25 for the next one, and so
on. But it seems somewhat arbitrary (I certainly don't have some sound
theoretical justification ...).
I've merged most of the fixes you reported. I've skipped the path_save
removal in planner.c, because that seems incorrect - if there are
multiple pathkeys, we must start with the original path, not the
modified one we built in the last iteration. Or am I missing something?
regards
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
Attachments:
[text/x-patch] v20220121-0001-GROUP-BY-reordering.patch (99.2K, ../../[email protected]/2-v20220121-0001-GROUP-BY-reordering.patch)
download | inline diff:
From 590589f4388499d41d2c1bb42678b794c044d045 Mon Sep 17 00:00:00 2001
From: Tomas Vondra <[email protected]>
Date: Fri, 21 Jan 2022 20:22:14 +0100
Subject: [PATCH v20220121] GROUP BY reordering
---
.../postgres_fdw/expected/postgres_fdw.out | 15 +-
src/backend/optimizer/path/costsize.c | 369 +++++++++-
src/backend/optimizer/path/equivclass.c | 13 +-
src/backend/optimizer/path/pathkeys.c | 580 ++++++++++++++++
src/backend/optimizer/plan/planner.c | 653 ++++++++++--------
src/backend/optimizer/util/pathnode.c | 2 +-
src/backend/utils/adt/selfuncs.c | 37 +-
src/backend/utils/misc/guc.c | 29 +
src/include/nodes/nodes.h | 1 +
src/include/nodes/pathnodes.h | 10 +
src/include/optimizer/cost.h | 4 +-
src/include/optimizer/paths.h | 11 +
src/include/utils/selfuncs.h | 5 +
src/test/regress/expected/aggregates.out | 244 ++++++-
.../regress/expected/incremental_sort.out | 2 +-
src/test/regress/expected/join.out | 51 +-
.../regress/expected/partition_aggregate.out | 136 ++--
src/test/regress/expected/partition_join.out | 75 +-
src/test/regress/expected/union.out | 60 +-
src/test/regress/sql/aggregates.sql | 99 +++
src/test/regress/sql/incremental_sort.sql | 2 +-
21 files changed, 1904 insertions(+), 494 deletions(-)
diff --git a/contrib/postgres_fdw/expected/postgres_fdw.out b/contrib/postgres_fdw/expected/postgres_fdw.out
index 7d6f7d9e3df..9926b6bba07 100644
--- a/contrib/postgres_fdw/expected/postgres_fdw.out
+++ b/contrib/postgres_fdw/expected/postgres_fdw.out
@@ -2741,16 +2741,13 @@ select c2 * (random() <= 1)::int as c2 from ft2 group by c2 * (random() <= 1)::i
-- GROUP BY clause in various forms, cardinal, alias and constant expression
explain (verbose, costs off)
select count(c2) w, c2 x, 5 y, 7.0 z from ft1 group by 2, y, 9.0::int order by 2;
- QUERY PLAN
----------------------------------------------------------------------------------------
- Sort
+ QUERY PLAN
+------------------------------------------------------------------------------------------------------------
+ Foreign Scan
Output: (count(c2)), c2, 5, 7.0, 9
- Sort Key: ft1.c2
- -> Foreign Scan
- Output: (count(c2)), c2, 5, 7.0, 9
- Relations: Aggregate on (public.ft1)
- Remote SQL: SELECT count(c2), c2, 5, 7.0, 9 FROM "S 1"."T 1" GROUP BY 2, 3, 5
-(7 rows)
+ Relations: Aggregate on (public.ft1)
+ Remote SQL: SELECT count(c2), c2, 5, 7.0, 9 FROM "S 1"."T 1" GROUP BY 2, 3, 5 ORDER BY c2 ASC NULLS LAST
+(4 rows)
select count(c2) w, c2 x, 5 y, 7.0 z from ft1 group by 2, y, 9.0::int order by 2;
w | x | y | z
diff --git a/src/backend/optimizer/path/costsize.c b/src/backend/optimizer/path/costsize.c
index 8dc7dd4ca26..68a32740d7e 100644
--- a/src/backend/optimizer/path/costsize.c
+++ b/src/backend/optimizer/path/costsize.c
@@ -1754,6 +1754,325 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
rterm->pathtarget->width);
}
+/*
+ * is_fake_var
+ * Workaround for generate_append_tlist() which generates fake Vars with
+ * varno == 0, that will cause a fail of estimate_num_group() call
+ *
+ * XXX Ummm, why would estimate_num_group fail with this?
+ */
+static bool
+is_fake_var(Expr *expr)
+{
+ if (IsA(expr, RelabelType))
+ expr = (Expr *) ((RelabelType *) expr)->arg;
+
+ return (IsA(expr, Var) && ((Var *) expr)->varno == 0);
+}
+
+/*
+ * get_width_cost_multiplier
+ * Returns relative complexity of comparing two values based on it's width.
+ * The idea behind - long values have more expensive comparison. Return value is
+ * in cpu_operator_cost unit.
+ */
+static double
+get_width_cost_multiplier(PlannerInfo *root, Expr *expr)
+{
+ double width = -1.0; /* fake value */
+
+ if (IsA(expr, RelabelType))
+ expr = (Expr *) ((RelabelType *) expr)->arg;
+
+ /* Try to find actual stat in corresponding relation */
+ if (IsA(expr, Var))
+ {
+ Var *var = (Var *) expr;
+
+ if (var->varno > 0 && var->varno < root->simple_rel_array_size)
+ {
+ RelOptInfo *rel = root->simple_rel_array[var->varno];
+
+ if (rel != NULL &&
+ var->varattno >= rel->min_attr &&
+ var->varattno <= rel->max_attr)
+ {
+ int ndx = var->varattno - rel->min_attr;
+
+ if (rel->attr_widths[ndx] > 0)
+ width = rel->attr_widths[ndx];
+ }
+ }
+ }
+
+ /* Didn't find any actual stats, use estimation by type */
+ if (width < 0.0)
+ {
+ Node *node = (Node*) expr;
+
+ width = get_typavgwidth(exprType(node), exprTypmod(node));
+ }
+
+ /*
+ * Any value in pgsql is passed by Datum type, so any operation with value
+ * could not be cheaper than operation with Datum type
+ */
+ if (width <= sizeof(Datum))
+ return 1.0;
+
+ /*
+ * Seems, cost of comparision is not directly proportional to args width,
+ * because comparing args could be differ width (we known only average over
+ * column) and difference often could be defined only by looking on first
+ * bytes. So, use log16(width) as estimation.
+ */
+ return 1.0 + 0.125 * LOG2(width / sizeof(Datum));
+}
+
+/*
+ * compute_cpu_sort_cost
+ * compute CPU cost of sort (i.e. in-memory)
+ *
+ * The main thing we need to calculate to estimate sort CPU costs is the number
+ * of calls to the comparator functions. The difficulty is that for multi-column
+ * sorts there may be different data types involved (for some of which the calls
+ * may be much more expensive). Furthermore, the columns may have very different
+ * number of distinct values - the higher the number, the fewer comparisons will
+ * be needed for the following columns.
+ *
+ * The algoritm is incremental - we add pathkeys one by one, and at each step we
+ * estimate the number of necessary comparisons (based on the number of distinct
+ * groups in the current pathkey prefix and the new pathkey), and the comparison
+ * costs (which is data type specific).
+ *
+ * Estimation of the number of comparisons is based on ideas from:
+ *
+ * "Quicksort Is Optimal", Robert Sedgewick, Jon Bentley, 2002
+ * [https://www.cs.princeton.edu/~rs/talks/QuicksortIsOptimal.pdf]
+ *
+ * In term of that paper, let N - number of tuples, Xi - number of identical
+ * tuples with value Ki, then the estimate of number of comparisons is:
+ *
+ * log(N! / (X1! * X2! * ..)) ~ sum(Xi * log(N/Xi))
+ *
+ * We assume all Xi the same because now we don't have any estimation of
+ * group sizes, we have only know the estimate of number of groups (distinct
+ * values). In that case, formula becomes:
+ *
+ * N * log(NumberOfGroups)
+ *
+ * For multi-column sorts we need to estimate the number of comparisons for
+ * each individual column - for example with columns (c1, c2, ..., ck) we
+ * can estimate that number of comparisons on ck is roughly
+ *
+ * ncomparisons(c1, c2, ..., ck) / ncomparisons(c1, c2, ..., c(k-1))
+ *
+ * Let k be a column number, Gk - number of groups defined by k columns, and Fk
+ * the cost of the comparison is
+ *
+ * N * sum( Fk * log(Gk) )
+ *
+ * Note: We also consider column width, not just the comparator cost.
+ *
+ * NOTE: some callers currently pass NIL for pathkeys because they
+ * can't conveniently supply the sort keys. In this case, it will fallback to
+ * simple comparison cost estimate.
+ */
+static Cost
+compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
+ Cost comparison_cost, double tuples, double output_tuples,
+ bool heapSort)
+{
+ Cost per_tuple_cost = 0.0;
+ ListCell *lc;
+ List *pathkeyExprs = NIL;
+ double tuplesPerPrevGroup = tuples;
+ double totalFuncCost = 1.0;
+ bool has_fake_var = false;
+ int i = 0;
+ Oid prev_datatype = InvalidOid;
+ Cost funcCost = 0.0;
+ List *cache_varinfos = NIL;
+
+ /* fallback if pathkeys is unknown */
+ if (list_length(pathkeys) == 0)
+ {
+ /*
+ * If we'll use a bounded heap-sort keeping just K tuples in memory, for
+ * a total number of tuple comparisons of N log2 K; but the constant
+ * factor is a bit higher than for quicksort. Tweak it so that the
+ * cost curve is continuous at the crossover point.
+ *
+ * XXX I suppose the "quicksort factor" references to 1.5 at the end
+ * of this function, but I'm not sure. I suggest we introduce some simple
+ * constants for that, instead of magic values.
+ */
+ output_tuples = (heapSort) ? 2.0 * output_tuples : tuples;
+ per_tuple_cost += 2.0 * cpu_operator_cost * LOG2(output_tuples);
+
+ /* add cost provided by caller */
+ per_tuple_cost += comparison_cost;
+
+ return per_tuple_cost * tuples;
+ }
+
+ /*
+ * Computing total cost of sorting takes into account:
+ * - per column comparison function cost
+ * - we try to compute needed number of comparison per column
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey*) lfirst(lc);
+ EquivalenceMember *em;
+ double nGroups,
+ correctedNGroups;
+
+ /*
+ * We believe that equivalence members aren't very different, so, to
+ * estimate cost we take just first member
+ */
+ em = (EquivalenceMember *) linitial(pathkey->pk_eclass->ec_members);
+
+ if (em->em_datatype != InvalidOid)
+ {
+ /* do not lookup funcCost if data type is the same as previous */
+ if (prev_datatype != em->em_datatype)
+ {
+ Oid sortop;
+ QualCost cost;
+
+ sortop = get_opfamily_member(pathkey->pk_opfamily,
+ em->em_datatype, em->em_datatype,
+ pathkey->pk_strategy);
+
+ cost.startup = 0;
+ cost.per_tuple = 0;
+ add_function_cost(root, get_opcode(sortop), NULL, &cost);
+ /* we need procost, not product of procost and cpu_operator_cost */
+ funcCost = cost.per_tuple / cpu_operator_cost;
+ prev_datatype = em->em_datatype;
+ }
+ }
+ else
+ funcCost = 1.0; /* fallback */
+
+ /* Try to take into account actual width fee */
+ funcCost *= get_width_cost_multiplier(root, em->em_expr);
+
+ totalFuncCost += funcCost;
+
+ /* Remember if we have a fake var in pathkeys */
+ has_fake_var |= is_fake_var(em->em_expr);
+ pathkeyExprs = lappend(pathkeyExprs, em->em_expr);
+
+ /*
+ * Prevent call estimate_num_groups() with fake Var. Note,
+ * pathkeyExprs contains only previous columns
+ */
+ if (has_fake_var == false)
+ /*
+ * Recursively compute number of groups in a group from previous step
+ */
+ nGroups = estimate_num_groups_incremental(root, pathkeyExprs,
+ tuplesPerPrevGroup, NULL, NULL,
+ &cache_varinfos,
+ list_length(pathkeyExprs) - 1);
+ else if (tuples > 4.0)
+ /*
+ * Use geometric mean as estimation if there is no any stats.
+ * Don't use DEFAULT_NUM_DISTINCT because it used for only one
+ * column while here we try to estimate number of groups over
+ * set of columns.
+ *
+ * XXX Perhaps this should use DEFAULT_NUM_DISTINCT at least to
+ * limit the calculated values, somehow?
+ *
+ * XXX What's the logic of the following formula?
+ */
+ nGroups = ceil(2.0 + sqrt(tuples) * (i + 1) / list_length(pathkeys));
+ else
+ nGroups = tuples;
+
+ /*
+ * Presorted keys aren't participated in comparison but still checked
+ * by qsort comparator.
+ */
+ if (i >= nPresortedKeys)
+ {
+ if (heapSort)
+ {
+ double heap_tuples;
+
+ /* have to keep at least one group, and a multiple of group size */
+ heap_tuples = Max(ceil(output_tuples / tuplesPerPrevGroup) * tuplesPerPrevGroup,
+ tuplesPerPrevGroup);
+
+ /* so how many (whole) groups is that? */
+ correctedNGroups = ceil(heap_tuples / tuplesPerPrevGroup);
+ }
+ else
+ /* all groups in the input */
+ correctedNGroups = nGroups;
+
+ correctedNGroups = Max(1.0, ceil(correctedNGroups));
+
+ per_tuple_cost += totalFuncCost * LOG2(correctedNGroups);
+ }
+
+ i++;
+
+ /*
+ * Real-world distribution isn't uniform but now we don't have a way to
+ * determine that, so, add multiplier to get closer to worst case.
+ * But ensure the number of tuples does not exceed the group size in the
+ * preceding step.
+ */
+ tuplesPerPrevGroup = Min(tuplesPerPrevGroup,
+ ceil(1.5 * tuplesPerPrevGroup / nGroups));
+
+ /*
+ * We could skip all following columns for cost estimation, because we
+ * believe that tuples are unique by the set of previous columns
+ */
+ if (tuplesPerPrevGroup <= 1.0)
+ break;
+ }
+
+ list_free(pathkeyExprs);
+
+ /* per_tuple_cost is in cpu_operator_cost units */
+ per_tuple_cost *= cpu_operator_cost;
+
+ /*
+ * Accordingly to "Introduction to algorithms", Thomas H. Cormen, Charles E.
+ * Leiserson, Ronald L. Rivest, ISBN 0-07-013143-0, quicksort estimation
+ * formula has additional term proportional to number of tuples (See Chapter
+ * 8.2 and Theorem 4.1). It has meaning with low number of tuples,
+ * approximately less that 1e4. Of course, it could be implemented as
+ * additional multiplier under logarithm, but use more complicated formula
+ * which takes into account number of unique tuples and it isn't clear how
+ * to combine multiplier with groups. Estimate it as 10 in cpu_operator_cost
+ * unit.
+ */
+ per_tuple_cost += 10 * cpu_operator_cost;
+
+ per_tuple_cost += comparison_cost;
+
+ return tuples * per_tuple_cost;
+}
+
+/*
+ * simple wrapper just to estimate best sort path
+ */
+Cost
+cost_sort_estimate(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
+ double tuples)
+{
+ return compute_cpu_sort_cost(root, pathkeys, nPresortedKeys,
+ 0, tuples, tuples, false);
+}
+
/*
* cost_tuplesort
* Determines and returns the cost of sorting a relation using tuplesort,
@@ -1770,7 +2089,7 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
* number of initial runs formed and M is the merge order used by tuplesort.c.
* Since the average initial run should be about sort_mem, we have
* disk traffic = 2 * relsize * ceil(logM(p / sort_mem))
- * cpu = comparison_cost * t * log2(t)
+ * and cpu cost (computed by compute_cpu_sort_cost()).
*
* If the sort is bounded (i.e., only the first k result tuples are needed)
* and k tuples can fit into sort_mem, we use a heap method that keeps only
@@ -1789,9 +2108,11 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
* 'comparison_cost' is the extra cost per comparison, if any
* 'sort_mem' is the number of kilobytes of work memory allowed for the sort
* 'limit_tuples' is the bound on the number of output tuples; -1 if no bound
+ * 'startup_cost' is expected to be 0 at input. If there is "input cost" it should
+ * be added by caller later.
*/
static void
-cost_tuplesort(Cost *startup_cost, Cost *run_cost,
+cost_tuplesort(PlannerInfo *root, List *pathkeys, Cost *startup_cost, Cost *run_cost,
double tuples, int width,
Cost comparison_cost, int sort_mem,
double limit_tuples)
@@ -1808,9 +2129,6 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
if (tuples < 2.0)
tuples = 2.0;
- /* Include the default cost-per-comparison */
- comparison_cost += 2.0 * cpu_operator_cost;
-
/* Do we have a useful LIMIT? */
if (limit_tuples > 0 && limit_tuples < tuples)
{
@@ -1834,12 +2152,10 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
double log_runs;
double npageaccesses;
- /*
- * CPU costs
- *
- * Assume about N log2 N comparisons
- */
- *startup_cost = comparison_cost * tuples * LOG2(tuples);
+ /* CPU costs */
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ tuples, false);
/* Disk costs */
@@ -1855,18 +2171,17 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
}
else if (tuples > 2 * output_tuples || input_bytes > sort_mem_bytes)
{
- /*
- * We'll use a bounded heap-sort keeping just K tuples in memory, for
- * a total number of tuple comparisons of N log2 K; but the constant
- * factor is a bit higher than for quicksort. Tweak it so that the
- * cost curve is continuous at the crossover point.
- */
- *startup_cost = comparison_cost * tuples * LOG2(2.0 * output_tuples);
+ /* We'll use a bounded heap-sort keeping just K tuples in memory. */
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ output_tuples, true);
}
else
{
/* We'll use plain quicksort on all the input tuples */
- *startup_cost = comparison_cost * tuples * LOG2(tuples);
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ tuples, false);
}
/*
@@ -1899,8 +2214,8 @@ cost_incremental_sort(Path *path,
double input_tuples, int width, Cost comparison_cost, int sort_mem,
double limit_tuples)
{
- Cost startup_cost = 0,
- run_cost = 0,
+ Cost startup_cost,
+ run_cost,
input_run_cost = input_total_cost - input_startup_cost;
double group_tuples,
input_groups;
@@ -1985,7 +2300,7 @@ cost_incremental_sort(Path *path,
* pessimistic about incremental sort performance and increase its average
* group size by half.
*/
- cost_tuplesort(&group_startup_cost, &group_run_cost,
+ cost_tuplesort(root, pathkeys, &group_startup_cost, &group_run_cost,
1.5 * group_tuples, width, comparison_cost, sort_mem,
limit_tuples);
@@ -1993,7 +2308,7 @@ cost_incremental_sort(Path *path,
* Startup cost of incremental sort is the startup cost of its first group
* plus the cost of its input.
*/
- startup_cost += group_startup_cost
+ startup_cost = group_startup_cost
+ input_startup_cost + group_input_run_cost;
/*
@@ -2002,7 +2317,7 @@ cost_incremental_sort(Path *path,
* group, plus the total cost to process the remaining groups, plus the
* remaining cost of input.
*/
- run_cost += group_run_cost
+ run_cost = group_run_cost
+ (group_run_cost + group_startup_cost) * (input_groups - 1)
+ group_input_run_cost * (input_groups - 1);
@@ -2042,7 +2357,7 @@ cost_sort(Path *path, PlannerInfo *root,
Cost startup_cost;
Cost run_cost;
- cost_tuplesort(&startup_cost, &run_cost,
+ cost_tuplesort(root, pathkeys, &startup_cost, &run_cost,
tuples, width,
comparison_cost, sort_mem,
limit_tuples);
@@ -2140,7 +2455,7 @@ append_nonpartial_cost(List *subpaths, int numpaths, int parallel_workers)
* Determines and returns the cost of an Append node.
*/
void
-cost_append(AppendPath *apath)
+cost_append(AppendPath *apath, PlannerInfo *root)
{
ListCell *l;
@@ -2208,7 +2523,7 @@ cost_append(AppendPath *apath)
* any child.
*/
cost_sort(&sort_path,
- NULL, /* doesn't currently need root */
+ root,
pathkeys,
subpath->total_cost,
subpath->rows,
diff --git a/src/backend/optimizer/path/equivclass.c b/src/backend/optimizer/path/equivclass.c
index 8c6770de972..5487ae2ee4c 100644
--- a/src/backend/optimizer/path/equivclass.c
+++ b/src/backend/optimizer/path/equivclass.c
@@ -681,7 +681,18 @@ get_eclass_for_sort_expr(PlannerInfo *root,
if (opcintype == cur_em->em_datatype &&
equal(expr, cur_em->em_expr))
- return cur_ec; /* Match! */
+ {
+ /*
+ * Match!
+ *
+ * Copy sortref if it wasn't set yet, it's possible if ec was
+ * constructed from WHERE clause, ie it doesn't have target
+ * reference at all
+ */
+ if (cur_ec->ec_sortref == 0 && sortref > 0)
+ cur_ec->ec_sortref = sortref;
+ return cur_ec;
+ }
}
}
diff --git a/src/backend/optimizer/path/pathkeys.c b/src/backend/optimizer/path/pathkeys.c
index 86a35cdef17..f18d4c8e4e7 100644
--- a/src/backend/optimizer/path/pathkeys.c
+++ b/src/backend/optimizer/path/pathkeys.c
@@ -17,16 +17,19 @@
*/
#include "postgres.h"
+#include "miscadmin.h"
#include "access/stratnum.h"
#include "catalog/pg_opfamily.h"
#include "nodes/makefuncs.h"
#include "nodes/nodeFuncs.h"
#include "nodes/plannodes.h"
+#include "optimizer/cost.h"
#include "optimizer/optimizer.h"
#include "optimizer/pathnode.h"
#include "optimizer/paths.h"
#include "partitioning/partbounds.h"
#include "utils/lsyscache.h"
+#include "utils/selfuncs.h"
static bool pathkey_is_redundant(PathKey *new_pathkey, List *pathkeys);
@@ -334,6 +337,530 @@ pathkeys_contained_in(List *keys1, List *keys2)
return false;
}
+/************************<DEBUG PART>*************************************/
+bool debug_group_by_reorder_by_pathkeys = true;
+bool debug_group_by_match_order_by = true;
+bool debug_cheapest_group_by = true;
+/************************</DEBUG PART>************************************/
+
+/*
+ * group_keys_reorder_by_pathkeys
+ * Reorder GROUP BY keys to match pathkeys of input path.
+ *
+ * Function returns new lists (pathkeys and clauses), original GROUP BY lists
+ * stay untouched.
+ *
+ * Returns the number of GROUP BY keys with a matching pathkey.
+ */
+int
+group_keys_reorder_by_pathkeys(List *pathkeys, List **group_pathkeys,
+ List **group_clauses)
+{
+ List *new_group_pathkeys= NIL,
+ *new_group_clauses = NIL;
+ ListCell *lc;
+ int n;
+
+ if (debug_group_by_reorder_by_pathkeys == false)
+ return 0;
+
+ if (pathkeys == NIL || *group_pathkeys == NIL)
+ return 0;
+
+ /*
+ * Walk the pathkeys (determining ordering of the input path) and see if
+ * there's a matching GROUP BY key. If we find one, we append if to the
+ * list, and do the same for the clauses.
+ *
+ * Once we find first pathkey without a matching GROUP BY key, the rest of
+ * the pathkeys is useless and can't be used to evaluate the grouping, so
+ * we abort the loop and ignore the remaining pathkeys.
+ *
+ * XXX Pathkeys are built in a way to allow simply comparing pointers.
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(lc);
+ SortGroupClause *sgc;
+
+ /* abort on first mismatch */
+ if (!list_member_ptr(*group_pathkeys, pathkey))
+ break;
+
+ new_group_pathkeys = lappend(new_group_pathkeys, pathkey);
+
+ sgc = get_sortgroupref_clause(pathkey->pk_eclass->ec_sortref,
+ *group_clauses);
+
+ new_group_clauses = lappend(new_group_clauses, sgc);
+ }
+
+ /* remember the number of pathkeys with a matching GROUP BY key */
+ n = list_length(new_group_pathkeys);
+
+ /* XXX maybe return when (n == 0) */
+
+ /* append the remaining group pathkeys (will be treated as not sorted) */
+ *group_pathkeys = list_concat_unique_ptr(new_group_pathkeys,
+ *group_pathkeys);
+ *group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+
+ return n;
+}
+
+/*
+ * Used to generate all permutations of a pathkey list.
+ */
+typedef struct PathkeyMutatorState {
+ List *elemsList;
+ ListCell **elemCells;
+ void **elems;
+ int *positions;
+ int mutatorNColumns;
+ int count;
+} PathkeyMutatorState;
+
+
+/*
+ * PathkeyMutatorInit
+ * Initialize state of the permutation generator.
+ *
+ * We want to generate permutations of elements in the "elems" list. We may want
+ * to skip some number of elements at the beginning (when treating as presorted)
+ * or at the end (we only permute a limited number of group keys).
+ *
+ * The list is decomposed into elements, and we also keep pointers to individual
+ * cells. This allows us to build the permuted list quickly and cheaply, without
+ * creating any copies.
+ */
+static void
+PathkeyMutatorInit(PathkeyMutatorState *state, List *elems, int start, int end)
+{
+ int i;
+ int n = end - start;
+ ListCell *lc;
+
+ memset(state, 0, sizeof(*state));
+
+ state->mutatorNColumns = n;
+
+ state->elemsList = list_copy(elems);
+
+ state->elems = palloc(sizeof(void*) * n);
+ state->elemCells = palloc(sizeof(ListCell*) * n);
+ state->positions = palloc(sizeof(int) * n);
+
+ i = 0;
+ for_each_cell(lc, state->elemsList, list_nth_cell(state->elemsList, start))
+ {
+ state->elemCells[i] = lc;
+ state->elems[i] = lfirst(lc);
+ state->positions[i] = i + 1;
+ i++;
+
+ if (i >= n)
+ break;
+ }
+}
+
+/* Swap two elements of an array. */
+static void
+PathkeyMutatorSwap(int *a, int i, int j)
+{
+ int s = a[i];
+
+ a[i] = a[j];
+ a[j] = s;
+}
+
+/*
+ * Generate the next permutation of elements.
+ */
+static bool
+PathkeyMutatorNextSet(int *a, int n)
+{
+ int j, k, l, r;
+
+ j = n - 2;
+
+ while (j >= 0 && a[j] >= a[j + 1])
+ j--;
+
+ if (j < 0)
+ return false;
+
+ k = n - 1;
+
+ while (k >= 0 && a[j] >= a[k])
+ k--;
+
+ PathkeyMutatorSwap(a, j, k);
+
+ l = j + 1;
+ r = n - 1;
+
+ while (l < r)
+ PathkeyMutatorSwap(a, l++, r--);
+
+ return true;
+}
+
+/*
+ * PathkeyMutatorNext
+ * Generate the next permutation of list of elements.
+ *
+ * Returns the next permutation (as a list of elements) or NIL if there are no
+ * more permutations.
+ */
+static List *
+PathkeyMutatorNext(PathkeyMutatorState *state)
+{
+ int i;
+
+ state->count++;
+
+ /* first permutation is original list */
+ if (state->count == 1)
+ return state->elemsList;
+
+ /* when there are no more permutations, return NIL */
+ if (!PathkeyMutatorNextSet(state->positions, state->mutatorNColumns))
+ {
+ pfree(state->elems);
+ pfree(state->elemCells);
+ pfree(state->positions);
+
+ list_free(state->elemsList);
+
+ return NIL;
+ }
+
+ /* update the list cells to point to the right elements */
+ for(i=0; i<state->mutatorNColumns; i++)
+ lfirst(state->elemCells[i]) =
+ (void *) state->elems[ state->positions[i] - 1 ];
+
+ return state->elemsList;
+}
+
+/*
+ * Cost of comparing pathkeys.
+ */
+typedef struct PathkeySortCost
+{
+ Cost cost;
+ PathKey *pathkey;
+} PathkeySortCost;
+
+static int
+pathkey_sort_cost_comparator(const void *_a, const void *_b)
+{
+ const PathkeySortCost *a = (PathkeySortCost *) _a;
+ const PathkeySortCost *b = (PathkeySortCost *) _b;
+
+ if (a->cost < b->cost)
+ return -1;
+ else if (a->cost == b->cost)
+ return 0;
+ return 1;
+}
+
+/*
+ * get_cheapest_group_keys_order
+ * Returns the pathkeys in an order cheapest to evaluate.
+ *
+ * Given a list of pathkeys, we try to reorder them in a way that minimizes
+ * the CPU cost of sorting. This depends mainly on the cost of comparator
+ * function (the pathkeys may use different data types) and the number of
+ * distinct values in each column (which affects the number of comparator
+ * calls for the following pathkeys).
+ *
+ * In case the input is partially sorted, only the remaining pathkeys are
+ * considered.
+ *
+ * Returns newly allocated lists. If no reordering is possible (or needed),
+ * the lists are set to NIL.
+ */
+static bool
+get_cheapest_group_keys_order(PlannerInfo *root, double nrows,
+ List **group_pathkeys, List **group_clauses,
+ int n_preordered)
+{
+ List *new_group_pathkeys = NIL,
+ *new_group_clauses = NIL,
+ *var_group_pathkeys;
+
+ ListCell *cell;
+ PathkeyMutatorState mstate;
+ double cheapest_sort_cost = -1.0;
+
+ int nFreeKeys;
+ int nToPermute;
+
+ /* If this optimization is disabled, we're done. */
+ if (!debug_cheapest_group_by)
+ return false;
+
+ /* If there are less than 2 unsorted pathkeys, we're done. */
+ if (list_length(*group_pathkeys) - n_preordered < 2)
+ return false;
+
+ /*
+ * We could exhaustively cost all possible orderings of the pathkeys, but for
+ * large number of pathkeys that might be prohibitively expensive. So we try
+ * to apply a simple cheap heuristics first - we sort the pathkeys by sort
+ * cost (as if the pathkey was sorted independently) and then check only the
+ * four cheapest pathkeys. The remaining pathkeys are kept ordered by cost.
+ *
+ * XXX This is a very simple heuristics, and likely to work fine for most
+ * cases (because number of GROUP BY clauses tends to be lower than 4). But
+ * it ignores how the number of distinct values in each pathkey affects the
+ * following sorts. It may be better to use "more expensive" pathkey first
+ * if it has many distinct values, because it then limits the number of
+ * comparisons for the remaining pathkeys. But evaluating that is kinda the
+ * expensive bit we're trying to not do.
+ */
+ nFreeKeys = list_length(*group_pathkeys) - n_preordered;
+ nToPermute = 4;
+ if (nFreeKeys > nToPermute)
+ {
+ int i;
+ PathkeySortCost *costs = palloc(sizeof(PathkeySortCost) * nFreeKeys);
+
+ /* skip the pre-ordered pathkeys */
+ cell = list_nth_cell(*group_pathkeys, n_preordered);
+
+ /* estimate cost for sorting individual pathkeys */
+ for (i = 0; cell != NULL; i++, (cell = lnext(*group_pathkeys, cell)))
+ {
+ List *to_cost = list_make1(lfirst(cell));
+
+ Assert(i < nFreeKeys);
+
+ costs[i].pathkey = lfirst(cell);
+ costs[i].cost = cost_sort_estimate(root, to_cost, 0, nrows);
+
+ pfree(to_cost);
+ }
+
+ /* sort the pathkeys by sort cost in ascending order */
+ qsort(costs, nFreeKeys, sizeof(*costs), pathkey_sort_cost_comparator);
+
+ /*
+ * Rebuild the list of pathkeys - first the preordered ones, then the
+ * rest ordered by cost.
+ */
+ new_group_pathkeys = list_truncate(list_copy(*group_pathkeys), n_preordered);
+
+ for (i = 0; i < nFreeKeys; i++)
+ new_group_pathkeys = lappend(new_group_pathkeys, costs[i].pathkey);
+
+ pfree(costs);
+ }
+ else
+ {
+ /*
+ * Since v13 list_free() can clean list elements so for original list
+ * not to be modified it should be copied to a new one which can then
+ * be cleaned safely if needed.
+ */
+ new_group_pathkeys = list_copy(*group_pathkeys);
+ nToPermute = nFreeKeys;
+ }
+
+ Assert(list_length(new_group_pathkeys) == list_length(*group_pathkeys));
+
+ /*
+ * Generate pathkey lists with permutations of the first nToPermute pathkeys.
+ *
+ * XXX We simply calculate sort cost for each individual pathkey list, but
+ * there's room for two dynamic programming optimizations here. Firstly, we
+ * may pass the current "best" cost to cost_sort_estimate so that it can
+ * "abort" if the estimated pathkeys list exceeds it. Secondly, it could pass
+ * return information about the position when it exceeded the cost, and we
+ * could skip all permutations with the same prefix.
+ *
+ * Imagine we've already found ordering with cost C1, and we're evaluating
+ * another ordering - cost_sort_estimate() calculates cost by adding the
+ * pathkeys one by one (more or less), and the cost only grows. If at any
+ * point it exceeds C1, it can't possibly be "better" so we can discard it.
+ * But we also know that we can discard all ordering with the same prefix,
+ * because if we're estimating (a,b,c,d) and we exceeded C1 at (a,b) then
+ * the same thing will happen for any ordering with this prefix.
+ *
+ *
+ */
+ PathkeyMutatorInit(&mstate, new_group_pathkeys, n_preordered, n_preordered + nToPermute);
+
+ while((var_group_pathkeys = PathkeyMutatorNext(&mstate)) != NIL)
+ {
+ Cost cost;
+
+ cost = cost_sort_estimate(root, var_group_pathkeys, n_preordered, nrows);
+
+ if (cost < cheapest_sort_cost || cheapest_sort_cost < 0)
+ {
+ list_free(new_group_pathkeys);
+ new_group_pathkeys = list_copy(var_group_pathkeys);
+ cheapest_sort_cost = cost;
+ }
+ }
+
+ /* Reorder the group clauses according to the reordered pathkeys. */
+ foreach(cell, new_group_pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(cell);
+
+ new_group_clauses = lappend(new_group_clauses,
+ get_sortgroupref_clause(pathkey->pk_eclass->ec_sortref,
+ *group_clauses));
+ }
+
+ /* Just append the rest GROUP BY clauses */
+ new_group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+
+ *group_pathkeys = new_group_pathkeys;
+ *group_clauses = new_group_clauses;
+
+ return true;
+}
+
+/*
+ * get_useful_group_keys_orderings
+ * Determine which orderings of GROUP BY keys are potentially interesting.
+ *
+ * Returns list of PathKeyInfo items, each representing an interesting ordering
+ * of GROUP BY keys. Each items stores pathkeys and clauses in matching order.
+ *
+ * The function considers (and keeps) multiple group by orderings:
+ *
+ * - the original ordering, as specified by the GROUP BY clause
+ *
+ * - GROUP BY keys reordered to minimize the sort cost
+ *
+ * - GROUP BY keys reordered to match path ordering (as much as possible), with
+ * the tail reoredered to minimize the sort cost
+ *
+ * - GROUP BY keys to match target ORDER BY clause (as much as possible), with
+ * the tail reoredered to minimize the sort cost
+ *
+ * There are other potentially interesting orderings (e.g. it might be best to
+ * match the first ORDER BY key, order the remaining keys differently and then
+ * rely on incremental sort to fix this), but we ignore those for now. To make
+ * this work we'd have to pretty much generate all possible permutations.
+ */
+List *
+get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys,
+ List *group_pathkeys, List *group_clauses)
+{
+ Query *parse = root->parse;
+ List *infos = NIL;
+ PathKeyInfo *info;
+ int n_preordered = 0;
+
+ List *pathkeys = group_pathkeys;
+ List *clauses = group_clauses;
+
+ /* always return at least the original pathkeys/clauses */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+
+ /* for grouping sets we can't do any reordering */
+ if (parse->groupingSets)
+ return infos;
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost, ignoring both the
+ * target ordering (ORDER BY) and ordering of the input path.
+ */
+ if (get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered))
+ {
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * If the path is sorted in some way, try reordering the group keys to match
+ * as much of the ordering as possible - we get this sort for free (mostly).
+ *
+ * We must not do this when there are no grouping sets, because those use
+ * more complex logic to decide the ordering.
+ *
+ * XXX Isn't this somewhat redundant with presorted_keys? Actually, it's
+ * more a complement, because it allows benefiting from incremental sort
+ * as much as possible.
+ *
+ * XXX This does nothing if (n_preordered == 0). We shouldn't create the
+ * info in this case.
+ */
+ if (path_pathkeys)
+ {
+ n_preordered = group_keys_reorder_by_pathkeys(path_pathkeys,
+ &pathkeys,
+ &clauses);
+
+ /* reorder the tail to minimize sort cost */
+ get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, in
+ * which case get_cheapest_group_keys_order returns false. But we
+ * still want to keep the keys reordered to path_pathkeys.
+ */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost (this time consider
+ * the ORDER BY clause, but only if set debug_group_by_match_order_by).
+ *
+ * XXX This does nothing if (n_preordered == 0). We shouldn't create the
+ * info in this case.
+ */
+ if (root->sort_pathkeys && debug_group_by_match_order_by)
+ {
+ n_preordered = group_keys_reorder_by_pathkeys(root->sort_pathkeys,
+ &pathkeys,
+ &clauses);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, in
+ * which case get_cheapest_group_keys_order returns false. But we
+ * still want to keep the keys reordered to sort_pathkeys.
+ */
+ get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered);
+
+ /* keep the group keys reordered to match ordering of input path */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ return infos;
+}
+
/*
* pathkeys_count_contained_in
* Same as pathkeys_contained_in, but also sets length of longest
@@ -1862,6 +2389,54 @@ pathkeys_useful_for_ordering(PlannerInfo *root, List *pathkeys)
return n_common_pathkeys;
}
+/*
+ * pathkeys_useful_for_grouping
+ * Count the number of pathkeys that are useful for grouping (instead of
+ * explicit sort)
+ *
+ * Group pathkeys could be reordered to benefit from the odering. The ordering
+ * may not be "complete" and may require incremental sort, but that's fine. So
+ * we simply count prefix pathkeys with a matching group key, and stop once we
+ * find the first pathkey without a match.
+ *
+ * So e.g. with pathkeys (a,b,c) and group keys (a,b,e) this determines (a,b)
+ * pathkeys are useful for grouping, and we might do incremental sort to get
+ * path ordered by (a,b,e).
+ *
+ * This logic is necessary to retain paths with ordeding not matching grouping
+ * keys directly, without the reordering.
+ *
+ * Returns the length of pathkey prefix with matching group keys.
+ */
+static int
+pathkeys_useful_for_grouping(PlannerInfo *root, List *pathkeys)
+{
+ ListCell *key;
+ int n = 0;
+
+ /* no special ordering requested for grouping */
+ if (root->group_pathkeys == NIL)
+ return 0;
+
+ /* unordered path */
+ if (pathkeys == NIL)
+ return 0;
+
+ /* walk the pathkeys and search for matching group key */
+ foreach(key, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(key);
+
+ /* no matching group key, we're done */
+ if (!list_member_ptr(root->group_pathkeys, pathkey))
+ break;
+
+ n++;
+ }
+
+ return n;
+}
+
/*
* truncate_useless_pathkeys
* Shorten the given pathkey list to just the useful pathkeys.
@@ -1876,6 +2451,9 @@ truncate_useless_pathkeys(PlannerInfo *root,
nuseful = pathkeys_useful_for_merging(root, rel, pathkeys);
nuseful2 = pathkeys_useful_for_ordering(root, pathkeys);
+ if (nuseful2 > nuseful)
+ nuseful = nuseful2;
+ nuseful2 = pathkeys_useful_for_grouping(root, pathkeys);
if (nuseful2 > nuseful)
nuseful = nuseful2;
@@ -1911,6 +2489,8 @@ has_useful_pathkeys(PlannerInfo *root, RelOptInfo *rel)
{
if (rel->joininfo != NIL || rel->has_eclass_joins)
return true; /* might be able to use pathkeys for merging */
+ if (root->group_pathkeys != NIL)
+ return true; /* might be able to use pathkeys for grouping */
if (root->query_pathkeys != NIL)
return true; /* might be able to use them for ordering */
return false; /* definitely useless */
diff --git a/src/backend/optimizer/plan/planner.c b/src/backend/optimizer/plan/planner.c
index bd09f85aea1..fac5822e5b5 100644
--- a/src/backend/optimizer/plan/planner.c
+++ b/src/backend/optimizer/plan/planner.c
@@ -6156,24 +6156,124 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
+ Path *path_save = path;
Path *path_original = path;
- bool is_sorted;
- int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ List *pathkey_orderings = NIL;
+
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
+
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
- if (path == cheapest_path || is_sorted)
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- /* Sort the cheapest-total path if it isn't already sorted */
- if (!is_sorted)
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
+
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (path == cheapest_path || is_sorted)
+ {
+ /* Sort the cheapest-total path if it isn't already sorted */
+ if (!is_sorted)
+ {
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ }
+
+ /* Now decide what to stick atop it */
+ if (parse->groupingSets)
+ {
+ consider_groupingsets_paths(root, grouped_rel,
+ path, true, can_hash,
+ gd, agg_costs, dNumGroups);
+ }
+ else if (parse->hasAggs)
+ {
+ /*
+ * We have aggregation, possibly with plain GROUP BY. Make
+ * an AggPath.
+ */
+ add_path(grouped_rel, (Path *)
+ create_agg_path(root,
+ grouped_rel,
+ path,
+ grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_SIMPLE,
+ info->clauses,
+ havingQual,
+ agg_costs,
+ dNumGroups));
+ }
+ else if (group_clauses)
+ {
+ /*
+ * We have GROUP BY without aggregation or grouping sets.
+ * Make a GroupPath.
+ */
+ add_path(grouped_rel, (Path *)
+ create_group_path(root,
+ grouped_rel,
+ path,
+ info->clauses,
+ havingQual,
+ dNumGroups));
+ }
+ else
+ {
+ /* Other cases should have been handled above */
+ Assert(false);
+ }
+ }
+
+ /*
+ * Now we may consider incremental sort on this path, but only
+ * when the path is not already sorted and when incremental sort
+ * is enabled.
+ */
+ if (is_sorted || !enable_incremental_sort)
+ continue;
+
+ /* Restore the input path (we might have added Sort on top). */
+ path = path_original;
+
+ /* no shared prefix, no point in building incremental sort */
+ if (presorted_keys == 0)
+ continue;
+
+ /*
+ * We should have already excluded pathkeys of length 1 because
+ * then presorted_keys > 0 would imply is_sorted was true.
+ */
+ Assert(list_length(root->group_pathkeys) != 1);
+
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
/* Now decide what to stick atop it */
if (parse->groupingSets)
@@ -6185,17 +6285,17 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
else if (parse->hasAggs)
{
/*
- * We have aggregation, possibly with plain GROUP BY. Make
- * an AggPath.
+ * We have aggregation, possibly with plain GROUP BY. Make an
+ * AggPath.
*/
add_path(grouped_rel, (Path *)
create_agg_path(root,
grouped_rel,
path,
grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_SIMPLE,
- parse->groupClause,
+ info->clauses,
havingQual,
agg_costs,
dNumGroups));
@@ -6203,14 +6303,14 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
else if (parse->groupClause)
{
/*
- * We have GROUP BY without aggregation or grouping sets.
- * Make a GroupPath.
+ * We have GROUP BY without aggregation or grouping sets. Make
+ * a GroupPath.
*/
add_path(grouped_rel, (Path *)
create_group_path(root,
grouped_rel,
path,
- parse->groupClause,
+ info->clauses,
havingQual,
dNumGroups));
}
@@ -6220,79 +6320,6 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
Assert(false);
}
}
-
- /*
- * Now we may consider incremental sort on this path, but only
- * when the path is not already sorted and when incremental sort
- * is enabled.
- */
- if (is_sorted || !enable_incremental_sort)
- continue;
-
- /* Restore the input path (we might have added Sort on top). */
- path = path_original;
-
- /* no shared prefix, no point in building incremental sort */
- if (presorted_keys == 0)
- continue;
-
- /*
- * We should have already excluded pathkeys of length 1 because
- * then presorted_keys > 0 would imply is_sorted was true.
- */
- Assert(list_length(root->group_pathkeys) != 1);
-
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
-
- /* Now decide what to stick atop it */
- if (parse->groupingSets)
- {
- consider_groupingsets_paths(root, grouped_rel,
- path, true, can_hash,
- gd, agg_costs, dNumGroups);
- }
- else if (parse->hasAggs)
- {
- /*
- * We have aggregation, possibly with plain GROUP BY. Make an
- * AggPath.
- */
- add_path(grouped_rel, (Path *)
- create_agg_path(root,
- grouped_rel,
- path,
- grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_SIMPLE,
- parse->groupClause,
- havingQual,
- agg_costs,
- dNumGroups));
- }
- else if (parse->groupClause)
- {
- /*
- * We have GROUP BY without aggregation or grouping sets. Make
- * a GroupPath.
- */
- add_path(grouped_rel, (Path *)
- create_group_path(root,
- grouped_rel,
- path,
- parse->groupClause,
- havingQual,
- dNumGroups));
- }
- else
- {
- /* Other cases should have been handled above */
- Assert(false);
- }
}
/*
@@ -6303,100 +6330,125 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
{
foreach(lc, partially_grouped_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
Path *path_original = path;
- bool is_sorted;
- int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ List *pathkey_orderings = NIL;
- /*
- * Insert a Sort node, if required. But there's no point in
- * sorting anything but the cheapest path.
- */
- if (!is_sorted)
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
+
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- if (path != partially_grouped_rel->cheapest_total_path)
- continue;
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- }
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- if (parse->hasAggs)
- add_path(grouped_rel, (Path *)
- create_agg_path(root,
- grouped_rel,
- path,
- grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_FINAL_DESERIAL,
- parse->groupClause,
- havingQual,
- agg_final_costs,
- dNumGroups));
- else
- add_path(grouped_rel, (Path *)
- create_group_path(root,
- grouped_rel,
- path,
- parse->groupClause,
- havingQual,
- dNumGroups));
+ /* restore the path (we replace it in the loop) */
+ path = path_original;
- /*
- * Now we may consider incremental sort on this path, but only
- * when the path is not already sorted and when incremental
- * sort is enabled.
- */
- if (is_sorted || !enable_incremental_sort)
- continue;
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
- /* Restore the input path (we might have added Sort on top). */
- path = path_original;
+ /*
+ * Insert a Sort node, if required. But there's no point in
+ * sorting anything but the cheapest path.
+ */
+ if (!is_sorted)
+ {
+ if (path != partially_grouped_rel->cheapest_total_path)
+ continue;
- /* no shared prefix, not point in building incremental sort */
- if (presorted_keys == 0)
- continue;
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ }
- /*
- * We should have already excluded pathkeys of length 1
- * because then presorted_keys > 0 would imply is_sorted was
- * true.
- */
- Assert(list_length(root->group_pathkeys) != 1);
+ if (parse->hasAggs)
+ add_path(grouped_rel, (Path *)
+ create_agg_path(root,
+ grouped_rel,
+ path,
+ grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_FINAL_DESERIAL,
+ info->clauses,
+ havingQual,
+ agg_final_costs,
+ dNumGroups));
+ else
+ add_path(grouped_rel, (Path *)
+ create_group_path(root,
+ grouped_rel,
+ path,
+ info->clauses,
+ havingQual,
+ dNumGroups));
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
+ /*
+ * Now we may consider incremental sort on this path, but only
+ * when the path is not already sorted and when incremental
+ * sort is enabled.
+ */
+ if (is_sorted || !enable_incremental_sort)
+ continue;
- if (parse->hasAggs)
- add_path(grouped_rel, (Path *)
- create_agg_path(root,
- grouped_rel,
- path,
- grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_FINAL_DESERIAL,
- parse->groupClause,
- havingQual,
- agg_final_costs,
- dNumGroups));
- else
- add_path(grouped_rel, (Path *)
- create_group_path(root,
- grouped_rel,
- path,
- parse->groupClause,
- havingQual,
- dNumGroups));
+ /* Restore the input path (we might have added Sort on top). */
+ path = path_original;
+
+ /* no shared prefix, not point in building incremental sort */
+ if (presorted_keys == 0)
+ continue;
+
+ /*
+ * We should have already excluded pathkeys of length 1
+ * because then presorted_keys > 0 would imply is_sorted was
+ * true.
+ */
+ Assert(list_length(root->group_pathkeys) != 1);
+
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+
+ if (parse->hasAggs)
+ add_path(grouped_rel, (Path *)
+ create_agg_path(root,
+ grouped_rel,
+ path,
+ grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_FINAL_DESERIAL,
+ info->clauses,
+ havingQual,
+ agg_final_costs,
+ dNumGroups));
+ else
+ add_path(grouped_rel, (Path *)
+ create_group_path(root,
+ grouped_rel,
+ path,
+ info->clauses,
+ havingQual,
+ dNumGroups));
+ }
}
}
}
@@ -6599,41 +6651,71 @@ create_partial_grouping_paths(PlannerInfo *root,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
+ Path *path_save = path;
+
+ List *pathkey_orderings = NIL;
- is_sorted = pathkeys_contained_in(root->group_pathkeys,
- path->pathkeys);
- if (path == cheapest_total_path || is_sorted)
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
+
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- /* Sort the cheapest partial path, if it isn't already */
- if (!is_sorted)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- if (parse->hasAggs)
- add_path(partially_grouped_rel, (Path *)
- create_agg_path(root,
- partially_grouped_rel,
- path,
- partially_grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_INITIAL_SERIAL,
- parse->groupClause,
- NIL,
- agg_partial_costs,
- dNumPartialGroups));
- else
- add_path(partially_grouped_rel, (Path *)
- create_group_path(root,
- partially_grouped_rel,
- path,
- parse->groupClause,
- NIL,
- dNumPartialGroups));
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
+
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (path == cheapest_total_path || is_sorted)
+ {
+ /* Sort the cheapest partial path, if it isn't already */
+ if (!is_sorted)
+ {
+ path = (Path *) create_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_path(partially_grouped_rel, (Path *)
+ create_agg_path(root,
+ partially_grouped_rel,
+ path,
+ partially_grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_INITIAL_SERIAL,
+ info->clauses,
+ NIL,
+ agg_partial_costs,
+ dNumPartialGroups));
+ else
+ add_path(partially_grouped_rel, (Path *)
+ create_group_path(root,
+ partially_grouped_rel,
+ path,
+ info->clauses,
+ NIL,
+ dNumPartialGroups));
+ }
}
}
@@ -6643,6 +6725,8 @@ create_partial_grouping_paths(PlannerInfo *root,
* We can also skip the entire loop when we only have a single-item
* group_pathkeys because then we can't possibly have a presorted
* prefix of the list without having the list be fully sorted.
+ *
+ * XXX Shouldn't this also consider the group-key-reordering?
*/
if (enable_incremental_sort && list_length(root->group_pathkeys) > 1)
{
@@ -6701,24 +6785,100 @@ create_partial_grouping_paths(PlannerInfo *root,
/* Similar to above logic, but for partial paths. */
foreach(lc, input_rel->partial_pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
Path *path_original = path;
- bool is_sorted;
- int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ List *pathkey_orderings = NIL;
+
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
- if (path == cheapest_partial_path || is_sorted)
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- /* Sort the cheapest partial path, if it isn't already */
- if (!is_sorted)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+ /* restore the path (we replace it in the loop) */
+ path = path_original;
+
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (path == cheapest_partial_path || is_sorted)
+ {
+
+ /* Sort the cheapest partial path, if it isn't already */
+ if (!is_sorted)
+ {
+ path = (Path *) create_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_partial_path(partially_grouped_rel, (Path *)
+ create_agg_path(root,
+ partially_grouped_rel,
+ path,
+ partially_grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_INITIAL_SERIAL,
+ info->clauses,
+ NIL,
+ agg_partial_costs,
+ dNumPartialPartialGroups));
+ else
+ add_partial_path(partially_grouped_rel, (Path *)
+ create_group_path(root,
+ partially_grouped_rel,
+ path,
+ info->clauses,
+ NIL,
+ dNumPartialPartialGroups));
+ }
+
+ /*
+ * Now we may consider incremental sort on this path, but only
+ * when the path is not already sorted and when incremental sort
+ * is enabled.
+ */
+ if (is_sorted || !enable_incremental_sort)
+ continue;
+
+ /* Restore the input path (we might have added Sort on top). */
+ path = path_original;
+
+ /* no shared prefix, not point in building incremental sort */
+ if (presorted_keys == 0)
+ continue;
+
+ /*
+ * We should have already excluded pathkeys of length 1 because
+ * then presorted_keys > 0 would imply is_sorted was true.
+ */
+ Assert(list_length(root->group_pathkeys) != 1);
+
+ path = (Path *) create_incremental_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
if (parse->hasAggs)
add_partial_path(partially_grouped_rel, (Path *)
@@ -6726,9 +6886,9 @@ create_partial_grouping_paths(PlannerInfo *root,
partially_grouped_rel,
path,
partially_grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_INITIAL_SERIAL,
- parse->groupClause,
+ info->clauses,
NIL,
agg_partial_costs,
dNumPartialPartialGroups));
@@ -6737,59 +6897,10 @@ create_partial_grouping_paths(PlannerInfo *root,
create_group_path(root,
partially_grouped_rel,
path,
- parse->groupClause,
+ info->clauses,
NIL,
dNumPartialPartialGroups));
}
-
- /*
- * Now we may consider incremental sort on this path, but only
- * when the path is not already sorted and when incremental sort
- * is enabled.
- */
- if (is_sorted || !enable_incremental_sort)
- continue;
-
- /* Restore the input path (we might have added Sort on top). */
- path = path_original;
-
- /* no shared prefix, not point in building incremental sort */
- if (presorted_keys == 0)
- continue;
-
- /*
- * We should have already excluded pathkeys of length 1 because
- * then presorted_keys > 0 would imply is_sorted was true.
- */
- Assert(list_length(root->group_pathkeys) != 1);
-
- path = (Path *) create_incremental_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
-
- if (parse->hasAggs)
- add_partial_path(partially_grouped_rel, (Path *)
- create_agg_path(root,
- partially_grouped_rel,
- path,
- partially_grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_INITIAL_SERIAL,
- parse->groupClause,
- NIL,
- agg_partial_costs,
- dNumPartialPartialGroups));
- else
- add_partial_path(partially_grouped_rel, (Path *)
- create_group_path(root,
- partially_grouped_rel,
- path,
- parse->groupClause,
- NIL,
- dNumPartialPartialGroups));
}
}
diff --git a/src/backend/optimizer/util/pathnode.c b/src/backend/optimizer/util/pathnode.c
index 5c32c96b71c..cd7c671a706 100644
--- a/src/backend/optimizer/util/pathnode.c
+++ b/src/backend/optimizer/util/pathnode.c
@@ -1342,7 +1342,7 @@ create_append_path(PlannerInfo *root,
pathnode->path.pathkeys = child->pathkeys;
}
else
- cost_append(pathnode);
+ cost_append(pathnode, root);
/* If the caller provided a row estimate, override the computed value. */
if (rows >= 0)
diff --git a/src/backend/utils/adt/selfuncs.c b/src/backend/utils/adt/selfuncs.c
index 1fbb0b28c3b..36c303e45c5 100644
--- a/src/backend/utils/adt/selfuncs.c
+++ b/src/backend/utils/adt/selfuncs.c
@@ -3294,7 +3294,10 @@ add_unique_group_var(PlannerInfo *root, List *varinfos,
}
/*
- * estimate_num_groups - Estimate number of groups in a grouped query
+ * estimate_num_groups/estimate_num_groups_incremental
+ * - Estimate number of groups in a grouped query.
+ * _incremental variant is performance optimization for
+ * case of adding one-by-one column
*
* Given a query having a GROUP BY clause, estimate how many groups there
* will be --- ie, the number of distinct combinations of the GROUP BY
@@ -3368,11 +3371,22 @@ double
estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
List **pgset, EstimationInfo *estinfo)
{
- List *varinfos = NIL;
+ return estimate_num_groups_incremental(root, groupExprs,
+ input_rows, pgset, estinfo,
+ NULL, 0);
+}
+
+double
+estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
+ double input_rows,
+ List **pgset, EstimationInfo *estinfo,
+ List **cache_varinfos, int prevNExprs)
+{
+ List *varinfos = (cache_varinfos) ? *cache_varinfos : NIL;
double srf_multiplier = 1.0;
double numdistinct;
ListCell *l;
- int i;
+ int i, j;
/* Zero the estinfo output parameter, if non-NULL */
if (estinfo != NULL)
@@ -3403,7 +3417,7 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
*/
numdistinct = 1.0;
- i = 0;
+ i = j = 0;
foreach(l, groupExprs)
{
Node *groupexpr = (Node *) lfirst(l);
@@ -3412,6 +3426,14 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
List *varshere;
ListCell *l2;
+ /* was done on previous call */
+ if (cache_varinfos && j++ < prevNExprs)
+ {
+ if (pgset)
+ i++; /* to keep in sync with lines below */
+ continue;
+ }
+
/* is expression in this grouping set? */
if (pgset && !list_member_int(*pgset, i++))
continue;
@@ -3481,7 +3503,11 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
if (varshere == NIL)
{
if (contain_volatile_functions(groupexpr))
+ {
+ if (cache_varinfos)
+ *cache_varinfos = varinfos;
return input_rows;
+ }
continue;
}
@@ -3498,6 +3524,9 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
}
}
+ if (cache_varinfos)
+ *cache_varinfos = varinfos;
+
/*
* If now no Vars, we must have an all-constant or all-boolean GROUP BY
* list.
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index 4c94f09c645..8719f67c9d0 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2132,6 +2132,35 @@ static struct config_bool ConfigureNamesBool[] =
NULL, NULL, NULL
},
+/************************<DEBUG OPT GROUP BY>*********************************/
+ {
+ {"debug_group_by_reorder_by_pathkeys", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("enable reorder GROUP BY by pathkeys"),
+ NULL
+ },
+ &debug_group_by_reorder_by_pathkeys,
+ true,
+ NULL, NULL, NULL
+ },
+ {
+ {"debug_enable_group_by_match_order_by", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("enable matching GROUP BY by ORDER BY."),
+ NULL
+ },
+ &debug_group_by_match_order_by,
+ true,
+ NULL, NULL, NULL
+ },
+ {
+ {"debug_enable_cheapest_group_by", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("find a cheapest order of columns in GROUP BY."),
+ NULL
+ },
+ &debug_cheapest_group_by,
+ true,
+ NULL, NULL, NULL
+ },
+/************************</DEBUG OPT GROUP BY>********************************/
/* End-of-list marker */
{
{NULL, 0, 0, NULL, NULL}, NULL, false, NULL, NULL, NULL
diff --git a/src/include/nodes/nodes.h b/src/include/nodes/nodes.h
index f9ddafd3459..543b7f340dc 100644
--- a/src/include/nodes/nodes.h
+++ b/src/include/nodes/nodes.h
@@ -266,6 +266,7 @@ typedef enum NodeTag
T_EquivalenceClass,
T_EquivalenceMember,
T_PathKey,
+ T_PathKeyInfo,
T_PathTarget,
T_RestrictInfo,
T_IndexClause,
diff --git a/src/include/nodes/pathnodes.h b/src/include/nodes/pathnodes.h
index 1f3845b3fec..83582301908 100644
--- a/src/include/nodes/pathnodes.h
+++ b/src/include/nodes/pathnodes.h
@@ -1070,6 +1070,16 @@ typedef struct PathKey
bool pk_nulls_first; /* do NULLs come before normal values? */
} PathKey;
+/*
+ * Combines information about pathkeys and the associated clauses.
+ */
+typedef struct PathKeyInfo
+{
+ NodeTag type;
+ List *pathkeys;
+ List *clauses;
+} PathKeyInfo;
+
/*
* VolatileFunctionStatus -- allows nodes to cache their
* contain_volatile_functions properties. VOLATILITY_UNKNOWN means not yet
diff --git a/src/include/optimizer/cost.h b/src/include/optimizer/cost.h
index 356a51f370a..0b083a27b3c 100644
--- a/src/include/optimizer/cost.h
+++ b/src/include/optimizer/cost.h
@@ -112,7 +112,9 @@ extern void cost_incremental_sort(Path *path,
Cost input_startup_cost, Cost input_total_cost,
double input_tuples, int width, Cost comparison_cost, int sort_mem,
double limit_tuples);
-extern void cost_append(AppendPath *path);
+extern Cost cost_sort_estimate(PlannerInfo *root, List *pathkeys,
+ int nPresortedKeys, double tuples);
+extern void cost_append(AppendPath *path, PlannerInfo *root);
extern void cost_merge_append(Path *path, PlannerInfo *root,
List *pathkeys, int n_streams,
Cost input_startup_cost, Cost input_total_cost,
diff --git a/src/include/optimizer/paths.h b/src/include/optimizer/paths.h
index 0c3a0b90c85..525b594cbd0 100644
--- a/src/include/optimizer/paths.h
+++ b/src/include/optimizer/paths.h
@@ -203,6 +203,17 @@ typedef enum
extern PathKeysComparison compare_pathkeys(List *keys1, List *keys2);
extern bool pathkeys_contained_in(List *keys1, List *keys2);
extern bool pathkeys_count_contained_in(List *keys1, List *keys2, int *n_common);
+extern int group_keys_reorder_by_pathkeys(List *pathkeys,
+ List **group_pathkeys,
+ List **group_clauses);
+/************************<DEBUG OPT GROUP BY>*********************************/
+extern bool debug_group_by_reorder_by_pathkeys;
+extern bool debug_group_by_match_order_by;
+extern bool debug_cheapest_group_by;
+/************************</DEBUG OPT GROUP BY>********************************/
+extern List *get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys,
+ List *pathkeys, List *clauses);
extern Path *get_cheapest_path_for_pathkeys(List *paths, List *pathkeys,
Relids required_outer,
CostSelector cost_criterion,
diff --git a/src/include/utils/selfuncs.h b/src/include/utils/selfuncs.h
index 8f3d73edfb2..c313a08d541 100644
--- a/src/include/utils/selfuncs.h
+++ b/src/include/utils/selfuncs.h
@@ -214,6 +214,11 @@ extern double estimate_num_groups(PlannerInfo *root, List *groupExprs,
double input_rows, List **pgset,
EstimationInfo *estinfo);
+extern double estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
+ double input_rows, List **pgset,
+ EstimationInfo *estinfo,
+ List **cache_varinfos, int prevNExprs);
+
extern void estimate_hash_bucket_stats(PlannerInfo *root,
Node *hashkey, double nbuckets,
Selectivity *mcv_freq,
diff --git a/src/test/regress/expected/aggregates.out b/src/test/regress/expected/aggregates.out
index be5fa5727d5..34cf3d900c1 100644
--- a/src/test/regress/expected/aggregates.out
+++ b/src/test/regress/expected/aggregates.out
@@ -1200,7 +1200,8 @@ explain (costs off)
select distinct min(f1), max(f1) from minmaxtest;
QUERY PLAN
---------------------------------------------------------------------------------------------
- Unique
+ HashAggregate
+ Group Key: $0, $1
InitPlan 1 (returns $0)
-> Limit
-> Merge Append
@@ -1223,10 +1224,8 @@ explain (costs off)
-> Index Only Scan using minmaxtest2i on minmaxtest2 minmaxtest_8
Index Cond: (f1 IS NOT NULL)
-> Index Only Scan Backward using minmaxtest3i on minmaxtest3 minmaxtest_9
- -> Sort
- Sort Key: ($0), ($1)
- -> Result
-(26 rows)
+ -> Result
+(25 rows)
select distinct min(f1), max(f1) from minmaxtest;
min | max
@@ -2438,6 +2437,241 @@ SELECT balk(hundred) FROM tenk1;
(1 row)
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+SELECT
+ i AS id,
+ i/2 AS p,
+ format('%60s', i%2) AS v,
+ i/4 AS c,
+ i/8 AS d,
+ (random() * (10000/8))::int as e --the same as d but no correlation with p
+ INTO btg
+FROM
+ generate_series(1, 10000) i;
+VACUUM btg;
+ANALYZE btg;
+-- GROUP BY optimization by reorder columns by frequency
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Sort
+ Sort Key: p, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Sort
+ Sort Key: p, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, c, v
+ -> Sort
+ Sort Key: p, c, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: v, p, c
+ -> Sort
+ Sort Key: v, p, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: p, d, c, v
+ -> Sort
+ Sort Key: p, d, c, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: v, p, d, c
+ -> Sort
+ Sort Key: v, p, d, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: p, v, d, c
+ -> Sort
+ Sort Key: p, v, d, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, d, e
+ -> Sort
+ Sort Key: p, d, e
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, e, d
+ -> Sort
+ Sort Key: p, e, d
+ -> Seq Scan on btg
+(5 rows)
+
+CREATE STATISTICS btg_dep ON d, e, p FROM btg;
+ANALYZE btg;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, d, e
+ -> Sort
+ Sort Key: p, d, e
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, e, d
+ -> Sort
+ Sort Key: p, e, d
+ -> Seq Scan on btg
+(5 rows)
+
+-- GROUP BY optimization by reorder columns by index scan
+CREATE INDEX ON btg(p, v);
+SET enable_seqscan=off;
+SET enable_bitmapscan=off;
+VACUUM btg;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v ORDER BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p ORDER BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, c, v
+ -> Incremental Sort
+ Sort Key: p, c, v
+ Presorted Key: p
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY p, v;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, v, c
+ -> Incremental Sort
+ Sort Key: p, v, c
+ Presorted Key: p, v
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, c, d, v
+ -> Incremental Sort
+ Sort Key: p, c, d, v
+ Presorted Key: p
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d ORDER BY p, v;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, v, c, d
+ -> Incremental Sort
+ Sort Key: p, v, c, d
+ Presorted Key: p, v
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+DROP TABLE btg;
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+RESET enable_seqscan;
+RESET enable_bitmapscan;
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/expected/incremental_sort.out b/src/test/regress/expected/incremental_sort.out
index 545e301e482..21c429226f7 100644
--- a/src/test/regress/expected/incremental_sort.out
+++ b/src/test/regress/expected/incremental_sort.out
@@ -1439,7 +1439,7 @@ set parallel_setup_cost = 0;
set parallel_tuple_cost = 0;
set max_parallel_workers_per_gather = 2;
create table t (a int, b int, c int);
-insert into t select mod(i,10),mod(i,10),i from generate_series(1,10000) s(i);
+insert into t select mod(i,10),mod(i,10),i from generate_series(1,60000) s(i);
create index on t (a);
analyze t;
set enable_incremental_sort = off;
diff --git a/src/test/regress/expected/join.out b/src/test/regress/expected/join.out
index d5b5b775fdd..01d81292f5c 100644
--- a/src/test/regress/expected/join.out
+++ b/src/test/regress/expected/join.out
@@ -1984,8 +1984,8 @@ USING (name);
------+----+----
bb | 12 | 13
cc | 22 | 23
- dd | | 33
ee | 42 |
+ dd | | 33
(4 rows)
-- Cases with non-nullable expressions in subquery results;
@@ -2019,8 +2019,8 @@ NATURAL FULL JOIN
------+------+------+------+------
bb | 12 | 2 | 13 | 3
cc | 22 | 2 | 23 | 3
- dd | | | 33 | 3
ee | 42 | 2 | |
+ dd | | | 33 | 3
(4 rows)
SELECT * FROM
@@ -4618,18 +4618,20 @@ select d.* from d left join (select * from b group by b.id, b.c_id) s
explain (costs off)
select d.* from d left join (select distinct * from b) s
on d.a = s.id;
- QUERY PLAN
---------------------------------------
- Merge Right Join
- Merge Cond: (b.id = d.a)
- -> Unique
- -> Sort
- Sort Key: b.id, b.c_id
- -> Seq Scan on b
+ QUERY PLAN
+---------------------------------------------
+ Merge Left Join
+ Merge Cond: (d.a = s.id)
-> Sort
Sort Key: d.a
-> Seq Scan on d
-(9 rows)
+ -> Sort
+ Sort Key: s.id
+ -> Subquery Scan on s
+ -> HashAggregate
+ Group Key: b.id, b.c_id
+ -> Seq Scan on b
+(11 rows)
-- check join removal works when uniqueness of the join condition is enforced
-- by a UNION
@@ -6336,44 +6338,39 @@ select * from j1 natural join j2;
explain (verbose, costs off)
select * from j1
inner join (select distinct id from j3) j3 on j1.id = j3.id;
- QUERY PLAN
------------------------------------------
+ QUERY PLAN
+-----------------------------------
Nested Loop
Output: j1.id, j3.id
Inner Unique: true
Join Filter: (j1.id = j3.id)
- -> Unique
+ -> HashAggregate
Output: j3.id
- -> Sort
+ Group Key: j3.id
+ -> Seq Scan on public.j3
Output: j3.id
- Sort Key: j3.id
- -> Seq Scan on public.j3
- Output: j3.id
-> Seq Scan on public.j1
Output: j1.id
-(13 rows)
+(11 rows)
-- ensure group by clause allows the inner to become unique
explain (verbose, costs off)
select * from j1
inner join (select id from j3 group by id) j3 on j1.id = j3.id;
- QUERY PLAN
------------------------------------------
+ QUERY PLAN
+-----------------------------------
Nested Loop
Output: j1.id, j3.id
Inner Unique: true
Join Filter: (j1.id = j3.id)
- -> Group
+ -> HashAggregate
Output: j3.id
Group Key: j3.id
- -> Sort
+ -> Seq Scan on public.j3
Output: j3.id
- Sort Key: j3.id
- -> Seq Scan on public.j3
- Output: j3.id
-> Seq Scan on public.j1
Output: j1.id
-(14 rows)
+(11 rows)
drop table j1;
drop table j2;
diff --git a/src/test/regress/expected/partition_aggregate.out b/src/test/regress/expected/partition_aggregate.out
index dfa4b036b52..a08a3825ff6 100644
--- a/src/test/regress/expected/partition_aggregate.out
+++ b/src/test/regress/expected/partition_aggregate.out
@@ -952,32 +952,30 @@ SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HA
--------------------------------------------------------------------------------------
Sort
Sort Key: pagg_tab_ml.a, (sum(pagg_tab_ml.b)), (array_agg(DISTINCT pagg_tab_ml.c))
- -> Gather
- Workers Planned: 2
- -> Parallel Append
- -> GroupAggregate
- Group Key: pagg_tab_ml.a
- Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml.a
- -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
- -> GroupAggregate
- Group Key: pagg_tab_ml_5.a
- Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_5.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
- -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
- -> GroupAggregate
- Group Key: pagg_tab_ml_2.a
- Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_2.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
- -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
-(27 rows)
+ -> Append
+ -> GroupAggregate
+ Group Key: pagg_tab_ml.a
+ Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml.a
+ -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_2.a
+ Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_2.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
+ -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_5.a
+ Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_5.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
+ -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
+(25 rows)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3 ORDER BY 1, 2, 3;
a | sum | array_agg | count
@@ -996,34 +994,32 @@ SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HA
-- Without ORDER BY clause, to test Gather at top-most path
EXPLAIN (COSTS OFF)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3;
- QUERY PLAN
----------------------------------------------------------------------------
- Gather
- Workers Planned: 2
- -> Parallel Append
- -> GroupAggregate
- Group Key: pagg_tab_ml.a
- Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml.a
- -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
- -> GroupAggregate
- Group Key: pagg_tab_ml_5.a
- Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_5.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
- -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
- -> GroupAggregate
- Group Key: pagg_tab_ml_2.a
- Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_2.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
- -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
-(25 rows)
+ QUERY PLAN
+---------------------------------------------------------------------
+ Append
+ -> GroupAggregate
+ Group Key: pagg_tab_ml.a
+ Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml.a
+ -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_2.a
+ Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_2.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
+ -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_5.a
+ Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_5.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
+ -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
+(23 rows)
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
@@ -1379,28 +1375,26 @@ SELECT x, sum(y), avg(y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) <
-- When GROUP BY clause does not match; partial aggregation is performed for each partition.
EXPLAIN (COSTS OFF)
SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) < 12 ORDER BY 1, 2, 3;
- QUERY PLAN
--------------------------------------------------------------------------------------------
+ QUERY PLAN
+-------------------------------------------------------------------------------------
Sort
Sort Key: pagg_tab_para.y, (sum(pagg_tab_para.x)), (avg(pagg_tab_para.x))
- -> Finalize GroupAggregate
+ -> Finalize HashAggregate
Group Key: pagg_tab_para.y
Filter: (avg(pagg_tab_para.x) < '12'::numeric)
- -> Gather Merge
+ -> Gather
Workers Planned: 2
- -> Sort
- Sort Key: pagg_tab_para.y
- -> Parallel Append
- -> Partial HashAggregate
- Group Key: pagg_tab_para.y
- -> Parallel Seq Scan on pagg_tab_para_p1 pagg_tab_para
- -> Partial HashAggregate
- Group Key: pagg_tab_para_1.y
- -> Parallel Seq Scan on pagg_tab_para_p2 pagg_tab_para_1
- -> Partial HashAggregate
- Group Key: pagg_tab_para_2.y
- -> Parallel Seq Scan on pagg_tab_para_p3 pagg_tab_para_2
-(19 rows)
+ -> Parallel Append
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para.y
+ -> Parallel Seq Scan on pagg_tab_para_p1 pagg_tab_para
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para_1.y
+ -> Parallel Seq Scan on pagg_tab_para_p2 pagg_tab_para_1
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para_2.y
+ -> Parallel Seq Scan on pagg_tab_para_p3 pagg_tab_para_2
+(17 rows)
SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) < 12 ORDER BY 1, 2, 3;
y | sum | avg | count
diff --git a/src/test/regress/expected/partition_join.out b/src/test/regress/expected/partition_join.out
index bb5b7c47a45..03926a84138 100644
--- a/src/test/regress/expected/partition_join.out
+++ b/src/test/regress/expected/partition_join.out
@@ -466,52 +466,41 @@ EXPLAIN (COSTS OFF)
SELECT a, b FROM prt1 FULL JOIN prt2 p2(b,a,c) USING(a,b)
WHERE a BETWEEN 490 AND 510
GROUP BY 1, 2 ORDER BY 1, 2;
- QUERY PLAN
------------------------------------------------------------------------------------------------------------------
+ QUERY PLAN
+-----------------------------------------------------------------------------------------------------------
Group
Group Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Merge Append
+ -> Sort
Sort Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Group
- Group Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Sort
- Sort Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Merge Full Join
- Merge Cond: ((prt1.a = p2.a) AND (prt1.b = p2.b))
- Filter: ((COALESCE(prt1.a, p2.a) >= 490) AND (COALESCE(prt1.a, p2.a) <= 510))
- -> Sort
- Sort Key: prt1.a, prt1.b
- -> Seq Scan on prt1_p1 prt1
- -> Sort
- Sort Key: p2.a, p2.b
- -> Seq Scan on prt2_p1 p2
- -> Group
- Group Key: (COALESCE(prt1_1.a, p2_1.a)), (COALESCE(prt1_1.b, p2_1.b))
- -> Sort
- Sort Key: (COALESCE(prt1_1.a, p2_1.a)), (COALESCE(prt1_1.b, p2_1.b))
- -> Merge Full Join
- Merge Cond: ((prt1_1.a = p2_1.a) AND (prt1_1.b = p2_1.b))
- Filter: ((COALESCE(prt1_1.a, p2_1.a) >= 490) AND (COALESCE(prt1_1.a, p2_1.a) <= 510))
- -> Sort
- Sort Key: prt1_1.a, prt1_1.b
- -> Seq Scan on prt1_p2 prt1_1
- -> Sort
- Sort Key: p2_1.a, p2_1.b
- -> Seq Scan on prt2_p2 p2_1
- -> Group
- Group Key: (COALESCE(prt1_2.a, p2_2.a)), (COALESCE(prt1_2.b, p2_2.b))
- -> Sort
- Sort Key: (COALESCE(prt1_2.a, p2_2.a)), (COALESCE(prt1_2.b, p2_2.b))
- -> Merge Full Join
- Merge Cond: ((prt1_2.a = p2_2.a) AND (prt1_2.b = p2_2.b))
- Filter: ((COALESCE(prt1_2.a, p2_2.a) >= 490) AND (COALESCE(prt1_2.a, p2_2.a) <= 510))
- -> Sort
- Sort Key: prt1_2.a, prt1_2.b
- -> Seq Scan on prt1_p3 prt1_2
- -> Sort
- Sort Key: p2_2.a, p2_2.b
- -> Seq Scan on prt2_p3 p2_2
-(43 rows)
+ -> Append
+ -> Merge Full Join
+ Merge Cond: ((prt1_1.a = p2_1.a) AND (prt1_1.b = p2_1.b))
+ Filter: ((COALESCE(prt1_1.a, p2_1.a) >= 490) AND (COALESCE(prt1_1.a, p2_1.a) <= 510))
+ -> Sort
+ Sort Key: prt1_1.a, prt1_1.b
+ -> Seq Scan on prt1_p1 prt1_1
+ -> Sort
+ Sort Key: p2_1.a, p2_1.b
+ -> Seq Scan on prt2_p1 p2_1
+ -> Merge Full Join
+ Merge Cond: ((prt1_2.a = p2_2.a) AND (prt1_2.b = p2_2.b))
+ Filter: ((COALESCE(prt1_2.a, p2_2.a) >= 490) AND (COALESCE(prt1_2.a, p2_2.a) <= 510))
+ -> Sort
+ Sort Key: prt1_2.a, prt1_2.b
+ -> Seq Scan on prt1_p2 prt1_2
+ -> Sort
+ Sort Key: p2_2.a, p2_2.b
+ -> Seq Scan on prt2_p2 p2_2
+ -> Merge Full Join
+ Merge Cond: ((prt1_3.b = p2_3.b) AND (prt1_3.a = p2_3.a))
+ Filter: ((COALESCE(prt1_3.a, p2_3.a) >= 490) AND (COALESCE(prt1_3.a, p2_3.a) <= 510))
+ -> Sort
+ Sort Key: prt1_3.b, prt1_3.a
+ -> Seq Scan on prt1_p3 prt1_3
+ -> Sort
+ Sort Key: p2_3.b, p2_3.a
+ -> Seq Scan on prt2_p3 p2_3
+(32 rows)
SELECT a, b FROM prt1 FULL JOIN prt2 p2(b,a,c) USING(a,b)
WHERE a BETWEEN 490 AND 510
diff --git a/src/test/regress/expected/union.out b/src/test/regress/expected/union.out
index dece7310cfe..7ac4a9380e2 100644
--- a/src/test/regress/expected/union.out
+++ b/src/test/regress/expected/union.out
@@ -1303,24 +1303,22 @@ select distinct q1 from
union all
select distinct * from int8_tbl i82) ss
where q2 = q2;
- QUERY PLAN
-----------------------------------------------------------
- Unique
- -> Merge Append
- Sort Key: "*SELECT* 1".q1
+ QUERY PLAN
+----------------------------------------------------
+ HashAggregate
+ Group Key: "*SELECT* 1".q1
+ -> Append
-> Subquery Scan on "*SELECT* 1"
- -> Unique
- -> Sort
- Sort Key: i81.q1, i81.q2
- -> Seq Scan on int8_tbl i81
- Filter: (q2 IS NOT NULL)
+ -> HashAggregate
+ Group Key: i81.q1, i81.q2
+ -> Seq Scan on int8_tbl i81
+ Filter: (q2 IS NOT NULL)
-> Subquery Scan on "*SELECT* 2"
- -> Unique
- -> Sort
- Sort Key: i82.q1, i82.q2
- -> Seq Scan on int8_tbl i82
- Filter: (q2 IS NOT NULL)
-(15 rows)
+ -> HashAggregate
+ Group Key: i82.q1, i82.q2
+ -> Seq Scan on int8_tbl i82
+ Filter: (q2 IS NOT NULL)
+(13 rows)
select distinct q1 from
(select distinct * from int8_tbl i81
@@ -1339,24 +1337,22 @@ select distinct q1 from
union all
select distinct * from int8_tbl i82) ss
where -q1 = q2;
- QUERY PLAN
---------------------------------------------------------
- Unique
- -> Merge Append
- Sort Key: "*SELECT* 1".q1
+ QUERY PLAN
+--------------------------------------------------
+ HashAggregate
+ Group Key: "*SELECT* 1".q1
+ -> Append
-> Subquery Scan on "*SELECT* 1"
- -> Unique
- -> Sort
- Sort Key: i81.q1, i81.q2
- -> Seq Scan on int8_tbl i81
- Filter: ((- q1) = q2)
+ -> HashAggregate
+ Group Key: i81.q1, i81.q2
+ -> Seq Scan on int8_tbl i81
+ Filter: ((- q1) = q2)
-> Subquery Scan on "*SELECT* 2"
- -> Unique
- -> Sort
- Sort Key: i82.q1, i82.q2
- -> Seq Scan on int8_tbl i82
- Filter: ((- q1) = q2)
-(15 rows)
+ -> HashAggregate
+ Group Key: i82.q1, i82.q2
+ -> Seq Scan on int8_tbl i82
+ Filter: ((- q1) = q2)
+(13 rows)
select distinct q1 from
(select distinct * from int8_tbl i81
diff --git a/src/test/regress/sql/aggregates.sql b/src/test/regress/sql/aggregates.sql
index 7cf86465e94..a97d4058140 100644
--- a/src/test/regress/sql/aggregates.sql
+++ b/src/test/regress/sql/aggregates.sql
@@ -1010,6 +1010,105 @@ SELECT balk(hundred) FROM tenk1;
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+
+SELECT
+ i AS id,
+ i/2 AS p,
+ format('%60s', i%2) AS v,
+ i/4 AS c,
+ i/8 AS d,
+ (random() * (10000/8))::int as e --the same as d but no correlation with p
+ INTO btg
+FROM
+ generate_series(1, 10000) i;
+
+VACUUM btg;
+ANALYZE btg;
+
+-- GROUP BY optimization by reorder columns by frequency
+
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+
+CREATE STATISTICS btg_dep ON d, e, p FROM btg;
+ANALYZE btg;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+
+
+-- GROUP BY optimization by reorder columns by index scan
+
+CREATE INDEX ON btg(p, v);
+SET enable_seqscan=off;
+SET enable_bitmapscan=off;
+VACUUM btg;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d ORDER BY p, v;
+
+DROP TABLE btg;
+
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+RESET enable_seqscan;
+RESET enable_bitmapscan;
+
+
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/sql/incremental_sort.sql b/src/test/regress/sql/incremental_sort.sql
index d8768a6b54d..1de163e0395 100644
--- a/src/test/regress/sql/incremental_sort.sql
+++ b/src/test/regress/sql/incremental_sort.sql
@@ -213,7 +213,7 @@ set parallel_tuple_cost = 0;
set max_parallel_workers_per_gather = 2;
create table t (a int, b int, c int);
-insert into t select mod(i,10),mod(i,10),i from generate_series(1,10000) s(i);
+insert into t select mod(i,10),mod(i,10),i from generate_series(1,60000) s(i);
create index on t (a);
analyze t;
--
2.31.1
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-01-24 05:02 Andrey Lepikhov <[email protected]>
parent: Tomas Vondra <[email protected]>
1 sibling, 0 replies; 38+ messages in thread
From: Andrey Lepikhov @ 2022-01-24 05:02 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; Teodor Sigaev <[email protected]>; +Cc: Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On 22/1/2022 01:34, Tomas Vondra wrote:
> The other thing we could do is reduce the coefficient gradually - so
> it'd be 1.5 for the first pathkey, then 1.25 for the next one, and so
> on. But it seems somewhat arbitrary (I certainly don't have some sound
> theoretical justification ...).
I think, it hasn't a reason to increase complexity without any theory at
the bottom. Simple solution allows to realize problems much faster, if
they arise.
> ... I've skipped the path_save
> removal in planner.c, because that seems incorrect - if there are
> multiple pathkeys, we must start with the original path, not the
> modified one we built in the last iteration. Or am I missing something
You are right, I misunderstood the idea of path_save variable.
--
regards,
Andrey Lepikhov
Postgres Professional
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-02-10 09:00 Andrey V. Lepikhov <[email protected]>
parent: Tomas Vondra <[email protected]>
1 sibling, 1 reply; 38+ messages in thread
From: Andrey V. Lepikhov @ 2022-02-10 09:00 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; Teodor Sigaev <[email protected]>; +Cc: Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On 1/22/22 01:34, Tomas Vondra wrote:
>
I rebased (with minor fixes) this patch onto current master.
Also, second patch dedicated to a problem of "varno 0" (fake_var).
I think, this case should make the same estimations as in the case of
varno != 0, but no any stats found. So I suppose to restrict number of
groups with min of a number of incoming tuples and DEFAULT_NUM_DISTINCT
value.
--
regards,
Andrey Lepikhov
Postgres Professional
From d5fd0f8f981d9e457320c1007f21f2b9b74aab9e Mon Sep 17 00:00:00 2001
From: Andrey Lepikhov <[email protected]>
Date: Thu, 10 Feb 2022 13:51:51 +0500
Subject: [PATCH 2/2] Use default restriction for number of groups.
---
src/backend/optimizer/path/costsize.c | 22 +++++++---------------
1 file changed, 7 insertions(+), 15 deletions(-)
diff --git a/src/backend/optimizer/path/costsize.c b/src/backend/optimizer/path/costsize.c
index 68a32740d7..b9e975df10 100644
--- a/src/backend/optimizer/path/costsize.c
+++ b/src/backend/optimizer/path/costsize.c
@@ -1756,8 +1756,8 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
/*
* is_fake_var
- * Workaround for generate_append_tlist() which generates fake Vars with
- * varno == 0, that will cause a fail of estimate_num_group() call
+ * Workaround for generate_append_tlist() which generates fake Vars for the
+ * case of "varno 0", that will cause a fail of estimate_num_group() call
*
* XXX Ummm, why would estimate_num_group fail with this?
*/
@@ -1978,21 +1978,13 @@ compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
tuplesPerPrevGroup, NULL, NULL,
&cache_varinfos,
list_length(pathkeyExprs) - 1);
- else if (tuples > 4.0)
+ else
/*
- * Use geometric mean as estimation if there is no any stats.
- * Don't use DEFAULT_NUM_DISTINCT because it used for only one
- * column while here we try to estimate number of groups over
- * set of columns.
- *
- * XXX Perhaps this should use DEFAULT_NUM_DISTINCT at least to
- * limit the calculated values, somehow?
- *
- * XXX What's the logic of the following formula?
+ * In case of full uncertainity use default defensive approach. It
+ * means that any permutations of such vars are equivalent.
+ * Also, see comments for the cost_incremental_sort routine.
*/
- nGroups = ceil(2.0 + sqrt(tuples) * (i + 1) / list_length(pathkeys));
- else
- nGroups = tuples;
+ nGroups = Min(tuplesPerPrevGroup, DEFAULT_NUM_DISTINCT);
/*
* Presorted keys aren't participated in comparison but still checked
--
2.25.1
Attachments:
[text/plain] 0002-Use-default-restriction-for-number-of-groups.txt (2.1K, ../../[email protected]/2-0002-Use-default-restriction-for-number-of-groups.txt)
download | inline diff:
From d5fd0f8f981d9e457320c1007f21f2b9b74aab9e Mon Sep 17 00:00:00 2001
From: Andrey Lepikhov <[email protected]>
Date: Thu, 10 Feb 2022 13:51:51 +0500
Subject: [PATCH 2/2] Use default restriction for number of groups.
---
src/backend/optimizer/path/costsize.c | 22 +++++++---------------
1 file changed, 7 insertions(+), 15 deletions(-)
diff --git a/src/backend/optimizer/path/costsize.c b/src/backend/optimizer/path/costsize.c
index 68a32740d7..b9e975df10 100644
--- a/src/backend/optimizer/path/costsize.c
+++ b/src/backend/optimizer/path/costsize.c
@@ -1756,8 +1756,8 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
/*
* is_fake_var
- * Workaround for generate_append_tlist() which generates fake Vars with
- * varno == 0, that will cause a fail of estimate_num_group() call
+ * Workaround for generate_append_tlist() which generates fake Vars for the
+ * case of "varno 0", that will cause a fail of estimate_num_group() call
*
* XXX Ummm, why would estimate_num_group fail with this?
*/
@@ -1978,21 +1978,13 @@ compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
tuplesPerPrevGroup, NULL, NULL,
&cache_varinfos,
list_length(pathkeyExprs) - 1);
- else if (tuples > 4.0)
+ else
/*
- * Use geometric mean as estimation if there is no any stats.
- * Don't use DEFAULT_NUM_DISTINCT because it used for only one
- * column while here we try to estimate number of groups over
- * set of columns.
- *
- * XXX Perhaps this should use DEFAULT_NUM_DISTINCT at least to
- * limit the calculated values, somehow?
- *
- * XXX What's the logic of the following formula?
+ * In case of full uncertainity use default defensive approach. It
+ * means that any permutations of such vars are equivalent.
+ * Also, see comments for the cost_incremental_sort routine.
*/
- nGroups = ceil(2.0 + sqrt(tuples) * (i + 1) / list_length(pathkeys));
- else
- nGroups = tuples;
+ nGroups = Min(tuplesPerPrevGroup, DEFAULT_NUM_DISTINCT);
/*
* Presorted keys aren't participated in comparison but still checked
--
2.25.1
[text/x-patch] v20220210-0001-GROUP-BY-reordering.patch (100.0K, ../../[email protected]/3-v20220210-0001-GROUP-BY-reordering.patch)
download | inline diff:
From 7357142a0f45313aa09014c4413c011b61aafe5f Mon Sep 17 00:00:00 2001
From: Tomas Vondra <[email protected]>
Date: Fri, 21 Jan 2022 20:22:14 +0100
Subject: [PATCH 1/2] GROUP BY reordering
---
.../postgres_fdw/expected/postgres_fdw.out | 15 +-
src/backend/optimizer/path/costsize.c | 369 +++++++++-
src/backend/optimizer/path/equivclass.c | 13 +-
src/backend/optimizer/path/pathkeys.c | 580 ++++++++++++++++
src/backend/optimizer/plan/planner.c | 653 ++++++++++--------
src/backend/optimizer/util/pathnode.c | 2 +-
src/backend/utils/adt/selfuncs.c | 37 +-
src/backend/utils/misc/guc.c | 32 +
src/include/nodes/nodes.h | 1 +
src/include/nodes/pathnodes.h | 10 +
src/include/optimizer/cost.h | 4 +-
src/include/optimizer/paths.h | 11 +
src/include/utils/selfuncs.h | 5 +
src/test/regress/expected/aggregates.out | 244 ++++++-
src/test/regress/expected/guc.out | 9 +-
.../regress/expected/incremental_sort.out | 2 +-
src/test/regress/expected/join.out | 51 +-
.../regress/expected/partition_aggregate.out | 136 ++--
src/test/regress/expected/partition_join.out | 75 +-
src/test/regress/expected/union.out | 60 +-
src/test/regress/sql/aggregates.sql | 99 +++
src/test/regress/sql/incremental_sort.sql | 2 +-
22 files changed, 1913 insertions(+), 497 deletions(-)
diff --git a/contrib/postgres_fdw/expected/postgres_fdw.out b/contrib/postgres_fdw/expected/postgres_fdw.out
index b2e02caefe..2331d13858 100644
--- a/contrib/postgres_fdw/expected/postgres_fdw.out
+++ b/contrib/postgres_fdw/expected/postgres_fdw.out
@@ -2741,16 +2741,13 @@ select c2 * (random() <= 1)::int as c2 from ft2 group by c2 * (random() <= 1)::i
-- GROUP BY clause in various forms, cardinal, alias and constant expression
explain (verbose, costs off)
select count(c2) w, c2 x, 5 y, 7.0 z from ft1 group by 2, y, 9.0::int order by 2;
- QUERY PLAN
----------------------------------------------------------------------------------------
- Sort
+ QUERY PLAN
+------------------------------------------------------------------------------------------------------------
+ Foreign Scan
Output: (count(c2)), c2, 5, 7.0, 9
- Sort Key: ft1.c2
- -> Foreign Scan
- Output: (count(c2)), c2, 5, 7.0, 9
- Relations: Aggregate on (public.ft1)
- Remote SQL: SELECT count(c2), c2, 5, 7.0, 9 FROM "S 1"."T 1" GROUP BY 2, 3, 5
-(7 rows)
+ Relations: Aggregate on (public.ft1)
+ Remote SQL: SELECT count(c2), c2, 5, 7.0, 9 FROM "S 1"."T 1" GROUP BY 2, 3, 5 ORDER BY c2 ASC NULLS LAST
+(4 rows)
select count(c2) w, c2 x, 5 y, 7.0 z from ft1 group by 2, y, 9.0::int order by 2;
w | x | y | z
diff --git a/src/backend/optimizer/path/costsize.c b/src/backend/optimizer/path/costsize.c
index 8dc7dd4ca2..68a32740d7 100644
--- a/src/backend/optimizer/path/costsize.c
+++ b/src/backend/optimizer/path/costsize.c
@@ -1754,6 +1754,325 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
rterm->pathtarget->width);
}
+/*
+ * is_fake_var
+ * Workaround for generate_append_tlist() which generates fake Vars with
+ * varno == 0, that will cause a fail of estimate_num_group() call
+ *
+ * XXX Ummm, why would estimate_num_group fail with this?
+ */
+static bool
+is_fake_var(Expr *expr)
+{
+ if (IsA(expr, RelabelType))
+ expr = (Expr *) ((RelabelType *) expr)->arg;
+
+ return (IsA(expr, Var) && ((Var *) expr)->varno == 0);
+}
+
+/*
+ * get_width_cost_multiplier
+ * Returns relative complexity of comparing two values based on it's width.
+ * The idea behind - long values have more expensive comparison. Return value is
+ * in cpu_operator_cost unit.
+ */
+static double
+get_width_cost_multiplier(PlannerInfo *root, Expr *expr)
+{
+ double width = -1.0; /* fake value */
+
+ if (IsA(expr, RelabelType))
+ expr = (Expr *) ((RelabelType *) expr)->arg;
+
+ /* Try to find actual stat in corresponding relation */
+ if (IsA(expr, Var))
+ {
+ Var *var = (Var *) expr;
+
+ if (var->varno > 0 && var->varno < root->simple_rel_array_size)
+ {
+ RelOptInfo *rel = root->simple_rel_array[var->varno];
+
+ if (rel != NULL &&
+ var->varattno >= rel->min_attr &&
+ var->varattno <= rel->max_attr)
+ {
+ int ndx = var->varattno - rel->min_attr;
+
+ if (rel->attr_widths[ndx] > 0)
+ width = rel->attr_widths[ndx];
+ }
+ }
+ }
+
+ /* Didn't find any actual stats, use estimation by type */
+ if (width < 0.0)
+ {
+ Node *node = (Node*) expr;
+
+ width = get_typavgwidth(exprType(node), exprTypmod(node));
+ }
+
+ /*
+ * Any value in pgsql is passed by Datum type, so any operation with value
+ * could not be cheaper than operation with Datum type
+ */
+ if (width <= sizeof(Datum))
+ return 1.0;
+
+ /*
+ * Seems, cost of comparision is not directly proportional to args width,
+ * because comparing args could be differ width (we known only average over
+ * column) and difference often could be defined only by looking on first
+ * bytes. So, use log16(width) as estimation.
+ */
+ return 1.0 + 0.125 * LOG2(width / sizeof(Datum));
+}
+
+/*
+ * compute_cpu_sort_cost
+ * compute CPU cost of sort (i.e. in-memory)
+ *
+ * The main thing we need to calculate to estimate sort CPU costs is the number
+ * of calls to the comparator functions. The difficulty is that for multi-column
+ * sorts there may be different data types involved (for some of which the calls
+ * may be much more expensive). Furthermore, the columns may have very different
+ * number of distinct values - the higher the number, the fewer comparisons will
+ * be needed for the following columns.
+ *
+ * The algoritm is incremental - we add pathkeys one by one, and at each step we
+ * estimate the number of necessary comparisons (based on the number of distinct
+ * groups in the current pathkey prefix and the new pathkey), and the comparison
+ * costs (which is data type specific).
+ *
+ * Estimation of the number of comparisons is based on ideas from:
+ *
+ * "Quicksort Is Optimal", Robert Sedgewick, Jon Bentley, 2002
+ * [https://www.cs.princeton.edu/~rs/talks/QuicksortIsOptimal.pdf]
+ *
+ * In term of that paper, let N - number of tuples, Xi - number of identical
+ * tuples with value Ki, then the estimate of number of comparisons is:
+ *
+ * log(N! / (X1! * X2! * ..)) ~ sum(Xi * log(N/Xi))
+ *
+ * We assume all Xi the same because now we don't have any estimation of
+ * group sizes, we have only know the estimate of number of groups (distinct
+ * values). In that case, formula becomes:
+ *
+ * N * log(NumberOfGroups)
+ *
+ * For multi-column sorts we need to estimate the number of comparisons for
+ * each individual column - for example with columns (c1, c2, ..., ck) we
+ * can estimate that number of comparisons on ck is roughly
+ *
+ * ncomparisons(c1, c2, ..., ck) / ncomparisons(c1, c2, ..., c(k-1))
+ *
+ * Let k be a column number, Gk - number of groups defined by k columns, and Fk
+ * the cost of the comparison is
+ *
+ * N * sum( Fk * log(Gk) )
+ *
+ * Note: We also consider column width, not just the comparator cost.
+ *
+ * NOTE: some callers currently pass NIL for pathkeys because they
+ * can't conveniently supply the sort keys. In this case, it will fallback to
+ * simple comparison cost estimate.
+ */
+static Cost
+compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
+ Cost comparison_cost, double tuples, double output_tuples,
+ bool heapSort)
+{
+ Cost per_tuple_cost = 0.0;
+ ListCell *lc;
+ List *pathkeyExprs = NIL;
+ double tuplesPerPrevGroup = tuples;
+ double totalFuncCost = 1.0;
+ bool has_fake_var = false;
+ int i = 0;
+ Oid prev_datatype = InvalidOid;
+ Cost funcCost = 0.0;
+ List *cache_varinfos = NIL;
+
+ /* fallback if pathkeys is unknown */
+ if (list_length(pathkeys) == 0)
+ {
+ /*
+ * If we'll use a bounded heap-sort keeping just K tuples in memory, for
+ * a total number of tuple comparisons of N log2 K; but the constant
+ * factor is a bit higher than for quicksort. Tweak it so that the
+ * cost curve is continuous at the crossover point.
+ *
+ * XXX I suppose the "quicksort factor" references to 1.5 at the end
+ * of this function, but I'm not sure. I suggest we introduce some simple
+ * constants for that, instead of magic values.
+ */
+ output_tuples = (heapSort) ? 2.0 * output_tuples : tuples;
+ per_tuple_cost += 2.0 * cpu_operator_cost * LOG2(output_tuples);
+
+ /* add cost provided by caller */
+ per_tuple_cost += comparison_cost;
+
+ return per_tuple_cost * tuples;
+ }
+
+ /*
+ * Computing total cost of sorting takes into account:
+ * - per column comparison function cost
+ * - we try to compute needed number of comparison per column
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey*) lfirst(lc);
+ EquivalenceMember *em;
+ double nGroups,
+ correctedNGroups;
+
+ /*
+ * We believe that equivalence members aren't very different, so, to
+ * estimate cost we take just first member
+ */
+ em = (EquivalenceMember *) linitial(pathkey->pk_eclass->ec_members);
+
+ if (em->em_datatype != InvalidOid)
+ {
+ /* do not lookup funcCost if data type is the same as previous */
+ if (prev_datatype != em->em_datatype)
+ {
+ Oid sortop;
+ QualCost cost;
+
+ sortop = get_opfamily_member(pathkey->pk_opfamily,
+ em->em_datatype, em->em_datatype,
+ pathkey->pk_strategy);
+
+ cost.startup = 0;
+ cost.per_tuple = 0;
+ add_function_cost(root, get_opcode(sortop), NULL, &cost);
+ /* we need procost, not product of procost and cpu_operator_cost */
+ funcCost = cost.per_tuple / cpu_operator_cost;
+ prev_datatype = em->em_datatype;
+ }
+ }
+ else
+ funcCost = 1.0; /* fallback */
+
+ /* Try to take into account actual width fee */
+ funcCost *= get_width_cost_multiplier(root, em->em_expr);
+
+ totalFuncCost += funcCost;
+
+ /* Remember if we have a fake var in pathkeys */
+ has_fake_var |= is_fake_var(em->em_expr);
+ pathkeyExprs = lappend(pathkeyExprs, em->em_expr);
+
+ /*
+ * Prevent call estimate_num_groups() with fake Var. Note,
+ * pathkeyExprs contains only previous columns
+ */
+ if (has_fake_var == false)
+ /*
+ * Recursively compute number of groups in a group from previous step
+ */
+ nGroups = estimate_num_groups_incremental(root, pathkeyExprs,
+ tuplesPerPrevGroup, NULL, NULL,
+ &cache_varinfos,
+ list_length(pathkeyExprs) - 1);
+ else if (tuples > 4.0)
+ /*
+ * Use geometric mean as estimation if there is no any stats.
+ * Don't use DEFAULT_NUM_DISTINCT because it used for only one
+ * column while here we try to estimate number of groups over
+ * set of columns.
+ *
+ * XXX Perhaps this should use DEFAULT_NUM_DISTINCT at least to
+ * limit the calculated values, somehow?
+ *
+ * XXX What's the logic of the following formula?
+ */
+ nGroups = ceil(2.0 + sqrt(tuples) * (i + 1) / list_length(pathkeys));
+ else
+ nGroups = tuples;
+
+ /*
+ * Presorted keys aren't participated in comparison but still checked
+ * by qsort comparator.
+ */
+ if (i >= nPresortedKeys)
+ {
+ if (heapSort)
+ {
+ double heap_tuples;
+
+ /* have to keep at least one group, and a multiple of group size */
+ heap_tuples = Max(ceil(output_tuples / tuplesPerPrevGroup) * tuplesPerPrevGroup,
+ tuplesPerPrevGroup);
+
+ /* so how many (whole) groups is that? */
+ correctedNGroups = ceil(heap_tuples / tuplesPerPrevGroup);
+ }
+ else
+ /* all groups in the input */
+ correctedNGroups = nGroups;
+
+ correctedNGroups = Max(1.0, ceil(correctedNGroups));
+
+ per_tuple_cost += totalFuncCost * LOG2(correctedNGroups);
+ }
+
+ i++;
+
+ /*
+ * Real-world distribution isn't uniform but now we don't have a way to
+ * determine that, so, add multiplier to get closer to worst case.
+ * But ensure the number of tuples does not exceed the group size in the
+ * preceding step.
+ */
+ tuplesPerPrevGroup = Min(tuplesPerPrevGroup,
+ ceil(1.5 * tuplesPerPrevGroup / nGroups));
+
+ /*
+ * We could skip all following columns for cost estimation, because we
+ * believe that tuples are unique by the set of previous columns
+ */
+ if (tuplesPerPrevGroup <= 1.0)
+ break;
+ }
+
+ list_free(pathkeyExprs);
+
+ /* per_tuple_cost is in cpu_operator_cost units */
+ per_tuple_cost *= cpu_operator_cost;
+
+ /*
+ * Accordingly to "Introduction to algorithms", Thomas H. Cormen, Charles E.
+ * Leiserson, Ronald L. Rivest, ISBN 0-07-013143-0, quicksort estimation
+ * formula has additional term proportional to number of tuples (See Chapter
+ * 8.2 and Theorem 4.1). It has meaning with low number of tuples,
+ * approximately less that 1e4. Of course, it could be implemented as
+ * additional multiplier under logarithm, but use more complicated formula
+ * which takes into account number of unique tuples and it isn't clear how
+ * to combine multiplier with groups. Estimate it as 10 in cpu_operator_cost
+ * unit.
+ */
+ per_tuple_cost += 10 * cpu_operator_cost;
+
+ per_tuple_cost += comparison_cost;
+
+ return tuples * per_tuple_cost;
+}
+
+/*
+ * simple wrapper just to estimate best sort path
+ */
+Cost
+cost_sort_estimate(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
+ double tuples)
+{
+ return compute_cpu_sort_cost(root, pathkeys, nPresortedKeys,
+ 0, tuples, tuples, false);
+}
+
/*
* cost_tuplesort
* Determines and returns the cost of sorting a relation using tuplesort,
@@ -1770,7 +2089,7 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
* number of initial runs formed and M is the merge order used by tuplesort.c.
* Since the average initial run should be about sort_mem, we have
* disk traffic = 2 * relsize * ceil(logM(p / sort_mem))
- * cpu = comparison_cost * t * log2(t)
+ * and cpu cost (computed by compute_cpu_sort_cost()).
*
* If the sort is bounded (i.e., only the first k result tuples are needed)
* and k tuples can fit into sort_mem, we use a heap method that keeps only
@@ -1789,9 +2108,11 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
* 'comparison_cost' is the extra cost per comparison, if any
* 'sort_mem' is the number of kilobytes of work memory allowed for the sort
* 'limit_tuples' is the bound on the number of output tuples; -1 if no bound
+ * 'startup_cost' is expected to be 0 at input. If there is "input cost" it should
+ * be added by caller later.
*/
static void
-cost_tuplesort(Cost *startup_cost, Cost *run_cost,
+cost_tuplesort(PlannerInfo *root, List *pathkeys, Cost *startup_cost, Cost *run_cost,
double tuples, int width,
Cost comparison_cost, int sort_mem,
double limit_tuples)
@@ -1808,9 +2129,6 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
if (tuples < 2.0)
tuples = 2.0;
- /* Include the default cost-per-comparison */
- comparison_cost += 2.0 * cpu_operator_cost;
-
/* Do we have a useful LIMIT? */
if (limit_tuples > 0 && limit_tuples < tuples)
{
@@ -1834,12 +2152,10 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
double log_runs;
double npageaccesses;
- /*
- * CPU costs
- *
- * Assume about N log2 N comparisons
- */
- *startup_cost = comparison_cost * tuples * LOG2(tuples);
+ /* CPU costs */
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ tuples, false);
/* Disk costs */
@@ -1855,18 +2171,17 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
}
else if (tuples > 2 * output_tuples || input_bytes > sort_mem_bytes)
{
- /*
- * We'll use a bounded heap-sort keeping just K tuples in memory, for
- * a total number of tuple comparisons of N log2 K; but the constant
- * factor is a bit higher than for quicksort. Tweak it so that the
- * cost curve is continuous at the crossover point.
- */
- *startup_cost = comparison_cost * tuples * LOG2(2.0 * output_tuples);
+ /* We'll use a bounded heap-sort keeping just K tuples in memory. */
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ output_tuples, true);
}
else
{
/* We'll use plain quicksort on all the input tuples */
- *startup_cost = comparison_cost * tuples * LOG2(tuples);
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ tuples, false);
}
/*
@@ -1899,8 +2214,8 @@ cost_incremental_sort(Path *path,
double input_tuples, int width, Cost comparison_cost, int sort_mem,
double limit_tuples)
{
- Cost startup_cost = 0,
- run_cost = 0,
+ Cost startup_cost,
+ run_cost,
input_run_cost = input_total_cost - input_startup_cost;
double group_tuples,
input_groups;
@@ -1985,7 +2300,7 @@ cost_incremental_sort(Path *path,
* pessimistic about incremental sort performance and increase its average
* group size by half.
*/
- cost_tuplesort(&group_startup_cost, &group_run_cost,
+ cost_tuplesort(root, pathkeys, &group_startup_cost, &group_run_cost,
1.5 * group_tuples, width, comparison_cost, sort_mem,
limit_tuples);
@@ -1993,7 +2308,7 @@ cost_incremental_sort(Path *path,
* Startup cost of incremental sort is the startup cost of its first group
* plus the cost of its input.
*/
- startup_cost += group_startup_cost
+ startup_cost = group_startup_cost
+ input_startup_cost + group_input_run_cost;
/*
@@ -2002,7 +2317,7 @@ cost_incremental_sort(Path *path,
* group, plus the total cost to process the remaining groups, plus the
* remaining cost of input.
*/
- run_cost += group_run_cost
+ run_cost = group_run_cost
+ (group_run_cost + group_startup_cost) * (input_groups - 1)
+ group_input_run_cost * (input_groups - 1);
@@ -2042,7 +2357,7 @@ cost_sort(Path *path, PlannerInfo *root,
Cost startup_cost;
Cost run_cost;
- cost_tuplesort(&startup_cost, &run_cost,
+ cost_tuplesort(root, pathkeys, &startup_cost, &run_cost,
tuples, width,
comparison_cost, sort_mem,
limit_tuples);
@@ -2140,7 +2455,7 @@ append_nonpartial_cost(List *subpaths, int numpaths, int parallel_workers)
* Determines and returns the cost of an Append node.
*/
void
-cost_append(AppendPath *apath)
+cost_append(AppendPath *apath, PlannerInfo *root)
{
ListCell *l;
@@ -2208,7 +2523,7 @@ cost_append(AppendPath *apath)
* any child.
*/
cost_sort(&sort_path,
- NULL, /* doesn't currently need root */
+ root,
pathkeys,
subpath->total_cost,
subpath->rows,
diff --git a/src/backend/optimizer/path/equivclass.c b/src/backend/optimizer/path/equivclass.c
index 8c6770de97..5487ae2ee4 100644
--- a/src/backend/optimizer/path/equivclass.c
+++ b/src/backend/optimizer/path/equivclass.c
@@ -681,7 +681,18 @@ get_eclass_for_sort_expr(PlannerInfo *root,
if (opcintype == cur_em->em_datatype &&
equal(expr, cur_em->em_expr))
- return cur_ec; /* Match! */
+ {
+ /*
+ * Match!
+ *
+ * Copy sortref if it wasn't set yet, it's possible if ec was
+ * constructed from WHERE clause, ie it doesn't have target
+ * reference at all
+ */
+ if (cur_ec->ec_sortref == 0 && sortref > 0)
+ cur_ec->ec_sortref = sortref;
+ return cur_ec;
+ }
}
}
diff --git a/src/backend/optimizer/path/pathkeys.c b/src/backend/optimizer/path/pathkeys.c
index 86a35cdef1..f18d4c8e4e 100644
--- a/src/backend/optimizer/path/pathkeys.c
+++ b/src/backend/optimizer/path/pathkeys.c
@@ -17,16 +17,19 @@
*/
#include "postgres.h"
+#include "miscadmin.h"
#include "access/stratnum.h"
#include "catalog/pg_opfamily.h"
#include "nodes/makefuncs.h"
#include "nodes/nodeFuncs.h"
#include "nodes/plannodes.h"
+#include "optimizer/cost.h"
#include "optimizer/optimizer.h"
#include "optimizer/pathnode.h"
#include "optimizer/paths.h"
#include "partitioning/partbounds.h"
#include "utils/lsyscache.h"
+#include "utils/selfuncs.h"
static bool pathkey_is_redundant(PathKey *new_pathkey, List *pathkeys);
@@ -334,6 +337,530 @@ pathkeys_contained_in(List *keys1, List *keys2)
return false;
}
+/************************<DEBUG PART>*************************************/
+bool debug_group_by_reorder_by_pathkeys = true;
+bool debug_group_by_match_order_by = true;
+bool debug_cheapest_group_by = true;
+/************************</DEBUG PART>************************************/
+
+/*
+ * group_keys_reorder_by_pathkeys
+ * Reorder GROUP BY keys to match pathkeys of input path.
+ *
+ * Function returns new lists (pathkeys and clauses), original GROUP BY lists
+ * stay untouched.
+ *
+ * Returns the number of GROUP BY keys with a matching pathkey.
+ */
+int
+group_keys_reorder_by_pathkeys(List *pathkeys, List **group_pathkeys,
+ List **group_clauses)
+{
+ List *new_group_pathkeys= NIL,
+ *new_group_clauses = NIL;
+ ListCell *lc;
+ int n;
+
+ if (debug_group_by_reorder_by_pathkeys == false)
+ return 0;
+
+ if (pathkeys == NIL || *group_pathkeys == NIL)
+ return 0;
+
+ /*
+ * Walk the pathkeys (determining ordering of the input path) and see if
+ * there's a matching GROUP BY key. If we find one, we append if to the
+ * list, and do the same for the clauses.
+ *
+ * Once we find first pathkey without a matching GROUP BY key, the rest of
+ * the pathkeys is useless and can't be used to evaluate the grouping, so
+ * we abort the loop and ignore the remaining pathkeys.
+ *
+ * XXX Pathkeys are built in a way to allow simply comparing pointers.
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(lc);
+ SortGroupClause *sgc;
+
+ /* abort on first mismatch */
+ if (!list_member_ptr(*group_pathkeys, pathkey))
+ break;
+
+ new_group_pathkeys = lappend(new_group_pathkeys, pathkey);
+
+ sgc = get_sortgroupref_clause(pathkey->pk_eclass->ec_sortref,
+ *group_clauses);
+
+ new_group_clauses = lappend(new_group_clauses, sgc);
+ }
+
+ /* remember the number of pathkeys with a matching GROUP BY key */
+ n = list_length(new_group_pathkeys);
+
+ /* XXX maybe return when (n == 0) */
+
+ /* append the remaining group pathkeys (will be treated as not sorted) */
+ *group_pathkeys = list_concat_unique_ptr(new_group_pathkeys,
+ *group_pathkeys);
+ *group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+
+ return n;
+}
+
+/*
+ * Used to generate all permutations of a pathkey list.
+ */
+typedef struct PathkeyMutatorState {
+ List *elemsList;
+ ListCell **elemCells;
+ void **elems;
+ int *positions;
+ int mutatorNColumns;
+ int count;
+} PathkeyMutatorState;
+
+
+/*
+ * PathkeyMutatorInit
+ * Initialize state of the permutation generator.
+ *
+ * We want to generate permutations of elements in the "elems" list. We may want
+ * to skip some number of elements at the beginning (when treating as presorted)
+ * or at the end (we only permute a limited number of group keys).
+ *
+ * The list is decomposed into elements, and we also keep pointers to individual
+ * cells. This allows us to build the permuted list quickly and cheaply, without
+ * creating any copies.
+ */
+static void
+PathkeyMutatorInit(PathkeyMutatorState *state, List *elems, int start, int end)
+{
+ int i;
+ int n = end - start;
+ ListCell *lc;
+
+ memset(state, 0, sizeof(*state));
+
+ state->mutatorNColumns = n;
+
+ state->elemsList = list_copy(elems);
+
+ state->elems = palloc(sizeof(void*) * n);
+ state->elemCells = palloc(sizeof(ListCell*) * n);
+ state->positions = palloc(sizeof(int) * n);
+
+ i = 0;
+ for_each_cell(lc, state->elemsList, list_nth_cell(state->elemsList, start))
+ {
+ state->elemCells[i] = lc;
+ state->elems[i] = lfirst(lc);
+ state->positions[i] = i + 1;
+ i++;
+
+ if (i >= n)
+ break;
+ }
+}
+
+/* Swap two elements of an array. */
+static void
+PathkeyMutatorSwap(int *a, int i, int j)
+{
+ int s = a[i];
+
+ a[i] = a[j];
+ a[j] = s;
+}
+
+/*
+ * Generate the next permutation of elements.
+ */
+static bool
+PathkeyMutatorNextSet(int *a, int n)
+{
+ int j, k, l, r;
+
+ j = n - 2;
+
+ while (j >= 0 && a[j] >= a[j + 1])
+ j--;
+
+ if (j < 0)
+ return false;
+
+ k = n - 1;
+
+ while (k >= 0 && a[j] >= a[k])
+ k--;
+
+ PathkeyMutatorSwap(a, j, k);
+
+ l = j + 1;
+ r = n - 1;
+
+ while (l < r)
+ PathkeyMutatorSwap(a, l++, r--);
+
+ return true;
+}
+
+/*
+ * PathkeyMutatorNext
+ * Generate the next permutation of list of elements.
+ *
+ * Returns the next permutation (as a list of elements) or NIL if there are no
+ * more permutations.
+ */
+static List *
+PathkeyMutatorNext(PathkeyMutatorState *state)
+{
+ int i;
+
+ state->count++;
+
+ /* first permutation is original list */
+ if (state->count == 1)
+ return state->elemsList;
+
+ /* when there are no more permutations, return NIL */
+ if (!PathkeyMutatorNextSet(state->positions, state->mutatorNColumns))
+ {
+ pfree(state->elems);
+ pfree(state->elemCells);
+ pfree(state->positions);
+
+ list_free(state->elemsList);
+
+ return NIL;
+ }
+
+ /* update the list cells to point to the right elements */
+ for(i=0; i<state->mutatorNColumns; i++)
+ lfirst(state->elemCells[i]) =
+ (void *) state->elems[ state->positions[i] - 1 ];
+
+ return state->elemsList;
+}
+
+/*
+ * Cost of comparing pathkeys.
+ */
+typedef struct PathkeySortCost
+{
+ Cost cost;
+ PathKey *pathkey;
+} PathkeySortCost;
+
+static int
+pathkey_sort_cost_comparator(const void *_a, const void *_b)
+{
+ const PathkeySortCost *a = (PathkeySortCost *) _a;
+ const PathkeySortCost *b = (PathkeySortCost *) _b;
+
+ if (a->cost < b->cost)
+ return -1;
+ else if (a->cost == b->cost)
+ return 0;
+ return 1;
+}
+
+/*
+ * get_cheapest_group_keys_order
+ * Returns the pathkeys in an order cheapest to evaluate.
+ *
+ * Given a list of pathkeys, we try to reorder them in a way that minimizes
+ * the CPU cost of sorting. This depends mainly on the cost of comparator
+ * function (the pathkeys may use different data types) and the number of
+ * distinct values in each column (which affects the number of comparator
+ * calls for the following pathkeys).
+ *
+ * In case the input is partially sorted, only the remaining pathkeys are
+ * considered.
+ *
+ * Returns newly allocated lists. If no reordering is possible (or needed),
+ * the lists are set to NIL.
+ */
+static bool
+get_cheapest_group_keys_order(PlannerInfo *root, double nrows,
+ List **group_pathkeys, List **group_clauses,
+ int n_preordered)
+{
+ List *new_group_pathkeys = NIL,
+ *new_group_clauses = NIL,
+ *var_group_pathkeys;
+
+ ListCell *cell;
+ PathkeyMutatorState mstate;
+ double cheapest_sort_cost = -1.0;
+
+ int nFreeKeys;
+ int nToPermute;
+
+ /* If this optimization is disabled, we're done. */
+ if (!debug_cheapest_group_by)
+ return false;
+
+ /* If there are less than 2 unsorted pathkeys, we're done. */
+ if (list_length(*group_pathkeys) - n_preordered < 2)
+ return false;
+
+ /*
+ * We could exhaustively cost all possible orderings of the pathkeys, but for
+ * large number of pathkeys that might be prohibitively expensive. So we try
+ * to apply a simple cheap heuristics first - we sort the pathkeys by sort
+ * cost (as if the pathkey was sorted independently) and then check only the
+ * four cheapest pathkeys. The remaining pathkeys are kept ordered by cost.
+ *
+ * XXX This is a very simple heuristics, and likely to work fine for most
+ * cases (because number of GROUP BY clauses tends to be lower than 4). But
+ * it ignores how the number of distinct values in each pathkey affects the
+ * following sorts. It may be better to use "more expensive" pathkey first
+ * if it has many distinct values, because it then limits the number of
+ * comparisons for the remaining pathkeys. But evaluating that is kinda the
+ * expensive bit we're trying to not do.
+ */
+ nFreeKeys = list_length(*group_pathkeys) - n_preordered;
+ nToPermute = 4;
+ if (nFreeKeys > nToPermute)
+ {
+ int i;
+ PathkeySortCost *costs = palloc(sizeof(PathkeySortCost) * nFreeKeys);
+
+ /* skip the pre-ordered pathkeys */
+ cell = list_nth_cell(*group_pathkeys, n_preordered);
+
+ /* estimate cost for sorting individual pathkeys */
+ for (i = 0; cell != NULL; i++, (cell = lnext(*group_pathkeys, cell)))
+ {
+ List *to_cost = list_make1(lfirst(cell));
+
+ Assert(i < nFreeKeys);
+
+ costs[i].pathkey = lfirst(cell);
+ costs[i].cost = cost_sort_estimate(root, to_cost, 0, nrows);
+
+ pfree(to_cost);
+ }
+
+ /* sort the pathkeys by sort cost in ascending order */
+ qsort(costs, nFreeKeys, sizeof(*costs), pathkey_sort_cost_comparator);
+
+ /*
+ * Rebuild the list of pathkeys - first the preordered ones, then the
+ * rest ordered by cost.
+ */
+ new_group_pathkeys = list_truncate(list_copy(*group_pathkeys), n_preordered);
+
+ for (i = 0; i < nFreeKeys; i++)
+ new_group_pathkeys = lappend(new_group_pathkeys, costs[i].pathkey);
+
+ pfree(costs);
+ }
+ else
+ {
+ /*
+ * Since v13 list_free() can clean list elements so for original list
+ * not to be modified it should be copied to a new one which can then
+ * be cleaned safely if needed.
+ */
+ new_group_pathkeys = list_copy(*group_pathkeys);
+ nToPermute = nFreeKeys;
+ }
+
+ Assert(list_length(new_group_pathkeys) == list_length(*group_pathkeys));
+
+ /*
+ * Generate pathkey lists with permutations of the first nToPermute pathkeys.
+ *
+ * XXX We simply calculate sort cost for each individual pathkey list, but
+ * there's room for two dynamic programming optimizations here. Firstly, we
+ * may pass the current "best" cost to cost_sort_estimate so that it can
+ * "abort" if the estimated pathkeys list exceeds it. Secondly, it could pass
+ * return information about the position when it exceeded the cost, and we
+ * could skip all permutations with the same prefix.
+ *
+ * Imagine we've already found ordering with cost C1, and we're evaluating
+ * another ordering - cost_sort_estimate() calculates cost by adding the
+ * pathkeys one by one (more or less), and the cost only grows. If at any
+ * point it exceeds C1, it can't possibly be "better" so we can discard it.
+ * But we also know that we can discard all ordering with the same prefix,
+ * because if we're estimating (a,b,c,d) and we exceeded C1 at (a,b) then
+ * the same thing will happen for any ordering with this prefix.
+ *
+ *
+ */
+ PathkeyMutatorInit(&mstate, new_group_pathkeys, n_preordered, n_preordered + nToPermute);
+
+ while((var_group_pathkeys = PathkeyMutatorNext(&mstate)) != NIL)
+ {
+ Cost cost;
+
+ cost = cost_sort_estimate(root, var_group_pathkeys, n_preordered, nrows);
+
+ if (cost < cheapest_sort_cost || cheapest_sort_cost < 0)
+ {
+ list_free(new_group_pathkeys);
+ new_group_pathkeys = list_copy(var_group_pathkeys);
+ cheapest_sort_cost = cost;
+ }
+ }
+
+ /* Reorder the group clauses according to the reordered pathkeys. */
+ foreach(cell, new_group_pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(cell);
+
+ new_group_clauses = lappend(new_group_clauses,
+ get_sortgroupref_clause(pathkey->pk_eclass->ec_sortref,
+ *group_clauses));
+ }
+
+ /* Just append the rest GROUP BY clauses */
+ new_group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+
+ *group_pathkeys = new_group_pathkeys;
+ *group_clauses = new_group_clauses;
+
+ return true;
+}
+
+/*
+ * get_useful_group_keys_orderings
+ * Determine which orderings of GROUP BY keys are potentially interesting.
+ *
+ * Returns list of PathKeyInfo items, each representing an interesting ordering
+ * of GROUP BY keys. Each items stores pathkeys and clauses in matching order.
+ *
+ * The function considers (and keeps) multiple group by orderings:
+ *
+ * - the original ordering, as specified by the GROUP BY clause
+ *
+ * - GROUP BY keys reordered to minimize the sort cost
+ *
+ * - GROUP BY keys reordered to match path ordering (as much as possible), with
+ * the tail reoredered to minimize the sort cost
+ *
+ * - GROUP BY keys to match target ORDER BY clause (as much as possible), with
+ * the tail reoredered to minimize the sort cost
+ *
+ * There are other potentially interesting orderings (e.g. it might be best to
+ * match the first ORDER BY key, order the remaining keys differently and then
+ * rely on incremental sort to fix this), but we ignore those for now. To make
+ * this work we'd have to pretty much generate all possible permutations.
+ */
+List *
+get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys,
+ List *group_pathkeys, List *group_clauses)
+{
+ Query *parse = root->parse;
+ List *infos = NIL;
+ PathKeyInfo *info;
+ int n_preordered = 0;
+
+ List *pathkeys = group_pathkeys;
+ List *clauses = group_clauses;
+
+ /* always return at least the original pathkeys/clauses */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+
+ /* for grouping sets we can't do any reordering */
+ if (parse->groupingSets)
+ return infos;
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost, ignoring both the
+ * target ordering (ORDER BY) and ordering of the input path.
+ */
+ if (get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered))
+ {
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * If the path is sorted in some way, try reordering the group keys to match
+ * as much of the ordering as possible - we get this sort for free (mostly).
+ *
+ * We must not do this when there are no grouping sets, because those use
+ * more complex logic to decide the ordering.
+ *
+ * XXX Isn't this somewhat redundant with presorted_keys? Actually, it's
+ * more a complement, because it allows benefiting from incremental sort
+ * as much as possible.
+ *
+ * XXX This does nothing if (n_preordered == 0). We shouldn't create the
+ * info in this case.
+ */
+ if (path_pathkeys)
+ {
+ n_preordered = group_keys_reorder_by_pathkeys(path_pathkeys,
+ &pathkeys,
+ &clauses);
+
+ /* reorder the tail to minimize sort cost */
+ get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, in
+ * which case get_cheapest_group_keys_order returns false. But we
+ * still want to keep the keys reordered to path_pathkeys.
+ */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost (this time consider
+ * the ORDER BY clause, but only if set debug_group_by_match_order_by).
+ *
+ * XXX This does nothing if (n_preordered == 0). We shouldn't create the
+ * info in this case.
+ */
+ if (root->sort_pathkeys && debug_group_by_match_order_by)
+ {
+ n_preordered = group_keys_reorder_by_pathkeys(root->sort_pathkeys,
+ &pathkeys,
+ &clauses);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, in
+ * which case get_cheapest_group_keys_order returns false. But we
+ * still want to keep the keys reordered to sort_pathkeys.
+ */
+ get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered);
+
+ /* keep the group keys reordered to match ordering of input path */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ return infos;
+}
+
/*
* pathkeys_count_contained_in
* Same as pathkeys_contained_in, but also sets length of longest
@@ -1862,6 +2389,54 @@ pathkeys_useful_for_ordering(PlannerInfo *root, List *pathkeys)
return n_common_pathkeys;
}
+/*
+ * pathkeys_useful_for_grouping
+ * Count the number of pathkeys that are useful for grouping (instead of
+ * explicit sort)
+ *
+ * Group pathkeys could be reordered to benefit from the odering. The ordering
+ * may not be "complete" and may require incremental sort, but that's fine. So
+ * we simply count prefix pathkeys with a matching group key, and stop once we
+ * find the first pathkey without a match.
+ *
+ * So e.g. with pathkeys (a,b,c) and group keys (a,b,e) this determines (a,b)
+ * pathkeys are useful for grouping, and we might do incremental sort to get
+ * path ordered by (a,b,e).
+ *
+ * This logic is necessary to retain paths with ordeding not matching grouping
+ * keys directly, without the reordering.
+ *
+ * Returns the length of pathkey prefix with matching group keys.
+ */
+static int
+pathkeys_useful_for_grouping(PlannerInfo *root, List *pathkeys)
+{
+ ListCell *key;
+ int n = 0;
+
+ /* no special ordering requested for grouping */
+ if (root->group_pathkeys == NIL)
+ return 0;
+
+ /* unordered path */
+ if (pathkeys == NIL)
+ return 0;
+
+ /* walk the pathkeys and search for matching group key */
+ foreach(key, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(key);
+
+ /* no matching group key, we're done */
+ if (!list_member_ptr(root->group_pathkeys, pathkey))
+ break;
+
+ n++;
+ }
+
+ return n;
+}
+
/*
* truncate_useless_pathkeys
* Shorten the given pathkey list to just the useful pathkeys.
@@ -1876,6 +2451,9 @@ truncate_useless_pathkeys(PlannerInfo *root,
nuseful = pathkeys_useful_for_merging(root, rel, pathkeys);
nuseful2 = pathkeys_useful_for_ordering(root, pathkeys);
+ if (nuseful2 > nuseful)
+ nuseful = nuseful2;
+ nuseful2 = pathkeys_useful_for_grouping(root, pathkeys);
if (nuseful2 > nuseful)
nuseful = nuseful2;
@@ -1911,6 +2489,8 @@ has_useful_pathkeys(PlannerInfo *root, RelOptInfo *rel)
{
if (rel->joininfo != NIL || rel->has_eclass_joins)
return true; /* might be able to use pathkeys for merging */
+ if (root->group_pathkeys != NIL)
+ return true; /* might be able to use pathkeys for grouping */
if (root->query_pathkeys != NIL)
return true; /* might be able to use them for ordering */
return false; /* definitely useless */
diff --git a/src/backend/optimizer/plan/planner.c b/src/backend/optimizer/plan/planner.c
index bd09f85aea..fac5822e5b 100644
--- a/src/backend/optimizer/plan/planner.c
+++ b/src/backend/optimizer/plan/planner.c
@@ -6156,24 +6156,124 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
+ Path *path_save = path;
Path *path_original = path;
- bool is_sorted;
- int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ List *pathkey_orderings = NIL;
+
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
+
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
- if (path == cheapest_path || is_sorted)
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- /* Sort the cheapest-total path if it isn't already sorted */
- if (!is_sorted)
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
+
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (path == cheapest_path || is_sorted)
+ {
+ /* Sort the cheapest-total path if it isn't already sorted */
+ if (!is_sorted)
+ {
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ }
+
+ /* Now decide what to stick atop it */
+ if (parse->groupingSets)
+ {
+ consider_groupingsets_paths(root, grouped_rel,
+ path, true, can_hash,
+ gd, agg_costs, dNumGroups);
+ }
+ else if (parse->hasAggs)
+ {
+ /*
+ * We have aggregation, possibly with plain GROUP BY. Make
+ * an AggPath.
+ */
+ add_path(grouped_rel, (Path *)
+ create_agg_path(root,
+ grouped_rel,
+ path,
+ grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_SIMPLE,
+ info->clauses,
+ havingQual,
+ agg_costs,
+ dNumGroups));
+ }
+ else if (group_clauses)
+ {
+ /*
+ * We have GROUP BY without aggregation or grouping sets.
+ * Make a GroupPath.
+ */
+ add_path(grouped_rel, (Path *)
+ create_group_path(root,
+ grouped_rel,
+ path,
+ info->clauses,
+ havingQual,
+ dNumGroups));
+ }
+ else
+ {
+ /* Other cases should have been handled above */
+ Assert(false);
+ }
+ }
+
+ /*
+ * Now we may consider incremental sort on this path, but only
+ * when the path is not already sorted and when incremental sort
+ * is enabled.
+ */
+ if (is_sorted || !enable_incremental_sort)
+ continue;
+
+ /* Restore the input path (we might have added Sort on top). */
+ path = path_original;
+
+ /* no shared prefix, no point in building incremental sort */
+ if (presorted_keys == 0)
+ continue;
+
+ /*
+ * We should have already excluded pathkeys of length 1 because
+ * then presorted_keys > 0 would imply is_sorted was true.
+ */
+ Assert(list_length(root->group_pathkeys) != 1);
+
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
/* Now decide what to stick atop it */
if (parse->groupingSets)
@@ -6185,17 +6285,17 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
else if (parse->hasAggs)
{
/*
- * We have aggregation, possibly with plain GROUP BY. Make
- * an AggPath.
+ * We have aggregation, possibly with plain GROUP BY. Make an
+ * AggPath.
*/
add_path(grouped_rel, (Path *)
create_agg_path(root,
grouped_rel,
path,
grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_SIMPLE,
- parse->groupClause,
+ info->clauses,
havingQual,
agg_costs,
dNumGroups));
@@ -6203,14 +6303,14 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
else if (parse->groupClause)
{
/*
- * We have GROUP BY without aggregation or grouping sets.
- * Make a GroupPath.
+ * We have GROUP BY without aggregation or grouping sets. Make
+ * a GroupPath.
*/
add_path(grouped_rel, (Path *)
create_group_path(root,
grouped_rel,
path,
- parse->groupClause,
+ info->clauses,
havingQual,
dNumGroups));
}
@@ -6220,79 +6320,6 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
Assert(false);
}
}
-
- /*
- * Now we may consider incremental sort on this path, but only
- * when the path is not already sorted and when incremental sort
- * is enabled.
- */
- if (is_sorted || !enable_incremental_sort)
- continue;
-
- /* Restore the input path (we might have added Sort on top). */
- path = path_original;
-
- /* no shared prefix, no point in building incremental sort */
- if (presorted_keys == 0)
- continue;
-
- /*
- * We should have already excluded pathkeys of length 1 because
- * then presorted_keys > 0 would imply is_sorted was true.
- */
- Assert(list_length(root->group_pathkeys) != 1);
-
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
-
- /* Now decide what to stick atop it */
- if (parse->groupingSets)
- {
- consider_groupingsets_paths(root, grouped_rel,
- path, true, can_hash,
- gd, agg_costs, dNumGroups);
- }
- else if (parse->hasAggs)
- {
- /*
- * We have aggregation, possibly with plain GROUP BY. Make an
- * AggPath.
- */
- add_path(grouped_rel, (Path *)
- create_agg_path(root,
- grouped_rel,
- path,
- grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_SIMPLE,
- parse->groupClause,
- havingQual,
- agg_costs,
- dNumGroups));
- }
- else if (parse->groupClause)
- {
- /*
- * We have GROUP BY without aggregation or grouping sets. Make
- * a GroupPath.
- */
- add_path(grouped_rel, (Path *)
- create_group_path(root,
- grouped_rel,
- path,
- parse->groupClause,
- havingQual,
- dNumGroups));
- }
- else
- {
- /* Other cases should have been handled above */
- Assert(false);
- }
}
/*
@@ -6303,100 +6330,125 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
{
foreach(lc, partially_grouped_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
Path *path_original = path;
- bool is_sorted;
- int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ List *pathkey_orderings = NIL;
- /*
- * Insert a Sort node, if required. But there's no point in
- * sorting anything but the cheapest path.
- */
- if (!is_sorted)
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
+
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- if (path != partially_grouped_rel->cheapest_total_path)
- continue;
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- }
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- if (parse->hasAggs)
- add_path(grouped_rel, (Path *)
- create_agg_path(root,
- grouped_rel,
- path,
- grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_FINAL_DESERIAL,
- parse->groupClause,
- havingQual,
- agg_final_costs,
- dNumGroups));
- else
- add_path(grouped_rel, (Path *)
- create_group_path(root,
- grouped_rel,
- path,
- parse->groupClause,
- havingQual,
- dNumGroups));
+ /* restore the path (we replace it in the loop) */
+ path = path_original;
- /*
- * Now we may consider incremental sort on this path, but only
- * when the path is not already sorted and when incremental
- * sort is enabled.
- */
- if (is_sorted || !enable_incremental_sort)
- continue;
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
- /* Restore the input path (we might have added Sort on top). */
- path = path_original;
+ /*
+ * Insert a Sort node, if required. But there's no point in
+ * sorting anything but the cheapest path.
+ */
+ if (!is_sorted)
+ {
+ if (path != partially_grouped_rel->cheapest_total_path)
+ continue;
- /* no shared prefix, not point in building incremental sort */
- if (presorted_keys == 0)
- continue;
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ }
- /*
- * We should have already excluded pathkeys of length 1
- * because then presorted_keys > 0 would imply is_sorted was
- * true.
- */
- Assert(list_length(root->group_pathkeys) != 1);
+ if (parse->hasAggs)
+ add_path(grouped_rel, (Path *)
+ create_agg_path(root,
+ grouped_rel,
+ path,
+ grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_FINAL_DESERIAL,
+ info->clauses,
+ havingQual,
+ agg_final_costs,
+ dNumGroups));
+ else
+ add_path(grouped_rel, (Path *)
+ create_group_path(root,
+ grouped_rel,
+ path,
+ info->clauses,
+ havingQual,
+ dNumGroups));
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
+ /*
+ * Now we may consider incremental sort on this path, but only
+ * when the path is not already sorted and when incremental
+ * sort is enabled.
+ */
+ if (is_sorted || !enable_incremental_sort)
+ continue;
- if (parse->hasAggs)
- add_path(grouped_rel, (Path *)
- create_agg_path(root,
- grouped_rel,
- path,
- grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_FINAL_DESERIAL,
- parse->groupClause,
- havingQual,
- agg_final_costs,
- dNumGroups));
- else
- add_path(grouped_rel, (Path *)
- create_group_path(root,
- grouped_rel,
- path,
- parse->groupClause,
- havingQual,
- dNumGroups));
+ /* Restore the input path (we might have added Sort on top). */
+ path = path_original;
+
+ /* no shared prefix, not point in building incremental sort */
+ if (presorted_keys == 0)
+ continue;
+
+ /*
+ * We should have already excluded pathkeys of length 1
+ * because then presorted_keys > 0 would imply is_sorted was
+ * true.
+ */
+ Assert(list_length(root->group_pathkeys) != 1);
+
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+
+ if (parse->hasAggs)
+ add_path(grouped_rel, (Path *)
+ create_agg_path(root,
+ grouped_rel,
+ path,
+ grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_FINAL_DESERIAL,
+ info->clauses,
+ havingQual,
+ agg_final_costs,
+ dNumGroups));
+ else
+ add_path(grouped_rel, (Path *)
+ create_group_path(root,
+ grouped_rel,
+ path,
+ info->clauses,
+ havingQual,
+ dNumGroups));
+ }
}
}
}
@@ -6599,41 +6651,71 @@ create_partial_grouping_paths(PlannerInfo *root,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
+ Path *path_save = path;
+
+ List *pathkey_orderings = NIL;
- is_sorted = pathkeys_contained_in(root->group_pathkeys,
- path->pathkeys);
- if (path == cheapest_total_path || is_sorted)
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
+
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- /* Sort the cheapest partial path, if it isn't already */
- if (!is_sorted)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- if (parse->hasAggs)
- add_path(partially_grouped_rel, (Path *)
- create_agg_path(root,
- partially_grouped_rel,
- path,
- partially_grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_INITIAL_SERIAL,
- parse->groupClause,
- NIL,
- agg_partial_costs,
- dNumPartialGroups));
- else
- add_path(partially_grouped_rel, (Path *)
- create_group_path(root,
- partially_grouped_rel,
- path,
- parse->groupClause,
- NIL,
- dNumPartialGroups));
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
+
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (path == cheapest_total_path || is_sorted)
+ {
+ /* Sort the cheapest partial path, if it isn't already */
+ if (!is_sorted)
+ {
+ path = (Path *) create_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_path(partially_grouped_rel, (Path *)
+ create_agg_path(root,
+ partially_grouped_rel,
+ path,
+ partially_grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_INITIAL_SERIAL,
+ info->clauses,
+ NIL,
+ agg_partial_costs,
+ dNumPartialGroups));
+ else
+ add_path(partially_grouped_rel, (Path *)
+ create_group_path(root,
+ partially_grouped_rel,
+ path,
+ info->clauses,
+ NIL,
+ dNumPartialGroups));
+ }
}
}
@@ -6643,6 +6725,8 @@ create_partial_grouping_paths(PlannerInfo *root,
* We can also skip the entire loop when we only have a single-item
* group_pathkeys because then we can't possibly have a presorted
* prefix of the list without having the list be fully sorted.
+ *
+ * XXX Shouldn't this also consider the group-key-reordering?
*/
if (enable_incremental_sort && list_length(root->group_pathkeys) > 1)
{
@@ -6701,24 +6785,100 @@ create_partial_grouping_paths(PlannerInfo *root,
/* Similar to above logic, but for partial paths. */
foreach(lc, input_rel->partial_pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
Path *path_original = path;
- bool is_sorted;
- int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ List *pathkey_orderings = NIL;
+
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
- if (path == cheapest_partial_path || is_sorted)
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- /* Sort the cheapest partial path, if it isn't already */
- if (!is_sorted)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+ /* restore the path (we replace it in the loop) */
+ path = path_original;
+
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (path == cheapest_partial_path || is_sorted)
+ {
+
+ /* Sort the cheapest partial path, if it isn't already */
+ if (!is_sorted)
+ {
+ path = (Path *) create_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_partial_path(partially_grouped_rel, (Path *)
+ create_agg_path(root,
+ partially_grouped_rel,
+ path,
+ partially_grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_INITIAL_SERIAL,
+ info->clauses,
+ NIL,
+ agg_partial_costs,
+ dNumPartialPartialGroups));
+ else
+ add_partial_path(partially_grouped_rel, (Path *)
+ create_group_path(root,
+ partially_grouped_rel,
+ path,
+ info->clauses,
+ NIL,
+ dNumPartialPartialGroups));
+ }
+
+ /*
+ * Now we may consider incremental sort on this path, but only
+ * when the path is not already sorted and when incremental sort
+ * is enabled.
+ */
+ if (is_sorted || !enable_incremental_sort)
+ continue;
+
+ /* Restore the input path (we might have added Sort on top). */
+ path = path_original;
+
+ /* no shared prefix, not point in building incremental sort */
+ if (presorted_keys == 0)
+ continue;
+
+ /*
+ * We should have already excluded pathkeys of length 1 because
+ * then presorted_keys > 0 would imply is_sorted was true.
+ */
+ Assert(list_length(root->group_pathkeys) != 1);
+
+ path = (Path *) create_incremental_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
if (parse->hasAggs)
add_partial_path(partially_grouped_rel, (Path *)
@@ -6726,9 +6886,9 @@ create_partial_grouping_paths(PlannerInfo *root,
partially_grouped_rel,
path,
partially_grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_INITIAL_SERIAL,
- parse->groupClause,
+ info->clauses,
NIL,
agg_partial_costs,
dNumPartialPartialGroups));
@@ -6737,59 +6897,10 @@ create_partial_grouping_paths(PlannerInfo *root,
create_group_path(root,
partially_grouped_rel,
path,
- parse->groupClause,
+ info->clauses,
NIL,
dNumPartialPartialGroups));
}
-
- /*
- * Now we may consider incremental sort on this path, but only
- * when the path is not already sorted and when incremental sort
- * is enabled.
- */
- if (is_sorted || !enable_incremental_sort)
- continue;
-
- /* Restore the input path (we might have added Sort on top). */
- path = path_original;
-
- /* no shared prefix, not point in building incremental sort */
- if (presorted_keys == 0)
- continue;
-
- /*
- * We should have already excluded pathkeys of length 1 because
- * then presorted_keys > 0 would imply is_sorted was true.
- */
- Assert(list_length(root->group_pathkeys) != 1);
-
- path = (Path *) create_incremental_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
-
- if (parse->hasAggs)
- add_partial_path(partially_grouped_rel, (Path *)
- create_agg_path(root,
- partially_grouped_rel,
- path,
- partially_grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_INITIAL_SERIAL,
- parse->groupClause,
- NIL,
- agg_partial_costs,
- dNumPartialPartialGroups));
- else
- add_partial_path(partially_grouped_rel, (Path *)
- create_group_path(root,
- partially_grouped_rel,
- path,
- parse->groupClause,
- NIL,
- dNumPartialPartialGroups));
}
}
diff --git a/src/backend/optimizer/util/pathnode.c b/src/backend/optimizer/util/pathnode.c
index 5c32c96b71..cd7c671a70 100644
--- a/src/backend/optimizer/util/pathnode.c
+++ b/src/backend/optimizer/util/pathnode.c
@@ -1342,7 +1342,7 @@ create_append_path(PlannerInfo *root,
pathnode->path.pathkeys = child->pathkeys;
}
else
- cost_append(pathnode);
+ cost_append(pathnode, root);
/* If the caller provided a row estimate, override the computed value. */
if (rows >= 0)
diff --git a/src/backend/utils/adt/selfuncs.c b/src/backend/utils/adt/selfuncs.c
index 1fbb0b28c3..36c303e45c 100644
--- a/src/backend/utils/adt/selfuncs.c
+++ b/src/backend/utils/adt/selfuncs.c
@@ -3294,7 +3294,10 @@ add_unique_group_var(PlannerInfo *root, List *varinfos,
}
/*
- * estimate_num_groups - Estimate number of groups in a grouped query
+ * estimate_num_groups/estimate_num_groups_incremental
+ * - Estimate number of groups in a grouped query.
+ * _incremental variant is performance optimization for
+ * case of adding one-by-one column
*
* Given a query having a GROUP BY clause, estimate how many groups there
* will be --- ie, the number of distinct combinations of the GROUP BY
@@ -3368,11 +3371,22 @@ double
estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
List **pgset, EstimationInfo *estinfo)
{
- List *varinfos = NIL;
+ return estimate_num_groups_incremental(root, groupExprs,
+ input_rows, pgset, estinfo,
+ NULL, 0);
+}
+
+double
+estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
+ double input_rows,
+ List **pgset, EstimationInfo *estinfo,
+ List **cache_varinfos, int prevNExprs)
+{
+ List *varinfos = (cache_varinfos) ? *cache_varinfos : NIL;
double srf_multiplier = 1.0;
double numdistinct;
ListCell *l;
- int i;
+ int i, j;
/* Zero the estinfo output parameter, if non-NULL */
if (estinfo != NULL)
@@ -3403,7 +3417,7 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
*/
numdistinct = 1.0;
- i = 0;
+ i = j = 0;
foreach(l, groupExprs)
{
Node *groupexpr = (Node *) lfirst(l);
@@ -3412,6 +3426,14 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
List *varshere;
ListCell *l2;
+ /* was done on previous call */
+ if (cache_varinfos && j++ < prevNExprs)
+ {
+ if (pgset)
+ i++; /* to keep in sync with lines below */
+ continue;
+ }
+
/* is expression in this grouping set? */
if (pgset && !list_member_int(*pgset, i++))
continue;
@@ -3481,7 +3503,11 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
if (varshere == NIL)
{
if (contain_volatile_functions(groupexpr))
+ {
+ if (cache_varinfos)
+ *cache_varinfos = varinfos;
return input_rows;
+ }
continue;
}
@@ -3498,6 +3524,9 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
}
}
+ if (cache_varinfos)
+ *cache_varinfos = varinfos;
+
/*
* If now no Vars, we must have an all-constant or all-boolean GROUP BY
* list.
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index f505413a7f..bba05a1595 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2132,6 +2132,38 @@ static struct config_bool ConfigureNamesBool[] =
NULL, NULL, NULL
},
+/************************<DEBUG OPT GROUP BY>*********************************/
+ {
+ {"debug_group_by_reorder_by_pathkeys", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("enable reorder GROUP BY by pathkeys"),
+ NULL,
+ GUC_NOT_IN_SAMPLE
+ },
+ &debug_group_by_reorder_by_pathkeys,
+ true,
+ NULL, NULL, NULL
+ },
+ {
+ {"debug_enable_group_by_match_order_by", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("enable matching GROUP BY by ORDER BY."),
+ NULL,
+ GUC_NOT_IN_SAMPLE
+ },
+ &debug_group_by_match_order_by,
+ true,
+ NULL, NULL, NULL
+ },
+ {
+ {"debug_enable_cheapest_group_by", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("find a cheapest order of columns in GROUP BY."),
+ NULL,
+ GUC_NOT_IN_SAMPLE
+ },
+ &debug_cheapest_group_by,
+ true,
+ NULL, NULL, NULL
+ },
+/************************</DEBUG OPT GROUP BY>********************************/
/* End-of-list marker */
{
{NULL, 0, 0, NULL, NULL}, NULL, false, NULL, NULL, NULL
diff --git a/src/include/nodes/nodes.h b/src/include/nodes/nodes.h
index da35f2c272..90a0c4fccb 100644
--- a/src/include/nodes/nodes.h
+++ b/src/include/nodes/nodes.h
@@ -266,6 +266,7 @@ typedef enum NodeTag
T_EquivalenceClass,
T_EquivalenceMember,
T_PathKey,
+ T_PathKeyInfo,
T_PathTarget,
T_RestrictInfo,
T_IndexClause,
diff --git a/src/include/nodes/pathnodes.h b/src/include/nodes/pathnodes.h
index 1f3845b3fe..8358230190 100644
--- a/src/include/nodes/pathnodes.h
+++ b/src/include/nodes/pathnodes.h
@@ -1070,6 +1070,16 @@ typedef struct PathKey
bool pk_nulls_first; /* do NULLs come before normal values? */
} PathKey;
+/*
+ * Combines information about pathkeys and the associated clauses.
+ */
+typedef struct PathKeyInfo
+{
+ NodeTag type;
+ List *pathkeys;
+ List *clauses;
+} PathKeyInfo;
+
/*
* VolatileFunctionStatus -- allows nodes to cache their
* contain_volatile_functions properties. VOLATILITY_UNKNOWN means not yet
diff --git a/src/include/optimizer/cost.h b/src/include/optimizer/cost.h
index 356a51f370..0b083a27b3 100644
--- a/src/include/optimizer/cost.h
+++ b/src/include/optimizer/cost.h
@@ -112,7 +112,9 @@ extern void cost_incremental_sort(Path *path,
Cost input_startup_cost, Cost input_total_cost,
double input_tuples, int width, Cost comparison_cost, int sort_mem,
double limit_tuples);
-extern void cost_append(AppendPath *path);
+extern Cost cost_sort_estimate(PlannerInfo *root, List *pathkeys,
+ int nPresortedKeys, double tuples);
+extern void cost_append(AppendPath *path, PlannerInfo *root);
extern void cost_merge_append(Path *path, PlannerInfo *root,
List *pathkeys, int n_streams,
Cost input_startup_cost, Cost input_total_cost,
diff --git a/src/include/optimizer/paths.h b/src/include/optimizer/paths.h
index 0c3a0b90c8..525b594cbd 100644
--- a/src/include/optimizer/paths.h
+++ b/src/include/optimizer/paths.h
@@ -203,6 +203,17 @@ typedef enum
extern PathKeysComparison compare_pathkeys(List *keys1, List *keys2);
extern bool pathkeys_contained_in(List *keys1, List *keys2);
extern bool pathkeys_count_contained_in(List *keys1, List *keys2, int *n_common);
+extern int group_keys_reorder_by_pathkeys(List *pathkeys,
+ List **group_pathkeys,
+ List **group_clauses);
+/************************<DEBUG OPT GROUP BY>*********************************/
+extern bool debug_group_by_reorder_by_pathkeys;
+extern bool debug_group_by_match_order_by;
+extern bool debug_cheapest_group_by;
+/************************</DEBUG OPT GROUP BY>********************************/
+extern List *get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys,
+ List *pathkeys, List *clauses);
extern Path *get_cheapest_path_for_pathkeys(List *paths, List *pathkeys,
Relids required_outer,
CostSelector cost_criterion,
diff --git a/src/include/utils/selfuncs.h b/src/include/utils/selfuncs.h
index 8f3d73edfb..c313a08d54 100644
--- a/src/include/utils/selfuncs.h
+++ b/src/include/utils/selfuncs.h
@@ -214,6 +214,11 @@ extern double estimate_num_groups(PlannerInfo *root, List *groupExprs,
double input_rows, List **pgset,
EstimationInfo *estinfo);
+extern double estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
+ double input_rows, List **pgset,
+ EstimationInfo *estinfo,
+ List **cache_varinfos, int prevNExprs);
+
extern void estimate_hash_bucket_stats(PlannerInfo *root,
Node *hashkey, double nbuckets,
Selectivity *mcv_freq,
diff --git a/src/test/regress/expected/aggregates.out b/src/test/regress/expected/aggregates.out
index 0a23a39aa2..601047fa3d 100644
--- a/src/test/regress/expected/aggregates.out
+++ b/src/test/regress/expected/aggregates.out
@@ -1210,7 +1210,8 @@ explain (costs off)
select distinct min(f1), max(f1) from minmaxtest;
QUERY PLAN
---------------------------------------------------------------------------------------------
- Unique
+ HashAggregate
+ Group Key: $0, $1
InitPlan 1 (returns $0)
-> Limit
-> Merge Append
@@ -1233,10 +1234,8 @@ explain (costs off)
-> Index Only Scan using minmaxtest2i on minmaxtest2 minmaxtest_8
Index Cond: (f1 IS NOT NULL)
-> Index Only Scan Backward using minmaxtest3i on minmaxtest3 minmaxtest_9
- -> Sort
- Sort Key: ($0), ($1)
- -> Result
-(26 rows)
+ -> Result
+(25 rows)
select distinct min(f1), max(f1) from minmaxtest;
min | max
@@ -2448,6 +2447,241 @@ SELECT balk(hundred) FROM tenk1;
(1 row)
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+SELECT
+ i AS id,
+ i/2 AS p,
+ format('%60s', i%2) AS v,
+ i/4 AS c,
+ i/8 AS d,
+ (random() * (10000/8))::int as e --the same as d but no correlation with p
+ INTO btg
+FROM
+ generate_series(1, 10000) i;
+VACUUM btg;
+ANALYZE btg;
+-- GROUP BY optimization by reorder columns by frequency
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Sort
+ Sort Key: p, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Sort
+ Sort Key: p, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, c, v
+ -> Sort
+ Sort Key: p, c, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: v, p, c
+ -> Sort
+ Sort Key: v, p, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: p, d, c, v
+ -> Sort
+ Sort Key: p, d, c, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: v, p, d, c
+ -> Sort
+ Sort Key: v, p, d, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: p, v, d, c
+ -> Sort
+ Sort Key: p, v, d, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, d, e
+ -> Sort
+ Sort Key: p, d, e
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, e, d
+ -> Sort
+ Sort Key: p, e, d
+ -> Seq Scan on btg
+(5 rows)
+
+CREATE STATISTICS btg_dep ON d, e, p FROM btg;
+ANALYZE btg;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, d, e
+ -> Sort
+ Sort Key: p, d, e
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, e, d
+ -> Sort
+ Sort Key: p, e, d
+ -> Seq Scan on btg
+(5 rows)
+
+-- GROUP BY optimization by reorder columns by index scan
+CREATE INDEX ON btg(p, v);
+SET enable_seqscan=off;
+SET enable_bitmapscan=off;
+VACUUM btg;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v ORDER BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p ORDER BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, c, v
+ -> Incremental Sort
+ Sort Key: p, c, v
+ Presorted Key: p
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY p, v;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, v, c
+ -> Incremental Sort
+ Sort Key: p, v, c
+ Presorted Key: p, v
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, c, d, v
+ -> Incremental Sort
+ Sort Key: p, c, d, v
+ Presorted Key: p
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d ORDER BY p, v;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, v, c, d
+ -> Incremental Sort
+ Sort Key: p, v, c, d
+ Presorted Key: p, v
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+DROP TABLE btg;
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+RESET enable_seqscan;
+RESET enable_bitmapscan;
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/expected/guc.out b/src/test/regress/expected/guc.out
index 75b6bfbf11..2137c0d8a2 100644
--- a/src/test/regress/expected/guc.out
+++ b/src/test/regress/expected/guc.out
@@ -847,10 +847,13 @@ SELECT name FROM tab_settings_flags
SELECT name FROM tab_settings_flags
WHERE category ~ '^Query Tuning' AND NOT explain
ORDER BY 1;
- name
----------------------------
+ name
+--------------------------------------
+ debug_enable_cheapest_group_by
+ debug_enable_group_by_match_order_by
+ debug_group_by_reorder_by_pathkeys
default_statistics_target
-(1 row)
+(4 rows)
-- Runtime-computed GUCs should be part of the preset category.
SELECT name FROM tab_settings_flags
diff --git a/src/test/regress/expected/incremental_sort.out b/src/test/regress/expected/incremental_sort.out
index 545e301e48..21c429226f 100644
--- a/src/test/regress/expected/incremental_sort.out
+++ b/src/test/regress/expected/incremental_sort.out
@@ -1439,7 +1439,7 @@ set parallel_setup_cost = 0;
set parallel_tuple_cost = 0;
set max_parallel_workers_per_gather = 2;
create table t (a int, b int, c int);
-insert into t select mod(i,10),mod(i,10),i from generate_series(1,10000) s(i);
+insert into t select mod(i,10),mod(i,10),i from generate_series(1,60000) s(i);
create index on t (a);
analyze t;
set enable_incremental_sort = off;
diff --git a/src/test/regress/expected/join.out b/src/test/regress/expected/join.out
index 19caebabd0..bf1a2db2cf 100644
--- a/src/test/regress/expected/join.out
+++ b/src/test/regress/expected/join.out
@@ -1984,8 +1984,8 @@ USING (name);
------+----+----
bb | 12 | 13
cc | 22 | 23
- dd | | 33
ee | 42 |
+ dd | | 33
(4 rows)
-- Cases with non-nullable expressions in subquery results;
@@ -2019,8 +2019,8 @@ NATURAL FULL JOIN
------+------+------+------+------
bb | 12 | 2 | 13 | 3
cc | 22 | 2 | 23 | 3
- dd | | | 33 | 3
ee | 42 | 2 | |
+ dd | | | 33 | 3
(4 rows)
SELECT * FROM
@@ -4618,18 +4618,20 @@ select d.* from d left join (select * from b group by b.id, b.c_id) s
explain (costs off)
select d.* from d left join (select distinct * from b) s
on d.a = s.id;
- QUERY PLAN
---------------------------------------
- Merge Right Join
- Merge Cond: (b.id = d.a)
- -> Unique
- -> Sort
- Sort Key: b.id, b.c_id
- -> Seq Scan on b
+ QUERY PLAN
+---------------------------------------------
+ Merge Left Join
+ Merge Cond: (d.a = s.id)
-> Sort
Sort Key: d.a
-> Seq Scan on d
-(9 rows)
+ -> Sort
+ Sort Key: s.id
+ -> Subquery Scan on s
+ -> HashAggregate
+ Group Key: b.id, b.c_id
+ -> Seq Scan on b
+(11 rows)
-- check join removal works when uniqueness of the join condition is enforced
-- by a UNION
@@ -6336,44 +6338,39 @@ select * from j1 natural join j2;
explain (verbose, costs off)
select * from j1
inner join (select distinct id from j3) j3 on j1.id = j3.id;
- QUERY PLAN
------------------------------------------
+ QUERY PLAN
+-----------------------------------
Nested Loop
Output: j1.id, j3.id
Inner Unique: true
Join Filter: (j1.id = j3.id)
- -> Unique
+ -> HashAggregate
Output: j3.id
- -> Sort
+ Group Key: j3.id
+ -> Seq Scan on public.j3
Output: j3.id
- Sort Key: j3.id
- -> Seq Scan on public.j3
- Output: j3.id
-> Seq Scan on public.j1
Output: j1.id
-(13 rows)
+(11 rows)
-- ensure group by clause allows the inner to become unique
explain (verbose, costs off)
select * from j1
inner join (select id from j3 group by id) j3 on j1.id = j3.id;
- QUERY PLAN
------------------------------------------
+ QUERY PLAN
+-----------------------------------
Nested Loop
Output: j1.id, j3.id
Inner Unique: true
Join Filter: (j1.id = j3.id)
- -> Group
+ -> HashAggregate
Output: j3.id
Group Key: j3.id
- -> Sort
+ -> Seq Scan on public.j3
Output: j3.id
- Sort Key: j3.id
- -> Seq Scan on public.j3
- Output: j3.id
-> Seq Scan on public.j1
Output: j1.id
-(14 rows)
+(11 rows)
drop table j1;
drop table j2;
diff --git a/src/test/regress/expected/partition_aggregate.out b/src/test/regress/expected/partition_aggregate.out
index dfa4b036b5..a08a3825ff 100644
--- a/src/test/regress/expected/partition_aggregate.out
+++ b/src/test/regress/expected/partition_aggregate.out
@@ -952,32 +952,30 @@ SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HA
--------------------------------------------------------------------------------------
Sort
Sort Key: pagg_tab_ml.a, (sum(pagg_tab_ml.b)), (array_agg(DISTINCT pagg_tab_ml.c))
- -> Gather
- Workers Planned: 2
- -> Parallel Append
- -> GroupAggregate
- Group Key: pagg_tab_ml.a
- Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml.a
- -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
- -> GroupAggregate
- Group Key: pagg_tab_ml_5.a
- Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_5.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
- -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
- -> GroupAggregate
- Group Key: pagg_tab_ml_2.a
- Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_2.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
- -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
-(27 rows)
+ -> Append
+ -> GroupAggregate
+ Group Key: pagg_tab_ml.a
+ Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml.a
+ -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_2.a
+ Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_2.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
+ -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_5.a
+ Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_5.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
+ -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
+(25 rows)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3 ORDER BY 1, 2, 3;
a | sum | array_agg | count
@@ -996,34 +994,32 @@ SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HA
-- Without ORDER BY clause, to test Gather at top-most path
EXPLAIN (COSTS OFF)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3;
- QUERY PLAN
----------------------------------------------------------------------------
- Gather
- Workers Planned: 2
- -> Parallel Append
- -> GroupAggregate
- Group Key: pagg_tab_ml.a
- Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml.a
- -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
- -> GroupAggregate
- Group Key: pagg_tab_ml_5.a
- Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_5.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
- -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
- -> GroupAggregate
- Group Key: pagg_tab_ml_2.a
- Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_2.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
- -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
-(25 rows)
+ QUERY PLAN
+---------------------------------------------------------------------
+ Append
+ -> GroupAggregate
+ Group Key: pagg_tab_ml.a
+ Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml.a
+ -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_2.a
+ Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_2.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
+ -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_5.a
+ Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_5.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
+ -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
+(23 rows)
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
@@ -1379,28 +1375,26 @@ SELECT x, sum(y), avg(y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) <
-- When GROUP BY clause does not match; partial aggregation is performed for each partition.
EXPLAIN (COSTS OFF)
SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) < 12 ORDER BY 1, 2, 3;
- QUERY PLAN
--------------------------------------------------------------------------------------------
+ QUERY PLAN
+-------------------------------------------------------------------------------------
Sort
Sort Key: pagg_tab_para.y, (sum(pagg_tab_para.x)), (avg(pagg_tab_para.x))
- -> Finalize GroupAggregate
+ -> Finalize HashAggregate
Group Key: pagg_tab_para.y
Filter: (avg(pagg_tab_para.x) < '12'::numeric)
- -> Gather Merge
+ -> Gather
Workers Planned: 2
- -> Sort
- Sort Key: pagg_tab_para.y
- -> Parallel Append
- -> Partial HashAggregate
- Group Key: pagg_tab_para.y
- -> Parallel Seq Scan on pagg_tab_para_p1 pagg_tab_para
- -> Partial HashAggregate
- Group Key: pagg_tab_para_1.y
- -> Parallel Seq Scan on pagg_tab_para_p2 pagg_tab_para_1
- -> Partial HashAggregate
- Group Key: pagg_tab_para_2.y
- -> Parallel Seq Scan on pagg_tab_para_p3 pagg_tab_para_2
-(19 rows)
+ -> Parallel Append
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para.y
+ -> Parallel Seq Scan on pagg_tab_para_p1 pagg_tab_para
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para_1.y
+ -> Parallel Seq Scan on pagg_tab_para_p2 pagg_tab_para_1
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para_2.y
+ -> Parallel Seq Scan on pagg_tab_para_p3 pagg_tab_para_2
+(17 rows)
SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) < 12 ORDER BY 1, 2, 3;
y | sum | avg | count
diff --git a/src/test/regress/expected/partition_join.out b/src/test/regress/expected/partition_join.out
index bb5b7c47a4..03926a8413 100644
--- a/src/test/regress/expected/partition_join.out
+++ b/src/test/regress/expected/partition_join.out
@@ -466,52 +466,41 @@ EXPLAIN (COSTS OFF)
SELECT a, b FROM prt1 FULL JOIN prt2 p2(b,a,c) USING(a,b)
WHERE a BETWEEN 490 AND 510
GROUP BY 1, 2 ORDER BY 1, 2;
- QUERY PLAN
------------------------------------------------------------------------------------------------------------------
+ QUERY PLAN
+-----------------------------------------------------------------------------------------------------------
Group
Group Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Merge Append
+ -> Sort
Sort Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Group
- Group Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Sort
- Sort Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Merge Full Join
- Merge Cond: ((prt1.a = p2.a) AND (prt1.b = p2.b))
- Filter: ((COALESCE(prt1.a, p2.a) >= 490) AND (COALESCE(prt1.a, p2.a) <= 510))
- -> Sort
- Sort Key: prt1.a, prt1.b
- -> Seq Scan on prt1_p1 prt1
- -> Sort
- Sort Key: p2.a, p2.b
- -> Seq Scan on prt2_p1 p2
- -> Group
- Group Key: (COALESCE(prt1_1.a, p2_1.a)), (COALESCE(prt1_1.b, p2_1.b))
- -> Sort
- Sort Key: (COALESCE(prt1_1.a, p2_1.a)), (COALESCE(prt1_1.b, p2_1.b))
- -> Merge Full Join
- Merge Cond: ((prt1_1.a = p2_1.a) AND (prt1_1.b = p2_1.b))
- Filter: ((COALESCE(prt1_1.a, p2_1.a) >= 490) AND (COALESCE(prt1_1.a, p2_1.a) <= 510))
- -> Sort
- Sort Key: prt1_1.a, prt1_1.b
- -> Seq Scan on prt1_p2 prt1_1
- -> Sort
- Sort Key: p2_1.a, p2_1.b
- -> Seq Scan on prt2_p2 p2_1
- -> Group
- Group Key: (COALESCE(prt1_2.a, p2_2.a)), (COALESCE(prt1_2.b, p2_2.b))
- -> Sort
- Sort Key: (COALESCE(prt1_2.a, p2_2.a)), (COALESCE(prt1_2.b, p2_2.b))
- -> Merge Full Join
- Merge Cond: ((prt1_2.a = p2_2.a) AND (prt1_2.b = p2_2.b))
- Filter: ((COALESCE(prt1_2.a, p2_2.a) >= 490) AND (COALESCE(prt1_2.a, p2_2.a) <= 510))
- -> Sort
- Sort Key: prt1_2.a, prt1_2.b
- -> Seq Scan on prt1_p3 prt1_2
- -> Sort
- Sort Key: p2_2.a, p2_2.b
- -> Seq Scan on prt2_p3 p2_2
-(43 rows)
+ -> Append
+ -> Merge Full Join
+ Merge Cond: ((prt1_1.a = p2_1.a) AND (prt1_1.b = p2_1.b))
+ Filter: ((COALESCE(prt1_1.a, p2_1.a) >= 490) AND (COALESCE(prt1_1.a, p2_1.a) <= 510))
+ -> Sort
+ Sort Key: prt1_1.a, prt1_1.b
+ -> Seq Scan on prt1_p1 prt1_1
+ -> Sort
+ Sort Key: p2_1.a, p2_1.b
+ -> Seq Scan on prt2_p1 p2_1
+ -> Merge Full Join
+ Merge Cond: ((prt1_2.a = p2_2.a) AND (prt1_2.b = p2_2.b))
+ Filter: ((COALESCE(prt1_2.a, p2_2.a) >= 490) AND (COALESCE(prt1_2.a, p2_2.a) <= 510))
+ -> Sort
+ Sort Key: prt1_2.a, prt1_2.b
+ -> Seq Scan on prt1_p2 prt1_2
+ -> Sort
+ Sort Key: p2_2.a, p2_2.b
+ -> Seq Scan on prt2_p2 p2_2
+ -> Merge Full Join
+ Merge Cond: ((prt1_3.b = p2_3.b) AND (prt1_3.a = p2_3.a))
+ Filter: ((COALESCE(prt1_3.a, p2_3.a) >= 490) AND (COALESCE(prt1_3.a, p2_3.a) <= 510))
+ -> Sort
+ Sort Key: prt1_3.b, prt1_3.a
+ -> Seq Scan on prt1_p3 prt1_3
+ -> Sort
+ Sort Key: p2_3.b, p2_3.a
+ -> Seq Scan on prt2_p3 p2_3
+(32 rows)
SELECT a, b FROM prt1 FULL JOIN prt2 p2(b,a,c) USING(a,b)
WHERE a BETWEEN 490 AND 510
diff --git a/src/test/regress/expected/union.out b/src/test/regress/expected/union.out
index dece7310cf..7ac4a9380e 100644
--- a/src/test/regress/expected/union.out
+++ b/src/test/regress/expected/union.out
@@ -1303,24 +1303,22 @@ select distinct q1 from
union all
select distinct * from int8_tbl i82) ss
where q2 = q2;
- QUERY PLAN
-----------------------------------------------------------
- Unique
- -> Merge Append
- Sort Key: "*SELECT* 1".q1
+ QUERY PLAN
+----------------------------------------------------
+ HashAggregate
+ Group Key: "*SELECT* 1".q1
+ -> Append
-> Subquery Scan on "*SELECT* 1"
- -> Unique
- -> Sort
- Sort Key: i81.q1, i81.q2
- -> Seq Scan on int8_tbl i81
- Filter: (q2 IS NOT NULL)
+ -> HashAggregate
+ Group Key: i81.q1, i81.q2
+ -> Seq Scan on int8_tbl i81
+ Filter: (q2 IS NOT NULL)
-> Subquery Scan on "*SELECT* 2"
- -> Unique
- -> Sort
- Sort Key: i82.q1, i82.q2
- -> Seq Scan on int8_tbl i82
- Filter: (q2 IS NOT NULL)
-(15 rows)
+ -> HashAggregate
+ Group Key: i82.q1, i82.q2
+ -> Seq Scan on int8_tbl i82
+ Filter: (q2 IS NOT NULL)
+(13 rows)
select distinct q1 from
(select distinct * from int8_tbl i81
@@ -1339,24 +1337,22 @@ select distinct q1 from
union all
select distinct * from int8_tbl i82) ss
where -q1 = q2;
- QUERY PLAN
---------------------------------------------------------
- Unique
- -> Merge Append
- Sort Key: "*SELECT* 1".q1
+ QUERY PLAN
+--------------------------------------------------
+ HashAggregate
+ Group Key: "*SELECT* 1".q1
+ -> Append
-> Subquery Scan on "*SELECT* 1"
- -> Unique
- -> Sort
- Sort Key: i81.q1, i81.q2
- -> Seq Scan on int8_tbl i81
- Filter: ((- q1) = q2)
+ -> HashAggregate
+ Group Key: i81.q1, i81.q2
+ -> Seq Scan on int8_tbl i81
+ Filter: ((- q1) = q2)
-> Subquery Scan on "*SELECT* 2"
- -> Unique
- -> Sort
- Sort Key: i82.q1, i82.q2
- -> Seq Scan on int8_tbl i82
- Filter: ((- q1) = q2)
-(15 rows)
+ -> HashAggregate
+ Group Key: i82.q1, i82.q2
+ -> Seq Scan on int8_tbl i82
+ Filter: ((- q1) = q2)
+(13 rows)
select distinct q1 from
(select distinct * from int8_tbl i81
diff --git a/src/test/regress/sql/aggregates.sql b/src/test/regress/sql/aggregates.sql
index 2f5d0e00f3..c6e0d7ba2b 100644
--- a/src/test/regress/sql/aggregates.sql
+++ b/src/test/regress/sql/aggregates.sql
@@ -1025,6 +1025,105 @@ SELECT balk(hundred) FROM tenk1;
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+
+SELECT
+ i AS id,
+ i/2 AS p,
+ format('%60s', i%2) AS v,
+ i/4 AS c,
+ i/8 AS d,
+ (random() * (10000/8))::int as e --the same as d but no correlation with p
+ INTO btg
+FROM
+ generate_series(1, 10000) i;
+
+VACUUM btg;
+ANALYZE btg;
+
+-- GROUP BY optimization by reorder columns by frequency
+
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+
+CREATE STATISTICS btg_dep ON d, e, p FROM btg;
+ANALYZE btg;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+
+
+-- GROUP BY optimization by reorder columns by index scan
+
+CREATE INDEX ON btg(p, v);
+SET enable_seqscan=off;
+SET enable_bitmapscan=off;
+VACUUM btg;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d ORDER BY p, v;
+
+DROP TABLE btg;
+
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+RESET enable_seqscan;
+RESET enable_bitmapscan;
+
+
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/sql/incremental_sort.sql b/src/test/regress/sql/incremental_sort.sql
index d8768a6b54..1de163e039 100644
--- a/src/test/regress/sql/incremental_sort.sql
+++ b/src/test/regress/sql/incremental_sort.sql
@@ -213,7 +213,7 @@ set parallel_tuple_cost = 0;
set max_parallel_workers_per_gather = 2;
create table t (a int, b int, c int);
-insert into t select mod(i,10),mod(i,10),i from generate_series(1,10000) s(i);
+insert into t select mod(i,10),mod(i,10),i from generate_series(1,60000) s(i);
create index on t (a);
analyze t;
--
2.25.1
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-03-15 08:26 Tomas Vondra <[email protected]>
parent: Andrey V. Lepikhov <[email protected]>
0 siblings, 1 reply; 38+ messages in thread
From: Tomas Vondra @ 2022-03-15 08:26 UTC (permalink / raw)
To: Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; +Cc: Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On 2/10/22 10:00, Andrey V. Lepikhov wrote:
> On 1/22/22 01:34, Tomas Vondra wrote:
>>
>
> I rebased (with minor fixes) this patch onto current master.
>
> Also, second patch dedicated to a problem of "varno 0" (fake_var).
> I think, this case should make the same estimations as in the case of
> varno != 0, but no any stats found. So I suppose to restrict number of
> groups with min of a number of incoming tuples and DEFAULT_NUM_DISTINCT
> value.
>
Thanks for the rebase. The two proposed changes (tweaked costing and
simplified fake_var handling) seem fine to me. I think the last thing
that needs to be done is cleanup of the debug GUCs, which I added to
allow easier experimentation with the patch.
I probably won't remove the GUCs entirely, though. I plan to add a
single GUC that would enable/disable this optimization. I'm not a huge
fan of adding more and more GUCs, but in this case it's probably the
right thing to do given the complexity of estimating cost with
correlated columns etc.
regards
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
Attachments:
[text/x-patch] 0001-Optimize-order-of-GROUP-BY-keys-20220315.patch (102.1K, ../../[email protected]/2-0001-Optimize-order-of-GROUP-BY-keys-20220315.patch)
download | inline diff:
From dca50a28b0be21decde3628302a2f8546875e884 Mon Sep 17 00:00:00 2001
From: Tomas Vondra <[email protected]>
Date: Fri, 21 Jan 2022 20:22:14 +0100
Subject: [PATCH 1/3] Optimize order of GROUP BY keys
When evaluating a query with a multi-column GROUP BY clause using sort,
the cost may be heavily dependent on the order in which the keys are
compared when building the groups. Grouping does not imply any ordering,
so we're allowed to compare the keys in arbitrary order, and a Hash Agg
leverages this. But for Group Agg, we simply compared keys in the order
as specified in the query. This commit explores alternative ordering of
the keys, trying to find a cheaper one.
To pick the optimal order of comparisons, multiple parametes need to be
considered, both local and global ones. Intuitively, it's better to
leave more expensive comparions (for complex data types etc.) at the
very end, because maybe one of the earlier comparisons will be enough to
make a decision, and the expensive comparison will be unnecessary. The
other local parameter is cardinality - the higher the cardinality of a
key, the lower the probability we'll need to compare additional keys. We
explore possible orderings by a combination of exhaustive and greedy
approach, and pick the cheapest ordering.
But there are global parameters too, in the sense that we can't fully
evaluate them at the point when creating the sort. For example, there
might be an ORDER BY clause, or another ordering-dependent operation,
higher up in the query plan. The path may be already sorted by (some
of) the keys, in which case we can either use incremental sort or even
skip the sort entirely. So we consider various additional pathkeys in
addition to the "cheapers comparion" order.
The original patch was proposed by Teodor Sigaev, and later reworked and
improved by Dmitry Dolgov. Reviews by a number of people, including me,
Andrey Lepikhov, Claudio Freire and Ibrar Ahmed.
Author: Dmitry Dolgov, Teodor Sigaev
Reviewed-by: Tomas Vondra, Andrey Lepikhov, Claudio Freire, Ibrar Ahmed
Discussion: https://postgr.es/m/7c79e6a5-8597-74e8-0671-1c39d124c9d6%40sigaev.ru
Discussion: https://postgr.es/m/CA%2Bq6zcW_4o2NC0zutLkOJPsFt80megSpX_dVRo6GK9PC-Jx_Ag%40mail.gmail.com
---
.../postgres_fdw/expected/postgres_fdw.out | 15 +-
src/backend/optimizer/path/costsize.c | 369 +++++++++-
src/backend/optimizer/path/equivclass.c | 13 +-
src/backend/optimizer/path/pathkeys.c | 580 ++++++++++++++++
src/backend/optimizer/plan/planner.c | 653 ++++++++++--------
src/backend/optimizer/util/pathnode.c | 2 +-
src/backend/utils/adt/selfuncs.c | 37 +-
src/backend/utils/misc/guc.c | 32 +
src/include/nodes/nodes.h | 1 +
src/include/nodes/pathnodes.h | 10 +
src/include/optimizer/cost.h | 4 +-
src/include/optimizer/paths.h | 11 +
src/include/utils/selfuncs.h | 5 +
src/test/regress/expected/aggregates.out | 244 ++++++-
src/test/regress/expected/guc.out | 9 +-
.../regress/expected/incremental_sort.out | 2 +-
src/test/regress/expected/join.out | 51 +-
.../regress/expected/partition_aggregate.out | 136 ++--
src/test/regress/expected/partition_join.out | 75 +-
src/test/regress/expected/union.out | 60 +-
src/test/regress/sql/aggregates.sql | 99 +++
src/test/regress/sql/incremental_sort.sql | 2 +-
22 files changed, 1913 insertions(+), 497 deletions(-)
diff --git a/contrib/postgres_fdw/expected/postgres_fdw.out b/contrib/postgres_fdw/expected/postgres_fdw.out
index f210f911880..2ffc836824a 100644
--- a/contrib/postgres_fdw/expected/postgres_fdw.out
+++ b/contrib/postgres_fdw/expected/postgres_fdw.out
@@ -2741,16 +2741,13 @@ select c2 * (random() <= 1)::int as c2 from ft2 group by c2 * (random() <= 1)::i
-- GROUP BY clause in various forms, cardinal, alias and constant expression
explain (verbose, costs off)
select count(c2) w, c2 x, 5 y, 7.0 z from ft1 group by 2, y, 9.0::int order by 2;
- QUERY PLAN
----------------------------------------------------------------------------------------
- Sort
+ QUERY PLAN
+------------------------------------------------------------------------------------------------------------
+ Foreign Scan
Output: (count(c2)), c2, 5, 7.0, 9
- Sort Key: ft1.c2
- -> Foreign Scan
- Output: (count(c2)), c2, 5, 7.0, 9
- Relations: Aggregate on (public.ft1)
- Remote SQL: SELECT count(c2), c2, 5, 7.0, 9 FROM "S 1"."T 1" GROUP BY 2, 3, 5
-(7 rows)
+ Relations: Aggregate on (public.ft1)
+ Remote SQL: SELECT count(c2), c2, 5, 7.0, 9 FROM "S 1"."T 1" GROUP BY 2, 3, 5 ORDER BY c2 ASC NULLS LAST
+(4 rows)
select count(c2) w, c2 x, 5 y, 7.0 z from ft1 group by 2, y, 9.0::int order by 2;
w | x | y | z
diff --git a/src/backend/optimizer/path/costsize.c b/src/backend/optimizer/path/costsize.c
index 8dc7dd4ca26..68a32740d7e 100644
--- a/src/backend/optimizer/path/costsize.c
+++ b/src/backend/optimizer/path/costsize.c
@@ -1754,6 +1754,325 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
rterm->pathtarget->width);
}
+/*
+ * is_fake_var
+ * Workaround for generate_append_tlist() which generates fake Vars with
+ * varno == 0, that will cause a fail of estimate_num_group() call
+ *
+ * XXX Ummm, why would estimate_num_group fail with this?
+ */
+static bool
+is_fake_var(Expr *expr)
+{
+ if (IsA(expr, RelabelType))
+ expr = (Expr *) ((RelabelType *) expr)->arg;
+
+ return (IsA(expr, Var) && ((Var *) expr)->varno == 0);
+}
+
+/*
+ * get_width_cost_multiplier
+ * Returns relative complexity of comparing two values based on it's width.
+ * The idea behind - long values have more expensive comparison. Return value is
+ * in cpu_operator_cost unit.
+ */
+static double
+get_width_cost_multiplier(PlannerInfo *root, Expr *expr)
+{
+ double width = -1.0; /* fake value */
+
+ if (IsA(expr, RelabelType))
+ expr = (Expr *) ((RelabelType *) expr)->arg;
+
+ /* Try to find actual stat in corresponding relation */
+ if (IsA(expr, Var))
+ {
+ Var *var = (Var *) expr;
+
+ if (var->varno > 0 && var->varno < root->simple_rel_array_size)
+ {
+ RelOptInfo *rel = root->simple_rel_array[var->varno];
+
+ if (rel != NULL &&
+ var->varattno >= rel->min_attr &&
+ var->varattno <= rel->max_attr)
+ {
+ int ndx = var->varattno - rel->min_attr;
+
+ if (rel->attr_widths[ndx] > 0)
+ width = rel->attr_widths[ndx];
+ }
+ }
+ }
+
+ /* Didn't find any actual stats, use estimation by type */
+ if (width < 0.0)
+ {
+ Node *node = (Node*) expr;
+
+ width = get_typavgwidth(exprType(node), exprTypmod(node));
+ }
+
+ /*
+ * Any value in pgsql is passed by Datum type, so any operation with value
+ * could not be cheaper than operation with Datum type
+ */
+ if (width <= sizeof(Datum))
+ return 1.0;
+
+ /*
+ * Seems, cost of comparision is not directly proportional to args width,
+ * because comparing args could be differ width (we known only average over
+ * column) and difference often could be defined only by looking on first
+ * bytes. So, use log16(width) as estimation.
+ */
+ return 1.0 + 0.125 * LOG2(width / sizeof(Datum));
+}
+
+/*
+ * compute_cpu_sort_cost
+ * compute CPU cost of sort (i.e. in-memory)
+ *
+ * The main thing we need to calculate to estimate sort CPU costs is the number
+ * of calls to the comparator functions. The difficulty is that for multi-column
+ * sorts there may be different data types involved (for some of which the calls
+ * may be much more expensive). Furthermore, the columns may have very different
+ * number of distinct values - the higher the number, the fewer comparisons will
+ * be needed for the following columns.
+ *
+ * The algoritm is incremental - we add pathkeys one by one, and at each step we
+ * estimate the number of necessary comparisons (based on the number of distinct
+ * groups in the current pathkey prefix and the new pathkey), and the comparison
+ * costs (which is data type specific).
+ *
+ * Estimation of the number of comparisons is based on ideas from:
+ *
+ * "Quicksort Is Optimal", Robert Sedgewick, Jon Bentley, 2002
+ * [https://www.cs.princeton.edu/~rs/talks/QuicksortIsOptimal.pdf]
+ *
+ * In term of that paper, let N - number of tuples, Xi - number of identical
+ * tuples with value Ki, then the estimate of number of comparisons is:
+ *
+ * log(N! / (X1! * X2! * ..)) ~ sum(Xi * log(N/Xi))
+ *
+ * We assume all Xi the same because now we don't have any estimation of
+ * group sizes, we have only know the estimate of number of groups (distinct
+ * values). In that case, formula becomes:
+ *
+ * N * log(NumberOfGroups)
+ *
+ * For multi-column sorts we need to estimate the number of comparisons for
+ * each individual column - for example with columns (c1, c2, ..., ck) we
+ * can estimate that number of comparisons on ck is roughly
+ *
+ * ncomparisons(c1, c2, ..., ck) / ncomparisons(c1, c2, ..., c(k-1))
+ *
+ * Let k be a column number, Gk - number of groups defined by k columns, and Fk
+ * the cost of the comparison is
+ *
+ * N * sum( Fk * log(Gk) )
+ *
+ * Note: We also consider column width, not just the comparator cost.
+ *
+ * NOTE: some callers currently pass NIL for pathkeys because they
+ * can't conveniently supply the sort keys. In this case, it will fallback to
+ * simple comparison cost estimate.
+ */
+static Cost
+compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
+ Cost comparison_cost, double tuples, double output_tuples,
+ bool heapSort)
+{
+ Cost per_tuple_cost = 0.0;
+ ListCell *lc;
+ List *pathkeyExprs = NIL;
+ double tuplesPerPrevGroup = tuples;
+ double totalFuncCost = 1.0;
+ bool has_fake_var = false;
+ int i = 0;
+ Oid prev_datatype = InvalidOid;
+ Cost funcCost = 0.0;
+ List *cache_varinfos = NIL;
+
+ /* fallback if pathkeys is unknown */
+ if (list_length(pathkeys) == 0)
+ {
+ /*
+ * If we'll use a bounded heap-sort keeping just K tuples in memory, for
+ * a total number of tuple comparisons of N log2 K; but the constant
+ * factor is a bit higher than for quicksort. Tweak it so that the
+ * cost curve is continuous at the crossover point.
+ *
+ * XXX I suppose the "quicksort factor" references to 1.5 at the end
+ * of this function, but I'm not sure. I suggest we introduce some simple
+ * constants for that, instead of magic values.
+ */
+ output_tuples = (heapSort) ? 2.0 * output_tuples : tuples;
+ per_tuple_cost += 2.0 * cpu_operator_cost * LOG2(output_tuples);
+
+ /* add cost provided by caller */
+ per_tuple_cost += comparison_cost;
+
+ return per_tuple_cost * tuples;
+ }
+
+ /*
+ * Computing total cost of sorting takes into account:
+ * - per column comparison function cost
+ * - we try to compute needed number of comparison per column
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey*) lfirst(lc);
+ EquivalenceMember *em;
+ double nGroups,
+ correctedNGroups;
+
+ /*
+ * We believe that equivalence members aren't very different, so, to
+ * estimate cost we take just first member
+ */
+ em = (EquivalenceMember *) linitial(pathkey->pk_eclass->ec_members);
+
+ if (em->em_datatype != InvalidOid)
+ {
+ /* do not lookup funcCost if data type is the same as previous */
+ if (prev_datatype != em->em_datatype)
+ {
+ Oid sortop;
+ QualCost cost;
+
+ sortop = get_opfamily_member(pathkey->pk_opfamily,
+ em->em_datatype, em->em_datatype,
+ pathkey->pk_strategy);
+
+ cost.startup = 0;
+ cost.per_tuple = 0;
+ add_function_cost(root, get_opcode(sortop), NULL, &cost);
+ /* we need procost, not product of procost and cpu_operator_cost */
+ funcCost = cost.per_tuple / cpu_operator_cost;
+ prev_datatype = em->em_datatype;
+ }
+ }
+ else
+ funcCost = 1.0; /* fallback */
+
+ /* Try to take into account actual width fee */
+ funcCost *= get_width_cost_multiplier(root, em->em_expr);
+
+ totalFuncCost += funcCost;
+
+ /* Remember if we have a fake var in pathkeys */
+ has_fake_var |= is_fake_var(em->em_expr);
+ pathkeyExprs = lappend(pathkeyExprs, em->em_expr);
+
+ /*
+ * Prevent call estimate_num_groups() with fake Var. Note,
+ * pathkeyExprs contains only previous columns
+ */
+ if (has_fake_var == false)
+ /*
+ * Recursively compute number of groups in a group from previous step
+ */
+ nGroups = estimate_num_groups_incremental(root, pathkeyExprs,
+ tuplesPerPrevGroup, NULL, NULL,
+ &cache_varinfos,
+ list_length(pathkeyExprs) - 1);
+ else if (tuples > 4.0)
+ /*
+ * Use geometric mean as estimation if there is no any stats.
+ * Don't use DEFAULT_NUM_DISTINCT because it used for only one
+ * column while here we try to estimate number of groups over
+ * set of columns.
+ *
+ * XXX Perhaps this should use DEFAULT_NUM_DISTINCT at least to
+ * limit the calculated values, somehow?
+ *
+ * XXX What's the logic of the following formula?
+ */
+ nGroups = ceil(2.0 + sqrt(tuples) * (i + 1) / list_length(pathkeys));
+ else
+ nGroups = tuples;
+
+ /*
+ * Presorted keys aren't participated in comparison but still checked
+ * by qsort comparator.
+ */
+ if (i >= nPresortedKeys)
+ {
+ if (heapSort)
+ {
+ double heap_tuples;
+
+ /* have to keep at least one group, and a multiple of group size */
+ heap_tuples = Max(ceil(output_tuples / tuplesPerPrevGroup) * tuplesPerPrevGroup,
+ tuplesPerPrevGroup);
+
+ /* so how many (whole) groups is that? */
+ correctedNGroups = ceil(heap_tuples / tuplesPerPrevGroup);
+ }
+ else
+ /* all groups in the input */
+ correctedNGroups = nGroups;
+
+ correctedNGroups = Max(1.0, ceil(correctedNGroups));
+
+ per_tuple_cost += totalFuncCost * LOG2(correctedNGroups);
+ }
+
+ i++;
+
+ /*
+ * Real-world distribution isn't uniform but now we don't have a way to
+ * determine that, so, add multiplier to get closer to worst case.
+ * But ensure the number of tuples does not exceed the group size in the
+ * preceding step.
+ */
+ tuplesPerPrevGroup = Min(tuplesPerPrevGroup,
+ ceil(1.5 * tuplesPerPrevGroup / nGroups));
+
+ /*
+ * We could skip all following columns for cost estimation, because we
+ * believe that tuples are unique by the set of previous columns
+ */
+ if (tuplesPerPrevGroup <= 1.0)
+ break;
+ }
+
+ list_free(pathkeyExprs);
+
+ /* per_tuple_cost is in cpu_operator_cost units */
+ per_tuple_cost *= cpu_operator_cost;
+
+ /*
+ * Accordingly to "Introduction to algorithms", Thomas H. Cormen, Charles E.
+ * Leiserson, Ronald L. Rivest, ISBN 0-07-013143-0, quicksort estimation
+ * formula has additional term proportional to number of tuples (See Chapter
+ * 8.2 and Theorem 4.1). It has meaning with low number of tuples,
+ * approximately less that 1e4. Of course, it could be implemented as
+ * additional multiplier under logarithm, but use more complicated formula
+ * which takes into account number of unique tuples and it isn't clear how
+ * to combine multiplier with groups. Estimate it as 10 in cpu_operator_cost
+ * unit.
+ */
+ per_tuple_cost += 10 * cpu_operator_cost;
+
+ per_tuple_cost += comparison_cost;
+
+ return tuples * per_tuple_cost;
+}
+
+/*
+ * simple wrapper just to estimate best sort path
+ */
+Cost
+cost_sort_estimate(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
+ double tuples)
+{
+ return compute_cpu_sort_cost(root, pathkeys, nPresortedKeys,
+ 0, tuples, tuples, false);
+}
+
/*
* cost_tuplesort
* Determines and returns the cost of sorting a relation using tuplesort,
@@ -1770,7 +2089,7 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
* number of initial runs formed and M is the merge order used by tuplesort.c.
* Since the average initial run should be about sort_mem, we have
* disk traffic = 2 * relsize * ceil(logM(p / sort_mem))
- * cpu = comparison_cost * t * log2(t)
+ * and cpu cost (computed by compute_cpu_sort_cost()).
*
* If the sort is bounded (i.e., only the first k result tuples are needed)
* and k tuples can fit into sort_mem, we use a heap method that keeps only
@@ -1789,9 +2108,11 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
* 'comparison_cost' is the extra cost per comparison, if any
* 'sort_mem' is the number of kilobytes of work memory allowed for the sort
* 'limit_tuples' is the bound on the number of output tuples; -1 if no bound
+ * 'startup_cost' is expected to be 0 at input. If there is "input cost" it should
+ * be added by caller later.
*/
static void
-cost_tuplesort(Cost *startup_cost, Cost *run_cost,
+cost_tuplesort(PlannerInfo *root, List *pathkeys, Cost *startup_cost, Cost *run_cost,
double tuples, int width,
Cost comparison_cost, int sort_mem,
double limit_tuples)
@@ -1808,9 +2129,6 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
if (tuples < 2.0)
tuples = 2.0;
- /* Include the default cost-per-comparison */
- comparison_cost += 2.0 * cpu_operator_cost;
-
/* Do we have a useful LIMIT? */
if (limit_tuples > 0 && limit_tuples < tuples)
{
@@ -1834,12 +2152,10 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
double log_runs;
double npageaccesses;
- /*
- * CPU costs
- *
- * Assume about N log2 N comparisons
- */
- *startup_cost = comparison_cost * tuples * LOG2(tuples);
+ /* CPU costs */
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ tuples, false);
/* Disk costs */
@@ -1855,18 +2171,17 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
}
else if (tuples > 2 * output_tuples || input_bytes > sort_mem_bytes)
{
- /*
- * We'll use a bounded heap-sort keeping just K tuples in memory, for
- * a total number of tuple comparisons of N log2 K; but the constant
- * factor is a bit higher than for quicksort. Tweak it so that the
- * cost curve is continuous at the crossover point.
- */
- *startup_cost = comparison_cost * tuples * LOG2(2.0 * output_tuples);
+ /* We'll use a bounded heap-sort keeping just K tuples in memory. */
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ output_tuples, true);
}
else
{
/* We'll use plain quicksort on all the input tuples */
- *startup_cost = comparison_cost * tuples * LOG2(tuples);
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ tuples, false);
}
/*
@@ -1899,8 +2214,8 @@ cost_incremental_sort(Path *path,
double input_tuples, int width, Cost comparison_cost, int sort_mem,
double limit_tuples)
{
- Cost startup_cost = 0,
- run_cost = 0,
+ Cost startup_cost,
+ run_cost,
input_run_cost = input_total_cost - input_startup_cost;
double group_tuples,
input_groups;
@@ -1985,7 +2300,7 @@ cost_incremental_sort(Path *path,
* pessimistic about incremental sort performance and increase its average
* group size by half.
*/
- cost_tuplesort(&group_startup_cost, &group_run_cost,
+ cost_tuplesort(root, pathkeys, &group_startup_cost, &group_run_cost,
1.5 * group_tuples, width, comparison_cost, sort_mem,
limit_tuples);
@@ -1993,7 +2308,7 @@ cost_incremental_sort(Path *path,
* Startup cost of incremental sort is the startup cost of its first group
* plus the cost of its input.
*/
- startup_cost += group_startup_cost
+ startup_cost = group_startup_cost
+ input_startup_cost + group_input_run_cost;
/*
@@ -2002,7 +2317,7 @@ cost_incremental_sort(Path *path,
* group, plus the total cost to process the remaining groups, plus the
* remaining cost of input.
*/
- run_cost += group_run_cost
+ run_cost = group_run_cost
+ (group_run_cost + group_startup_cost) * (input_groups - 1)
+ group_input_run_cost * (input_groups - 1);
@@ -2042,7 +2357,7 @@ cost_sort(Path *path, PlannerInfo *root,
Cost startup_cost;
Cost run_cost;
- cost_tuplesort(&startup_cost, &run_cost,
+ cost_tuplesort(root, pathkeys, &startup_cost, &run_cost,
tuples, width,
comparison_cost, sort_mem,
limit_tuples);
@@ -2140,7 +2455,7 @@ append_nonpartial_cost(List *subpaths, int numpaths, int parallel_workers)
* Determines and returns the cost of an Append node.
*/
void
-cost_append(AppendPath *apath)
+cost_append(AppendPath *apath, PlannerInfo *root)
{
ListCell *l;
@@ -2208,7 +2523,7 @@ cost_append(AppendPath *apath)
* any child.
*/
cost_sort(&sort_path,
- NULL, /* doesn't currently need root */
+ root,
pathkeys,
subpath->total_cost,
subpath->rows,
diff --git a/src/backend/optimizer/path/equivclass.c b/src/backend/optimizer/path/equivclass.c
index 8c6770de972..5487ae2ee4c 100644
--- a/src/backend/optimizer/path/equivclass.c
+++ b/src/backend/optimizer/path/equivclass.c
@@ -681,7 +681,18 @@ get_eclass_for_sort_expr(PlannerInfo *root,
if (opcintype == cur_em->em_datatype &&
equal(expr, cur_em->em_expr))
- return cur_ec; /* Match! */
+ {
+ /*
+ * Match!
+ *
+ * Copy sortref if it wasn't set yet, it's possible if ec was
+ * constructed from WHERE clause, ie it doesn't have target
+ * reference at all
+ */
+ if (cur_ec->ec_sortref == 0 && sortref > 0)
+ cur_ec->ec_sortref = sortref;
+ return cur_ec;
+ }
}
}
diff --git a/src/backend/optimizer/path/pathkeys.c b/src/backend/optimizer/path/pathkeys.c
index 86a35cdef17..f18d4c8e4e7 100644
--- a/src/backend/optimizer/path/pathkeys.c
+++ b/src/backend/optimizer/path/pathkeys.c
@@ -17,16 +17,19 @@
*/
#include "postgres.h"
+#include "miscadmin.h"
#include "access/stratnum.h"
#include "catalog/pg_opfamily.h"
#include "nodes/makefuncs.h"
#include "nodes/nodeFuncs.h"
#include "nodes/plannodes.h"
+#include "optimizer/cost.h"
#include "optimizer/optimizer.h"
#include "optimizer/pathnode.h"
#include "optimizer/paths.h"
#include "partitioning/partbounds.h"
#include "utils/lsyscache.h"
+#include "utils/selfuncs.h"
static bool pathkey_is_redundant(PathKey *new_pathkey, List *pathkeys);
@@ -334,6 +337,530 @@ pathkeys_contained_in(List *keys1, List *keys2)
return false;
}
+/************************<DEBUG PART>*************************************/
+bool debug_group_by_reorder_by_pathkeys = true;
+bool debug_group_by_match_order_by = true;
+bool debug_cheapest_group_by = true;
+/************************</DEBUG PART>************************************/
+
+/*
+ * group_keys_reorder_by_pathkeys
+ * Reorder GROUP BY keys to match pathkeys of input path.
+ *
+ * Function returns new lists (pathkeys and clauses), original GROUP BY lists
+ * stay untouched.
+ *
+ * Returns the number of GROUP BY keys with a matching pathkey.
+ */
+int
+group_keys_reorder_by_pathkeys(List *pathkeys, List **group_pathkeys,
+ List **group_clauses)
+{
+ List *new_group_pathkeys= NIL,
+ *new_group_clauses = NIL;
+ ListCell *lc;
+ int n;
+
+ if (debug_group_by_reorder_by_pathkeys == false)
+ return 0;
+
+ if (pathkeys == NIL || *group_pathkeys == NIL)
+ return 0;
+
+ /*
+ * Walk the pathkeys (determining ordering of the input path) and see if
+ * there's a matching GROUP BY key. If we find one, we append if to the
+ * list, and do the same for the clauses.
+ *
+ * Once we find first pathkey without a matching GROUP BY key, the rest of
+ * the pathkeys is useless and can't be used to evaluate the grouping, so
+ * we abort the loop and ignore the remaining pathkeys.
+ *
+ * XXX Pathkeys are built in a way to allow simply comparing pointers.
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(lc);
+ SortGroupClause *sgc;
+
+ /* abort on first mismatch */
+ if (!list_member_ptr(*group_pathkeys, pathkey))
+ break;
+
+ new_group_pathkeys = lappend(new_group_pathkeys, pathkey);
+
+ sgc = get_sortgroupref_clause(pathkey->pk_eclass->ec_sortref,
+ *group_clauses);
+
+ new_group_clauses = lappend(new_group_clauses, sgc);
+ }
+
+ /* remember the number of pathkeys with a matching GROUP BY key */
+ n = list_length(new_group_pathkeys);
+
+ /* XXX maybe return when (n == 0) */
+
+ /* append the remaining group pathkeys (will be treated as not sorted) */
+ *group_pathkeys = list_concat_unique_ptr(new_group_pathkeys,
+ *group_pathkeys);
+ *group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+
+ return n;
+}
+
+/*
+ * Used to generate all permutations of a pathkey list.
+ */
+typedef struct PathkeyMutatorState {
+ List *elemsList;
+ ListCell **elemCells;
+ void **elems;
+ int *positions;
+ int mutatorNColumns;
+ int count;
+} PathkeyMutatorState;
+
+
+/*
+ * PathkeyMutatorInit
+ * Initialize state of the permutation generator.
+ *
+ * We want to generate permutations of elements in the "elems" list. We may want
+ * to skip some number of elements at the beginning (when treating as presorted)
+ * or at the end (we only permute a limited number of group keys).
+ *
+ * The list is decomposed into elements, and we also keep pointers to individual
+ * cells. This allows us to build the permuted list quickly and cheaply, without
+ * creating any copies.
+ */
+static void
+PathkeyMutatorInit(PathkeyMutatorState *state, List *elems, int start, int end)
+{
+ int i;
+ int n = end - start;
+ ListCell *lc;
+
+ memset(state, 0, sizeof(*state));
+
+ state->mutatorNColumns = n;
+
+ state->elemsList = list_copy(elems);
+
+ state->elems = palloc(sizeof(void*) * n);
+ state->elemCells = palloc(sizeof(ListCell*) * n);
+ state->positions = palloc(sizeof(int) * n);
+
+ i = 0;
+ for_each_cell(lc, state->elemsList, list_nth_cell(state->elemsList, start))
+ {
+ state->elemCells[i] = lc;
+ state->elems[i] = lfirst(lc);
+ state->positions[i] = i + 1;
+ i++;
+
+ if (i >= n)
+ break;
+ }
+}
+
+/* Swap two elements of an array. */
+static void
+PathkeyMutatorSwap(int *a, int i, int j)
+{
+ int s = a[i];
+
+ a[i] = a[j];
+ a[j] = s;
+}
+
+/*
+ * Generate the next permutation of elements.
+ */
+static bool
+PathkeyMutatorNextSet(int *a, int n)
+{
+ int j, k, l, r;
+
+ j = n - 2;
+
+ while (j >= 0 && a[j] >= a[j + 1])
+ j--;
+
+ if (j < 0)
+ return false;
+
+ k = n - 1;
+
+ while (k >= 0 && a[j] >= a[k])
+ k--;
+
+ PathkeyMutatorSwap(a, j, k);
+
+ l = j + 1;
+ r = n - 1;
+
+ while (l < r)
+ PathkeyMutatorSwap(a, l++, r--);
+
+ return true;
+}
+
+/*
+ * PathkeyMutatorNext
+ * Generate the next permutation of list of elements.
+ *
+ * Returns the next permutation (as a list of elements) or NIL if there are no
+ * more permutations.
+ */
+static List *
+PathkeyMutatorNext(PathkeyMutatorState *state)
+{
+ int i;
+
+ state->count++;
+
+ /* first permutation is original list */
+ if (state->count == 1)
+ return state->elemsList;
+
+ /* when there are no more permutations, return NIL */
+ if (!PathkeyMutatorNextSet(state->positions, state->mutatorNColumns))
+ {
+ pfree(state->elems);
+ pfree(state->elemCells);
+ pfree(state->positions);
+
+ list_free(state->elemsList);
+
+ return NIL;
+ }
+
+ /* update the list cells to point to the right elements */
+ for(i=0; i<state->mutatorNColumns; i++)
+ lfirst(state->elemCells[i]) =
+ (void *) state->elems[ state->positions[i] - 1 ];
+
+ return state->elemsList;
+}
+
+/*
+ * Cost of comparing pathkeys.
+ */
+typedef struct PathkeySortCost
+{
+ Cost cost;
+ PathKey *pathkey;
+} PathkeySortCost;
+
+static int
+pathkey_sort_cost_comparator(const void *_a, const void *_b)
+{
+ const PathkeySortCost *a = (PathkeySortCost *) _a;
+ const PathkeySortCost *b = (PathkeySortCost *) _b;
+
+ if (a->cost < b->cost)
+ return -1;
+ else if (a->cost == b->cost)
+ return 0;
+ return 1;
+}
+
+/*
+ * get_cheapest_group_keys_order
+ * Returns the pathkeys in an order cheapest to evaluate.
+ *
+ * Given a list of pathkeys, we try to reorder them in a way that minimizes
+ * the CPU cost of sorting. This depends mainly on the cost of comparator
+ * function (the pathkeys may use different data types) and the number of
+ * distinct values in each column (which affects the number of comparator
+ * calls for the following pathkeys).
+ *
+ * In case the input is partially sorted, only the remaining pathkeys are
+ * considered.
+ *
+ * Returns newly allocated lists. If no reordering is possible (or needed),
+ * the lists are set to NIL.
+ */
+static bool
+get_cheapest_group_keys_order(PlannerInfo *root, double nrows,
+ List **group_pathkeys, List **group_clauses,
+ int n_preordered)
+{
+ List *new_group_pathkeys = NIL,
+ *new_group_clauses = NIL,
+ *var_group_pathkeys;
+
+ ListCell *cell;
+ PathkeyMutatorState mstate;
+ double cheapest_sort_cost = -1.0;
+
+ int nFreeKeys;
+ int nToPermute;
+
+ /* If this optimization is disabled, we're done. */
+ if (!debug_cheapest_group_by)
+ return false;
+
+ /* If there are less than 2 unsorted pathkeys, we're done. */
+ if (list_length(*group_pathkeys) - n_preordered < 2)
+ return false;
+
+ /*
+ * We could exhaustively cost all possible orderings of the pathkeys, but for
+ * large number of pathkeys that might be prohibitively expensive. So we try
+ * to apply a simple cheap heuristics first - we sort the pathkeys by sort
+ * cost (as if the pathkey was sorted independently) and then check only the
+ * four cheapest pathkeys. The remaining pathkeys are kept ordered by cost.
+ *
+ * XXX This is a very simple heuristics, and likely to work fine for most
+ * cases (because number of GROUP BY clauses tends to be lower than 4). But
+ * it ignores how the number of distinct values in each pathkey affects the
+ * following sorts. It may be better to use "more expensive" pathkey first
+ * if it has many distinct values, because it then limits the number of
+ * comparisons for the remaining pathkeys. But evaluating that is kinda the
+ * expensive bit we're trying to not do.
+ */
+ nFreeKeys = list_length(*group_pathkeys) - n_preordered;
+ nToPermute = 4;
+ if (nFreeKeys > nToPermute)
+ {
+ int i;
+ PathkeySortCost *costs = palloc(sizeof(PathkeySortCost) * nFreeKeys);
+
+ /* skip the pre-ordered pathkeys */
+ cell = list_nth_cell(*group_pathkeys, n_preordered);
+
+ /* estimate cost for sorting individual pathkeys */
+ for (i = 0; cell != NULL; i++, (cell = lnext(*group_pathkeys, cell)))
+ {
+ List *to_cost = list_make1(lfirst(cell));
+
+ Assert(i < nFreeKeys);
+
+ costs[i].pathkey = lfirst(cell);
+ costs[i].cost = cost_sort_estimate(root, to_cost, 0, nrows);
+
+ pfree(to_cost);
+ }
+
+ /* sort the pathkeys by sort cost in ascending order */
+ qsort(costs, nFreeKeys, sizeof(*costs), pathkey_sort_cost_comparator);
+
+ /*
+ * Rebuild the list of pathkeys - first the preordered ones, then the
+ * rest ordered by cost.
+ */
+ new_group_pathkeys = list_truncate(list_copy(*group_pathkeys), n_preordered);
+
+ for (i = 0; i < nFreeKeys; i++)
+ new_group_pathkeys = lappend(new_group_pathkeys, costs[i].pathkey);
+
+ pfree(costs);
+ }
+ else
+ {
+ /*
+ * Since v13 list_free() can clean list elements so for original list
+ * not to be modified it should be copied to a new one which can then
+ * be cleaned safely if needed.
+ */
+ new_group_pathkeys = list_copy(*group_pathkeys);
+ nToPermute = nFreeKeys;
+ }
+
+ Assert(list_length(new_group_pathkeys) == list_length(*group_pathkeys));
+
+ /*
+ * Generate pathkey lists with permutations of the first nToPermute pathkeys.
+ *
+ * XXX We simply calculate sort cost for each individual pathkey list, but
+ * there's room for two dynamic programming optimizations here. Firstly, we
+ * may pass the current "best" cost to cost_sort_estimate so that it can
+ * "abort" if the estimated pathkeys list exceeds it. Secondly, it could pass
+ * return information about the position when it exceeded the cost, and we
+ * could skip all permutations with the same prefix.
+ *
+ * Imagine we've already found ordering with cost C1, and we're evaluating
+ * another ordering - cost_sort_estimate() calculates cost by adding the
+ * pathkeys one by one (more or less), and the cost only grows. If at any
+ * point it exceeds C1, it can't possibly be "better" so we can discard it.
+ * But we also know that we can discard all ordering with the same prefix,
+ * because if we're estimating (a,b,c,d) and we exceeded C1 at (a,b) then
+ * the same thing will happen for any ordering with this prefix.
+ *
+ *
+ */
+ PathkeyMutatorInit(&mstate, new_group_pathkeys, n_preordered, n_preordered + nToPermute);
+
+ while((var_group_pathkeys = PathkeyMutatorNext(&mstate)) != NIL)
+ {
+ Cost cost;
+
+ cost = cost_sort_estimate(root, var_group_pathkeys, n_preordered, nrows);
+
+ if (cost < cheapest_sort_cost || cheapest_sort_cost < 0)
+ {
+ list_free(new_group_pathkeys);
+ new_group_pathkeys = list_copy(var_group_pathkeys);
+ cheapest_sort_cost = cost;
+ }
+ }
+
+ /* Reorder the group clauses according to the reordered pathkeys. */
+ foreach(cell, new_group_pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(cell);
+
+ new_group_clauses = lappend(new_group_clauses,
+ get_sortgroupref_clause(pathkey->pk_eclass->ec_sortref,
+ *group_clauses));
+ }
+
+ /* Just append the rest GROUP BY clauses */
+ new_group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+
+ *group_pathkeys = new_group_pathkeys;
+ *group_clauses = new_group_clauses;
+
+ return true;
+}
+
+/*
+ * get_useful_group_keys_orderings
+ * Determine which orderings of GROUP BY keys are potentially interesting.
+ *
+ * Returns list of PathKeyInfo items, each representing an interesting ordering
+ * of GROUP BY keys. Each items stores pathkeys and clauses in matching order.
+ *
+ * The function considers (and keeps) multiple group by orderings:
+ *
+ * - the original ordering, as specified by the GROUP BY clause
+ *
+ * - GROUP BY keys reordered to minimize the sort cost
+ *
+ * - GROUP BY keys reordered to match path ordering (as much as possible), with
+ * the tail reoredered to minimize the sort cost
+ *
+ * - GROUP BY keys to match target ORDER BY clause (as much as possible), with
+ * the tail reoredered to minimize the sort cost
+ *
+ * There are other potentially interesting orderings (e.g. it might be best to
+ * match the first ORDER BY key, order the remaining keys differently and then
+ * rely on incremental sort to fix this), but we ignore those for now. To make
+ * this work we'd have to pretty much generate all possible permutations.
+ */
+List *
+get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys,
+ List *group_pathkeys, List *group_clauses)
+{
+ Query *parse = root->parse;
+ List *infos = NIL;
+ PathKeyInfo *info;
+ int n_preordered = 0;
+
+ List *pathkeys = group_pathkeys;
+ List *clauses = group_clauses;
+
+ /* always return at least the original pathkeys/clauses */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+
+ /* for grouping sets we can't do any reordering */
+ if (parse->groupingSets)
+ return infos;
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost, ignoring both the
+ * target ordering (ORDER BY) and ordering of the input path.
+ */
+ if (get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered))
+ {
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * If the path is sorted in some way, try reordering the group keys to match
+ * as much of the ordering as possible - we get this sort for free (mostly).
+ *
+ * We must not do this when there are no grouping sets, because those use
+ * more complex logic to decide the ordering.
+ *
+ * XXX Isn't this somewhat redundant with presorted_keys? Actually, it's
+ * more a complement, because it allows benefiting from incremental sort
+ * as much as possible.
+ *
+ * XXX This does nothing if (n_preordered == 0). We shouldn't create the
+ * info in this case.
+ */
+ if (path_pathkeys)
+ {
+ n_preordered = group_keys_reorder_by_pathkeys(path_pathkeys,
+ &pathkeys,
+ &clauses);
+
+ /* reorder the tail to minimize sort cost */
+ get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, in
+ * which case get_cheapest_group_keys_order returns false. But we
+ * still want to keep the keys reordered to path_pathkeys.
+ */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost (this time consider
+ * the ORDER BY clause, but only if set debug_group_by_match_order_by).
+ *
+ * XXX This does nothing if (n_preordered == 0). We shouldn't create the
+ * info in this case.
+ */
+ if (root->sort_pathkeys && debug_group_by_match_order_by)
+ {
+ n_preordered = group_keys_reorder_by_pathkeys(root->sort_pathkeys,
+ &pathkeys,
+ &clauses);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, in
+ * which case get_cheapest_group_keys_order returns false. But we
+ * still want to keep the keys reordered to sort_pathkeys.
+ */
+ get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered);
+
+ /* keep the group keys reordered to match ordering of input path */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ return infos;
+}
+
/*
* pathkeys_count_contained_in
* Same as pathkeys_contained_in, but also sets length of longest
@@ -1862,6 +2389,54 @@ pathkeys_useful_for_ordering(PlannerInfo *root, List *pathkeys)
return n_common_pathkeys;
}
+/*
+ * pathkeys_useful_for_grouping
+ * Count the number of pathkeys that are useful for grouping (instead of
+ * explicit sort)
+ *
+ * Group pathkeys could be reordered to benefit from the odering. The ordering
+ * may not be "complete" and may require incremental sort, but that's fine. So
+ * we simply count prefix pathkeys with a matching group key, and stop once we
+ * find the first pathkey without a match.
+ *
+ * So e.g. with pathkeys (a,b,c) and group keys (a,b,e) this determines (a,b)
+ * pathkeys are useful for grouping, and we might do incremental sort to get
+ * path ordered by (a,b,e).
+ *
+ * This logic is necessary to retain paths with ordeding not matching grouping
+ * keys directly, without the reordering.
+ *
+ * Returns the length of pathkey prefix with matching group keys.
+ */
+static int
+pathkeys_useful_for_grouping(PlannerInfo *root, List *pathkeys)
+{
+ ListCell *key;
+ int n = 0;
+
+ /* no special ordering requested for grouping */
+ if (root->group_pathkeys == NIL)
+ return 0;
+
+ /* unordered path */
+ if (pathkeys == NIL)
+ return 0;
+
+ /* walk the pathkeys and search for matching group key */
+ foreach(key, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(key);
+
+ /* no matching group key, we're done */
+ if (!list_member_ptr(root->group_pathkeys, pathkey))
+ break;
+
+ n++;
+ }
+
+ return n;
+}
+
/*
* truncate_useless_pathkeys
* Shorten the given pathkey list to just the useful pathkeys.
@@ -1876,6 +2451,9 @@ truncate_useless_pathkeys(PlannerInfo *root,
nuseful = pathkeys_useful_for_merging(root, rel, pathkeys);
nuseful2 = pathkeys_useful_for_ordering(root, pathkeys);
+ if (nuseful2 > nuseful)
+ nuseful = nuseful2;
+ nuseful2 = pathkeys_useful_for_grouping(root, pathkeys);
if (nuseful2 > nuseful)
nuseful = nuseful2;
@@ -1911,6 +2489,8 @@ has_useful_pathkeys(PlannerInfo *root, RelOptInfo *rel)
{
if (rel->joininfo != NIL || rel->has_eclass_joins)
return true; /* might be able to use pathkeys for merging */
+ if (root->group_pathkeys != NIL)
+ return true; /* might be able to use pathkeys for grouping */
if (root->query_pathkeys != NIL)
return true; /* might be able to use them for ordering */
return false; /* definitely useless */
diff --git a/src/backend/optimizer/plan/planner.c b/src/backend/optimizer/plan/planner.c
index bd09f85aea1..fac5822e5b5 100644
--- a/src/backend/optimizer/plan/planner.c
+++ b/src/backend/optimizer/plan/planner.c
@@ -6156,24 +6156,124 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
+ Path *path_save = path;
Path *path_original = path;
- bool is_sorted;
- int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ List *pathkey_orderings = NIL;
+
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
+
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
- if (path == cheapest_path || is_sorted)
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- /* Sort the cheapest-total path if it isn't already sorted */
- if (!is_sorted)
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
+
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (path == cheapest_path || is_sorted)
+ {
+ /* Sort the cheapest-total path if it isn't already sorted */
+ if (!is_sorted)
+ {
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ }
+
+ /* Now decide what to stick atop it */
+ if (parse->groupingSets)
+ {
+ consider_groupingsets_paths(root, grouped_rel,
+ path, true, can_hash,
+ gd, agg_costs, dNumGroups);
+ }
+ else if (parse->hasAggs)
+ {
+ /*
+ * We have aggregation, possibly with plain GROUP BY. Make
+ * an AggPath.
+ */
+ add_path(grouped_rel, (Path *)
+ create_agg_path(root,
+ grouped_rel,
+ path,
+ grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_SIMPLE,
+ info->clauses,
+ havingQual,
+ agg_costs,
+ dNumGroups));
+ }
+ else if (group_clauses)
+ {
+ /*
+ * We have GROUP BY without aggregation or grouping sets.
+ * Make a GroupPath.
+ */
+ add_path(grouped_rel, (Path *)
+ create_group_path(root,
+ grouped_rel,
+ path,
+ info->clauses,
+ havingQual,
+ dNumGroups));
+ }
+ else
+ {
+ /* Other cases should have been handled above */
+ Assert(false);
+ }
+ }
+
+ /*
+ * Now we may consider incremental sort on this path, but only
+ * when the path is not already sorted and when incremental sort
+ * is enabled.
+ */
+ if (is_sorted || !enable_incremental_sort)
+ continue;
+
+ /* Restore the input path (we might have added Sort on top). */
+ path = path_original;
+
+ /* no shared prefix, no point in building incremental sort */
+ if (presorted_keys == 0)
+ continue;
+
+ /*
+ * We should have already excluded pathkeys of length 1 because
+ * then presorted_keys > 0 would imply is_sorted was true.
+ */
+ Assert(list_length(root->group_pathkeys) != 1);
+
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
/* Now decide what to stick atop it */
if (parse->groupingSets)
@@ -6185,17 +6285,17 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
else if (parse->hasAggs)
{
/*
- * We have aggregation, possibly with plain GROUP BY. Make
- * an AggPath.
+ * We have aggregation, possibly with plain GROUP BY. Make an
+ * AggPath.
*/
add_path(grouped_rel, (Path *)
create_agg_path(root,
grouped_rel,
path,
grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_SIMPLE,
- parse->groupClause,
+ info->clauses,
havingQual,
agg_costs,
dNumGroups));
@@ -6203,14 +6303,14 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
else if (parse->groupClause)
{
/*
- * We have GROUP BY without aggregation or grouping sets.
- * Make a GroupPath.
+ * We have GROUP BY without aggregation or grouping sets. Make
+ * a GroupPath.
*/
add_path(grouped_rel, (Path *)
create_group_path(root,
grouped_rel,
path,
- parse->groupClause,
+ info->clauses,
havingQual,
dNumGroups));
}
@@ -6220,79 +6320,6 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
Assert(false);
}
}
-
- /*
- * Now we may consider incremental sort on this path, but only
- * when the path is not already sorted and when incremental sort
- * is enabled.
- */
- if (is_sorted || !enable_incremental_sort)
- continue;
-
- /* Restore the input path (we might have added Sort on top). */
- path = path_original;
-
- /* no shared prefix, no point in building incremental sort */
- if (presorted_keys == 0)
- continue;
-
- /*
- * We should have already excluded pathkeys of length 1 because
- * then presorted_keys > 0 would imply is_sorted was true.
- */
- Assert(list_length(root->group_pathkeys) != 1);
-
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
-
- /* Now decide what to stick atop it */
- if (parse->groupingSets)
- {
- consider_groupingsets_paths(root, grouped_rel,
- path, true, can_hash,
- gd, agg_costs, dNumGroups);
- }
- else if (parse->hasAggs)
- {
- /*
- * We have aggregation, possibly with plain GROUP BY. Make an
- * AggPath.
- */
- add_path(grouped_rel, (Path *)
- create_agg_path(root,
- grouped_rel,
- path,
- grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_SIMPLE,
- parse->groupClause,
- havingQual,
- agg_costs,
- dNumGroups));
- }
- else if (parse->groupClause)
- {
- /*
- * We have GROUP BY without aggregation or grouping sets. Make
- * a GroupPath.
- */
- add_path(grouped_rel, (Path *)
- create_group_path(root,
- grouped_rel,
- path,
- parse->groupClause,
- havingQual,
- dNumGroups));
- }
- else
- {
- /* Other cases should have been handled above */
- Assert(false);
- }
}
/*
@@ -6303,100 +6330,125 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
{
foreach(lc, partially_grouped_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
Path *path_original = path;
- bool is_sorted;
- int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ List *pathkey_orderings = NIL;
- /*
- * Insert a Sort node, if required. But there's no point in
- * sorting anything but the cheapest path.
- */
- if (!is_sorted)
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
+
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- if (path != partially_grouped_rel->cheapest_total_path)
- continue;
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- }
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- if (parse->hasAggs)
- add_path(grouped_rel, (Path *)
- create_agg_path(root,
- grouped_rel,
- path,
- grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_FINAL_DESERIAL,
- parse->groupClause,
- havingQual,
- agg_final_costs,
- dNumGroups));
- else
- add_path(grouped_rel, (Path *)
- create_group_path(root,
- grouped_rel,
- path,
- parse->groupClause,
- havingQual,
- dNumGroups));
+ /* restore the path (we replace it in the loop) */
+ path = path_original;
- /*
- * Now we may consider incremental sort on this path, but only
- * when the path is not already sorted and when incremental
- * sort is enabled.
- */
- if (is_sorted || !enable_incremental_sort)
- continue;
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
- /* Restore the input path (we might have added Sort on top). */
- path = path_original;
+ /*
+ * Insert a Sort node, if required. But there's no point in
+ * sorting anything but the cheapest path.
+ */
+ if (!is_sorted)
+ {
+ if (path != partially_grouped_rel->cheapest_total_path)
+ continue;
- /* no shared prefix, not point in building incremental sort */
- if (presorted_keys == 0)
- continue;
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ }
- /*
- * We should have already excluded pathkeys of length 1
- * because then presorted_keys > 0 would imply is_sorted was
- * true.
- */
- Assert(list_length(root->group_pathkeys) != 1);
+ if (parse->hasAggs)
+ add_path(grouped_rel, (Path *)
+ create_agg_path(root,
+ grouped_rel,
+ path,
+ grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_FINAL_DESERIAL,
+ info->clauses,
+ havingQual,
+ agg_final_costs,
+ dNumGroups));
+ else
+ add_path(grouped_rel, (Path *)
+ create_group_path(root,
+ grouped_rel,
+ path,
+ info->clauses,
+ havingQual,
+ dNumGroups));
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
+ /*
+ * Now we may consider incremental sort on this path, but only
+ * when the path is not already sorted and when incremental
+ * sort is enabled.
+ */
+ if (is_sorted || !enable_incremental_sort)
+ continue;
- if (parse->hasAggs)
- add_path(grouped_rel, (Path *)
- create_agg_path(root,
- grouped_rel,
- path,
- grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_FINAL_DESERIAL,
- parse->groupClause,
- havingQual,
- agg_final_costs,
- dNumGroups));
- else
- add_path(grouped_rel, (Path *)
- create_group_path(root,
- grouped_rel,
- path,
- parse->groupClause,
- havingQual,
- dNumGroups));
+ /* Restore the input path (we might have added Sort on top). */
+ path = path_original;
+
+ /* no shared prefix, not point in building incremental sort */
+ if (presorted_keys == 0)
+ continue;
+
+ /*
+ * We should have already excluded pathkeys of length 1
+ * because then presorted_keys > 0 would imply is_sorted was
+ * true.
+ */
+ Assert(list_length(root->group_pathkeys) != 1);
+
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+
+ if (parse->hasAggs)
+ add_path(grouped_rel, (Path *)
+ create_agg_path(root,
+ grouped_rel,
+ path,
+ grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_FINAL_DESERIAL,
+ info->clauses,
+ havingQual,
+ agg_final_costs,
+ dNumGroups));
+ else
+ add_path(grouped_rel, (Path *)
+ create_group_path(root,
+ grouped_rel,
+ path,
+ info->clauses,
+ havingQual,
+ dNumGroups));
+ }
}
}
}
@@ -6599,41 +6651,71 @@ create_partial_grouping_paths(PlannerInfo *root,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
+ Path *path_save = path;
+
+ List *pathkey_orderings = NIL;
- is_sorted = pathkeys_contained_in(root->group_pathkeys,
- path->pathkeys);
- if (path == cheapest_total_path || is_sorted)
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
+
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- /* Sort the cheapest partial path, if it isn't already */
- if (!is_sorted)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- if (parse->hasAggs)
- add_path(partially_grouped_rel, (Path *)
- create_agg_path(root,
- partially_grouped_rel,
- path,
- partially_grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_INITIAL_SERIAL,
- parse->groupClause,
- NIL,
- agg_partial_costs,
- dNumPartialGroups));
- else
- add_path(partially_grouped_rel, (Path *)
- create_group_path(root,
- partially_grouped_rel,
- path,
- parse->groupClause,
- NIL,
- dNumPartialGroups));
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
+
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (path == cheapest_total_path || is_sorted)
+ {
+ /* Sort the cheapest partial path, if it isn't already */
+ if (!is_sorted)
+ {
+ path = (Path *) create_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_path(partially_grouped_rel, (Path *)
+ create_agg_path(root,
+ partially_grouped_rel,
+ path,
+ partially_grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_INITIAL_SERIAL,
+ info->clauses,
+ NIL,
+ agg_partial_costs,
+ dNumPartialGroups));
+ else
+ add_path(partially_grouped_rel, (Path *)
+ create_group_path(root,
+ partially_grouped_rel,
+ path,
+ info->clauses,
+ NIL,
+ dNumPartialGroups));
+ }
}
}
@@ -6643,6 +6725,8 @@ create_partial_grouping_paths(PlannerInfo *root,
* We can also skip the entire loop when we only have a single-item
* group_pathkeys because then we can't possibly have a presorted
* prefix of the list without having the list be fully sorted.
+ *
+ * XXX Shouldn't this also consider the group-key-reordering?
*/
if (enable_incremental_sort && list_length(root->group_pathkeys) > 1)
{
@@ -6701,24 +6785,100 @@ create_partial_grouping_paths(PlannerInfo *root,
/* Similar to above logic, but for partial paths. */
foreach(lc, input_rel->partial_pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
Path *path_original = path;
- bool is_sorted;
- int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ List *pathkey_orderings = NIL;
+
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
- if (path == cheapest_partial_path || is_sorted)
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- /* Sort the cheapest partial path, if it isn't already */
- if (!is_sorted)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+ /* restore the path (we replace it in the loop) */
+ path = path_original;
+
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (path == cheapest_partial_path || is_sorted)
+ {
+
+ /* Sort the cheapest partial path, if it isn't already */
+ if (!is_sorted)
+ {
+ path = (Path *) create_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_partial_path(partially_grouped_rel, (Path *)
+ create_agg_path(root,
+ partially_grouped_rel,
+ path,
+ partially_grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_INITIAL_SERIAL,
+ info->clauses,
+ NIL,
+ agg_partial_costs,
+ dNumPartialPartialGroups));
+ else
+ add_partial_path(partially_grouped_rel, (Path *)
+ create_group_path(root,
+ partially_grouped_rel,
+ path,
+ info->clauses,
+ NIL,
+ dNumPartialPartialGroups));
+ }
+
+ /*
+ * Now we may consider incremental sort on this path, but only
+ * when the path is not already sorted and when incremental sort
+ * is enabled.
+ */
+ if (is_sorted || !enable_incremental_sort)
+ continue;
+
+ /* Restore the input path (we might have added Sort on top). */
+ path = path_original;
+
+ /* no shared prefix, not point in building incremental sort */
+ if (presorted_keys == 0)
+ continue;
+
+ /*
+ * We should have already excluded pathkeys of length 1 because
+ * then presorted_keys > 0 would imply is_sorted was true.
+ */
+ Assert(list_length(root->group_pathkeys) != 1);
+
+ path = (Path *) create_incremental_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
if (parse->hasAggs)
add_partial_path(partially_grouped_rel, (Path *)
@@ -6726,9 +6886,9 @@ create_partial_grouping_paths(PlannerInfo *root,
partially_grouped_rel,
path,
partially_grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_INITIAL_SERIAL,
- parse->groupClause,
+ info->clauses,
NIL,
agg_partial_costs,
dNumPartialPartialGroups));
@@ -6737,59 +6897,10 @@ create_partial_grouping_paths(PlannerInfo *root,
create_group_path(root,
partially_grouped_rel,
path,
- parse->groupClause,
+ info->clauses,
NIL,
dNumPartialPartialGroups));
}
-
- /*
- * Now we may consider incremental sort on this path, but only
- * when the path is not already sorted and when incremental sort
- * is enabled.
- */
- if (is_sorted || !enable_incremental_sort)
- continue;
-
- /* Restore the input path (we might have added Sort on top). */
- path = path_original;
-
- /* no shared prefix, not point in building incremental sort */
- if (presorted_keys == 0)
- continue;
-
- /*
- * We should have already excluded pathkeys of length 1 because
- * then presorted_keys > 0 would imply is_sorted was true.
- */
- Assert(list_length(root->group_pathkeys) != 1);
-
- path = (Path *) create_incremental_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
-
- if (parse->hasAggs)
- add_partial_path(partially_grouped_rel, (Path *)
- create_agg_path(root,
- partially_grouped_rel,
- path,
- partially_grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_INITIAL_SERIAL,
- parse->groupClause,
- NIL,
- agg_partial_costs,
- dNumPartialPartialGroups));
- else
- add_partial_path(partially_grouped_rel, (Path *)
- create_group_path(root,
- partially_grouped_rel,
- path,
- parse->groupClause,
- NIL,
- dNumPartialPartialGroups));
}
}
diff --git a/src/backend/optimizer/util/pathnode.c b/src/backend/optimizer/util/pathnode.c
index 5c32c96b71c..cd7c671a706 100644
--- a/src/backend/optimizer/util/pathnode.c
+++ b/src/backend/optimizer/util/pathnode.c
@@ -1342,7 +1342,7 @@ create_append_path(PlannerInfo *root,
pathnode->path.pathkeys = child->pathkeys;
}
else
- cost_append(pathnode);
+ cost_append(pathnode, root);
/* If the caller provided a row estimate, override the computed value. */
if (rows >= 0)
diff --git a/src/backend/utils/adt/selfuncs.c b/src/backend/utils/adt/selfuncs.c
index 1fbb0b28c3b..36c303e45c5 100644
--- a/src/backend/utils/adt/selfuncs.c
+++ b/src/backend/utils/adt/selfuncs.c
@@ -3294,7 +3294,10 @@ add_unique_group_var(PlannerInfo *root, List *varinfos,
}
/*
- * estimate_num_groups - Estimate number of groups in a grouped query
+ * estimate_num_groups/estimate_num_groups_incremental
+ * - Estimate number of groups in a grouped query.
+ * _incremental variant is performance optimization for
+ * case of adding one-by-one column
*
* Given a query having a GROUP BY clause, estimate how many groups there
* will be --- ie, the number of distinct combinations of the GROUP BY
@@ -3368,11 +3371,22 @@ double
estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
List **pgset, EstimationInfo *estinfo)
{
- List *varinfos = NIL;
+ return estimate_num_groups_incremental(root, groupExprs,
+ input_rows, pgset, estinfo,
+ NULL, 0);
+}
+
+double
+estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
+ double input_rows,
+ List **pgset, EstimationInfo *estinfo,
+ List **cache_varinfos, int prevNExprs)
+{
+ List *varinfos = (cache_varinfos) ? *cache_varinfos : NIL;
double srf_multiplier = 1.0;
double numdistinct;
ListCell *l;
- int i;
+ int i, j;
/* Zero the estinfo output parameter, if non-NULL */
if (estinfo != NULL)
@@ -3403,7 +3417,7 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
*/
numdistinct = 1.0;
- i = 0;
+ i = j = 0;
foreach(l, groupExprs)
{
Node *groupexpr = (Node *) lfirst(l);
@@ -3412,6 +3426,14 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
List *varshere;
ListCell *l2;
+ /* was done on previous call */
+ if (cache_varinfos && j++ < prevNExprs)
+ {
+ if (pgset)
+ i++; /* to keep in sync with lines below */
+ continue;
+ }
+
/* is expression in this grouping set? */
if (pgset && !list_member_int(*pgset, i++))
continue;
@@ -3481,7 +3503,11 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
if (varshere == NIL)
{
if (contain_volatile_functions(groupexpr))
+ {
+ if (cache_varinfos)
+ *cache_varinfos = varinfos;
return input_rows;
+ }
continue;
}
@@ -3498,6 +3524,9 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
}
}
+ if (cache_varinfos)
+ *cache_varinfos = varinfos;
+
/*
* If now no Vars, we must have an all-constant or all-boolean GROUP BY
* list.
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index e7f0a380e60..935d5bd9f6d 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2139,6 +2139,38 @@ static struct config_bool ConfigureNamesBool[] =
NULL, NULL, NULL
},
+/************************<DEBUG OPT GROUP BY>*********************************/
+ {
+ {"debug_group_by_reorder_by_pathkeys", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("enable reorder GROUP BY by pathkeys"),
+ NULL,
+ GUC_NOT_IN_SAMPLE
+ },
+ &debug_group_by_reorder_by_pathkeys,
+ true,
+ NULL, NULL, NULL
+ },
+ {
+ {"debug_enable_group_by_match_order_by", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("enable matching GROUP BY by ORDER BY."),
+ NULL,
+ GUC_NOT_IN_SAMPLE
+ },
+ &debug_group_by_match_order_by,
+ true,
+ NULL, NULL, NULL
+ },
+ {
+ {"debug_enable_cheapest_group_by", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("find a cheapest order of columns in GROUP BY."),
+ NULL,
+ GUC_NOT_IN_SAMPLE
+ },
+ &debug_cheapest_group_by,
+ true,
+ NULL, NULL, NULL
+ },
+/************************</DEBUG OPT GROUP BY>********************************/
/* End-of-list marker */
{
{NULL, 0, 0, NULL, NULL}, NULL, false, NULL, NULL, NULL
diff --git a/src/include/nodes/nodes.h b/src/include/nodes/nodes.h
index 5d075f0c346..34b9fbcf69e 100644
--- a/src/include/nodes/nodes.h
+++ b/src/include/nodes/nodes.h
@@ -266,6 +266,7 @@ typedef enum NodeTag
T_EquivalenceClass,
T_EquivalenceMember,
T_PathKey,
+ T_PathKeyInfo,
T_PathTarget,
T_RestrictInfo,
T_IndexClause,
diff --git a/src/include/nodes/pathnodes.h b/src/include/nodes/pathnodes.h
index 1f3845b3fec..83582301908 100644
--- a/src/include/nodes/pathnodes.h
+++ b/src/include/nodes/pathnodes.h
@@ -1070,6 +1070,16 @@ typedef struct PathKey
bool pk_nulls_first; /* do NULLs come before normal values? */
} PathKey;
+/*
+ * Combines information about pathkeys and the associated clauses.
+ */
+typedef struct PathKeyInfo
+{
+ NodeTag type;
+ List *pathkeys;
+ List *clauses;
+} PathKeyInfo;
+
/*
* VolatileFunctionStatus -- allows nodes to cache their
* contain_volatile_functions properties. VOLATILITY_UNKNOWN means not yet
diff --git a/src/include/optimizer/cost.h b/src/include/optimizer/cost.h
index 356a51f370a..0b083a27b3c 100644
--- a/src/include/optimizer/cost.h
+++ b/src/include/optimizer/cost.h
@@ -112,7 +112,9 @@ extern void cost_incremental_sort(Path *path,
Cost input_startup_cost, Cost input_total_cost,
double input_tuples, int width, Cost comparison_cost, int sort_mem,
double limit_tuples);
-extern void cost_append(AppendPath *path);
+extern Cost cost_sort_estimate(PlannerInfo *root, List *pathkeys,
+ int nPresortedKeys, double tuples);
+extern void cost_append(AppendPath *path, PlannerInfo *root);
extern void cost_merge_append(Path *path, PlannerInfo *root,
List *pathkeys, int n_streams,
Cost input_startup_cost, Cost input_total_cost,
diff --git a/src/include/optimizer/paths.h b/src/include/optimizer/paths.h
index 0c3a0b90c85..525b594cbd0 100644
--- a/src/include/optimizer/paths.h
+++ b/src/include/optimizer/paths.h
@@ -203,6 +203,17 @@ typedef enum
extern PathKeysComparison compare_pathkeys(List *keys1, List *keys2);
extern bool pathkeys_contained_in(List *keys1, List *keys2);
extern bool pathkeys_count_contained_in(List *keys1, List *keys2, int *n_common);
+extern int group_keys_reorder_by_pathkeys(List *pathkeys,
+ List **group_pathkeys,
+ List **group_clauses);
+/************************<DEBUG OPT GROUP BY>*********************************/
+extern bool debug_group_by_reorder_by_pathkeys;
+extern bool debug_group_by_match_order_by;
+extern bool debug_cheapest_group_by;
+/************************</DEBUG OPT GROUP BY>********************************/
+extern List *get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys,
+ List *pathkeys, List *clauses);
extern Path *get_cheapest_path_for_pathkeys(List *paths, List *pathkeys,
Relids required_outer,
CostSelector cost_criterion,
diff --git a/src/include/utils/selfuncs.h b/src/include/utils/selfuncs.h
index 8f3d73edfb2..c313a08d541 100644
--- a/src/include/utils/selfuncs.h
+++ b/src/include/utils/selfuncs.h
@@ -214,6 +214,11 @@ extern double estimate_num_groups(PlannerInfo *root, List *groupExprs,
double input_rows, List **pgset,
EstimationInfo *estinfo);
+extern double estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
+ double input_rows, List **pgset,
+ EstimationInfo *estinfo,
+ List **cache_varinfos, int prevNExprs);
+
extern void estimate_hash_bucket_stats(PlannerInfo *root,
Node *hashkey, double nbuckets,
Selectivity *mcv_freq,
diff --git a/src/test/regress/expected/aggregates.out b/src/test/regress/expected/aggregates.out
index 0a23a39aa29..601047fa3dd 100644
--- a/src/test/regress/expected/aggregates.out
+++ b/src/test/regress/expected/aggregates.out
@@ -1210,7 +1210,8 @@ explain (costs off)
select distinct min(f1), max(f1) from minmaxtest;
QUERY PLAN
---------------------------------------------------------------------------------------------
- Unique
+ HashAggregate
+ Group Key: $0, $1
InitPlan 1 (returns $0)
-> Limit
-> Merge Append
@@ -1233,10 +1234,8 @@ explain (costs off)
-> Index Only Scan using minmaxtest2i on minmaxtest2 minmaxtest_8
Index Cond: (f1 IS NOT NULL)
-> Index Only Scan Backward using minmaxtest3i on minmaxtest3 minmaxtest_9
- -> Sort
- Sort Key: ($0), ($1)
- -> Result
-(26 rows)
+ -> Result
+(25 rows)
select distinct min(f1), max(f1) from minmaxtest;
min | max
@@ -2448,6 +2447,241 @@ SELECT balk(hundred) FROM tenk1;
(1 row)
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+SELECT
+ i AS id,
+ i/2 AS p,
+ format('%60s', i%2) AS v,
+ i/4 AS c,
+ i/8 AS d,
+ (random() * (10000/8))::int as e --the same as d but no correlation with p
+ INTO btg
+FROM
+ generate_series(1, 10000) i;
+VACUUM btg;
+ANALYZE btg;
+-- GROUP BY optimization by reorder columns by frequency
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Sort
+ Sort Key: p, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Sort
+ Sort Key: p, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, c, v
+ -> Sort
+ Sort Key: p, c, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: v, p, c
+ -> Sort
+ Sort Key: v, p, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: p, d, c, v
+ -> Sort
+ Sort Key: p, d, c, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: v, p, d, c
+ -> Sort
+ Sort Key: v, p, d, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: p, v, d, c
+ -> Sort
+ Sort Key: p, v, d, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, d, e
+ -> Sort
+ Sort Key: p, d, e
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, e, d
+ -> Sort
+ Sort Key: p, e, d
+ -> Seq Scan on btg
+(5 rows)
+
+CREATE STATISTICS btg_dep ON d, e, p FROM btg;
+ANALYZE btg;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, d, e
+ -> Sort
+ Sort Key: p, d, e
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, e, d
+ -> Sort
+ Sort Key: p, e, d
+ -> Seq Scan on btg
+(5 rows)
+
+-- GROUP BY optimization by reorder columns by index scan
+CREATE INDEX ON btg(p, v);
+SET enable_seqscan=off;
+SET enable_bitmapscan=off;
+VACUUM btg;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v ORDER BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p ORDER BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, c, v
+ -> Incremental Sort
+ Sort Key: p, c, v
+ Presorted Key: p
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY p, v;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, v, c
+ -> Incremental Sort
+ Sort Key: p, v, c
+ Presorted Key: p, v
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, c, d, v
+ -> Incremental Sort
+ Sort Key: p, c, d, v
+ Presorted Key: p
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d ORDER BY p, v;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, v, c, d
+ -> Incremental Sort
+ Sort Key: p, v, c, d
+ Presorted Key: p, v
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+DROP TABLE btg;
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+RESET enable_seqscan;
+RESET enable_bitmapscan;
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/expected/guc.out b/src/test/regress/expected/guc.out
index 3de6404ba5b..98a16c118f2 100644
--- a/src/test/regress/expected/guc.out
+++ b/src/test/regress/expected/guc.out
@@ -858,10 +858,13 @@ SELECT name FROM tab_settings_flags
SELECT name FROM tab_settings_flags
WHERE category ~ '^Query Tuning' AND NOT explain
ORDER BY 1;
- name
----------------------------
+ name
+--------------------------------------
+ debug_enable_cheapest_group_by
+ debug_enable_group_by_match_order_by
+ debug_group_by_reorder_by_pathkeys
default_statistics_target
-(1 row)
+(4 rows)
-- Runtime-computed GUCs should be part of the preset category.
SELECT name FROM tab_settings_flags
diff --git a/src/test/regress/expected/incremental_sort.out b/src/test/regress/expected/incremental_sort.out
index 545e301e482..21c429226f7 100644
--- a/src/test/regress/expected/incremental_sort.out
+++ b/src/test/regress/expected/incremental_sort.out
@@ -1439,7 +1439,7 @@ set parallel_setup_cost = 0;
set parallel_tuple_cost = 0;
set max_parallel_workers_per_gather = 2;
create table t (a int, b int, c int);
-insert into t select mod(i,10),mod(i,10),i from generate_series(1,10000) s(i);
+insert into t select mod(i,10),mod(i,10),i from generate_series(1,60000) s(i);
create index on t (a);
analyze t;
set enable_incremental_sort = off;
diff --git a/src/test/regress/expected/join.out b/src/test/regress/expected/join.out
index 19caebabd01..bf1a2db2cf0 100644
--- a/src/test/regress/expected/join.out
+++ b/src/test/regress/expected/join.out
@@ -1984,8 +1984,8 @@ USING (name);
------+----+----
bb | 12 | 13
cc | 22 | 23
- dd | | 33
ee | 42 |
+ dd | | 33
(4 rows)
-- Cases with non-nullable expressions in subquery results;
@@ -2019,8 +2019,8 @@ NATURAL FULL JOIN
------+------+------+------+------
bb | 12 | 2 | 13 | 3
cc | 22 | 2 | 23 | 3
- dd | | | 33 | 3
ee | 42 | 2 | |
+ dd | | | 33 | 3
(4 rows)
SELECT * FROM
@@ -4618,18 +4618,20 @@ select d.* from d left join (select * from b group by b.id, b.c_id) s
explain (costs off)
select d.* from d left join (select distinct * from b) s
on d.a = s.id;
- QUERY PLAN
---------------------------------------
- Merge Right Join
- Merge Cond: (b.id = d.a)
- -> Unique
- -> Sort
- Sort Key: b.id, b.c_id
- -> Seq Scan on b
+ QUERY PLAN
+---------------------------------------------
+ Merge Left Join
+ Merge Cond: (d.a = s.id)
-> Sort
Sort Key: d.a
-> Seq Scan on d
-(9 rows)
+ -> Sort
+ Sort Key: s.id
+ -> Subquery Scan on s
+ -> HashAggregate
+ Group Key: b.id, b.c_id
+ -> Seq Scan on b
+(11 rows)
-- check join removal works when uniqueness of the join condition is enforced
-- by a UNION
@@ -6336,44 +6338,39 @@ select * from j1 natural join j2;
explain (verbose, costs off)
select * from j1
inner join (select distinct id from j3) j3 on j1.id = j3.id;
- QUERY PLAN
------------------------------------------
+ QUERY PLAN
+-----------------------------------
Nested Loop
Output: j1.id, j3.id
Inner Unique: true
Join Filter: (j1.id = j3.id)
- -> Unique
+ -> HashAggregate
Output: j3.id
- -> Sort
+ Group Key: j3.id
+ -> Seq Scan on public.j3
Output: j3.id
- Sort Key: j3.id
- -> Seq Scan on public.j3
- Output: j3.id
-> Seq Scan on public.j1
Output: j1.id
-(13 rows)
+(11 rows)
-- ensure group by clause allows the inner to become unique
explain (verbose, costs off)
select * from j1
inner join (select id from j3 group by id) j3 on j1.id = j3.id;
- QUERY PLAN
------------------------------------------
+ QUERY PLAN
+-----------------------------------
Nested Loop
Output: j1.id, j3.id
Inner Unique: true
Join Filter: (j1.id = j3.id)
- -> Group
+ -> HashAggregate
Output: j3.id
Group Key: j3.id
- -> Sort
+ -> Seq Scan on public.j3
Output: j3.id
- Sort Key: j3.id
- -> Seq Scan on public.j3
- Output: j3.id
-> Seq Scan on public.j1
Output: j1.id
-(14 rows)
+(11 rows)
drop table j1;
drop table j2;
diff --git a/src/test/regress/expected/partition_aggregate.out b/src/test/regress/expected/partition_aggregate.out
index dfa4b036b52..a08a3825ff6 100644
--- a/src/test/regress/expected/partition_aggregate.out
+++ b/src/test/regress/expected/partition_aggregate.out
@@ -952,32 +952,30 @@ SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HA
--------------------------------------------------------------------------------------
Sort
Sort Key: pagg_tab_ml.a, (sum(pagg_tab_ml.b)), (array_agg(DISTINCT pagg_tab_ml.c))
- -> Gather
- Workers Planned: 2
- -> Parallel Append
- -> GroupAggregate
- Group Key: pagg_tab_ml.a
- Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml.a
- -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
- -> GroupAggregate
- Group Key: pagg_tab_ml_5.a
- Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_5.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
- -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
- -> GroupAggregate
- Group Key: pagg_tab_ml_2.a
- Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_2.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
- -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
-(27 rows)
+ -> Append
+ -> GroupAggregate
+ Group Key: pagg_tab_ml.a
+ Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml.a
+ -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_2.a
+ Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_2.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
+ -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_5.a
+ Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_5.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
+ -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
+(25 rows)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3 ORDER BY 1, 2, 3;
a | sum | array_agg | count
@@ -996,34 +994,32 @@ SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HA
-- Without ORDER BY clause, to test Gather at top-most path
EXPLAIN (COSTS OFF)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3;
- QUERY PLAN
----------------------------------------------------------------------------
- Gather
- Workers Planned: 2
- -> Parallel Append
- -> GroupAggregate
- Group Key: pagg_tab_ml.a
- Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml.a
- -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
- -> GroupAggregate
- Group Key: pagg_tab_ml_5.a
- Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_5.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
- -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
- -> GroupAggregate
- Group Key: pagg_tab_ml_2.a
- Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_2.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
- -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
-(25 rows)
+ QUERY PLAN
+---------------------------------------------------------------------
+ Append
+ -> GroupAggregate
+ Group Key: pagg_tab_ml.a
+ Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml.a
+ -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_2.a
+ Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_2.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
+ -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_5.a
+ Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_5.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
+ -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
+(23 rows)
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
@@ -1379,28 +1375,26 @@ SELECT x, sum(y), avg(y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) <
-- When GROUP BY clause does not match; partial aggregation is performed for each partition.
EXPLAIN (COSTS OFF)
SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) < 12 ORDER BY 1, 2, 3;
- QUERY PLAN
--------------------------------------------------------------------------------------------
+ QUERY PLAN
+-------------------------------------------------------------------------------------
Sort
Sort Key: pagg_tab_para.y, (sum(pagg_tab_para.x)), (avg(pagg_tab_para.x))
- -> Finalize GroupAggregate
+ -> Finalize HashAggregate
Group Key: pagg_tab_para.y
Filter: (avg(pagg_tab_para.x) < '12'::numeric)
- -> Gather Merge
+ -> Gather
Workers Planned: 2
- -> Sort
- Sort Key: pagg_tab_para.y
- -> Parallel Append
- -> Partial HashAggregate
- Group Key: pagg_tab_para.y
- -> Parallel Seq Scan on pagg_tab_para_p1 pagg_tab_para
- -> Partial HashAggregate
- Group Key: pagg_tab_para_1.y
- -> Parallel Seq Scan on pagg_tab_para_p2 pagg_tab_para_1
- -> Partial HashAggregate
- Group Key: pagg_tab_para_2.y
- -> Parallel Seq Scan on pagg_tab_para_p3 pagg_tab_para_2
-(19 rows)
+ -> Parallel Append
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para.y
+ -> Parallel Seq Scan on pagg_tab_para_p1 pagg_tab_para
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para_1.y
+ -> Parallel Seq Scan on pagg_tab_para_p2 pagg_tab_para_1
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para_2.y
+ -> Parallel Seq Scan on pagg_tab_para_p3 pagg_tab_para_2
+(17 rows)
SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) < 12 ORDER BY 1, 2, 3;
y | sum | avg | count
diff --git a/src/test/regress/expected/partition_join.out b/src/test/regress/expected/partition_join.out
index bb5b7c47a45..03926a84138 100644
--- a/src/test/regress/expected/partition_join.out
+++ b/src/test/regress/expected/partition_join.out
@@ -466,52 +466,41 @@ EXPLAIN (COSTS OFF)
SELECT a, b FROM prt1 FULL JOIN prt2 p2(b,a,c) USING(a,b)
WHERE a BETWEEN 490 AND 510
GROUP BY 1, 2 ORDER BY 1, 2;
- QUERY PLAN
------------------------------------------------------------------------------------------------------------------
+ QUERY PLAN
+-----------------------------------------------------------------------------------------------------------
Group
Group Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Merge Append
+ -> Sort
Sort Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Group
- Group Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Sort
- Sort Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Merge Full Join
- Merge Cond: ((prt1.a = p2.a) AND (prt1.b = p2.b))
- Filter: ((COALESCE(prt1.a, p2.a) >= 490) AND (COALESCE(prt1.a, p2.a) <= 510))
- -> Sort
- Sort Key: prt1.a, prt1.b
- -> Seq Scan on prt1_p1 prt1
- -> Sort
- Sort Key: p2.a, p2.b
- -> Seq Scan on prt2_p1 p2
- -> Group
- Group Key: (COALESCE(prt1_1.a, p2_1.a)), (COALESCE(prt1_1.b, p2_1.b))
- -> Sort
- Sort Key: (COALESCE(prt1_1.a, p2_1.a)), (COALESCE(prt1_1.b, p2_1.b))
- -> Merge Full Join
- Merge Cond: ((prt1_1.a = p2_1.a) AND (prt1_1.b = p2_1.b))
- Filter: ((COALESCE(prt1_1.a, p2_1.a) >= 490) AND (COALESCE(prt1_1.a, p2_1.a) <= 510))
- -> Sort
- Sort Key: prt1_1.a, prt1_1.b
- -> Seq Scan on prt1_p2 prt1_1
- -> Sort
- Sort Key: p2_1.a, p2_1.b
- -> Seq Scan on prt2_p2 p2_1
- -> Group
- Group Key: (COALESCE(prt1_2.a, p2_2.a)), (COALESCE(prt1_2.b, p2_2.b))
- -> Sort
- Sort Key: (COALESCE(prt1_2.a, p2_2.a)), (COALESCE(prt1_2.b, p2_2.b))
- -> Merge Full Join
- Merge Cond: ((prt1_2.a = p2_2.a) AND (prt1_2.b = p2_2.b))
- Filter: ((COALESCE(prt1_2.a, p2_2.a) >= 490) AND (COALESCE(prt1_2.a, p2_2.a) <= 510))
- -> Sort
- Sort Key: prt1_2.a, prt1_2.b
- -> Seq Scan on prt1_p3 prt1_2
- -> Sort
- Sort Key: p2_2.a, p2_2.b
- -> Seq Scan on prt2_p3 p2_2
-(43 rows)
+ -> Append
+ -> Merge Full Join
+ Merge Cond: ((prt1_1.a = p2_1.a) AND (prt1_1.b = p2_1.b))
+ Filter: ((COALESCE(prt1_1.a, p2_1.a) >= 490) AND (COALESCE(prt1_1.a, p2_1.a) <= 510))
+ -> Sort
+ Sort Key: prt1_1.a, prt1_1.b
+ -> Seq Scan on prt1_p1 prt1_1
+ -> Sort
+ Sort Key: p2_1.a, p2_1.b
+ -> Seq Scan on prt2_p1 p2_1
+ -> Merge Full Join
+ Merge Cond: ((prt1_2.a = p2_2.a) AND (prt1_2.b = p2_2.b))
+ Filter: ((COALESCE(prt1_2.a, p2_2.a) >= 490) AND (COALESCE(prt1_2.a, p2_2.a) <= 510))
+ -> Sort
+ Sort Key: prt1_2.a, prt1_2.b
+ -> Seq Scan on prt1_p2 prt1_2
+ -> Sort
+ Sort Key: p2_2.a, p2_2.b
+ -> Seq Scan on prt2_p2 p2_2
+ -> Merge Full Join
+ Merge Cond: ((prt1_3.b = p2_3.b) AND (prt1_3.a = p2_3.a))
+ Filter: ((COALESCE(prt1_3.a, p2_3.a) >= 490) AND (COALESCE(prt1_3.a, p2_3.a) <= 510))
+ -> Sort
+ Sort Key: prt1_3.b, prt1_3.a
+ -> Seq Scan on prt1_p3 prt1_3
+ -> Sort
+ Sort Key: p2_3.b, p2_3.a
+ -> Seq Scan on prt2_p3 p2_3
+(32 rows)
SELECT a, b FROM prt1 FULL JOIN prt2 p2(b,a,c) USING(a,b)
WHERE a BETWEEN 490 AND 510
diff --git a/src/test/regress/expected/union.out b/src/test/regress/expected/union.out
index dece7310cfe..7ac4a9380e2 100644
--- a/src/test/regress/expected/union.out
+++ b/src/test/regress/expected/union.out
@@ -1303,24 +1303,22 @@ select distinct q1 from
union all
select distinct * from int8_tbl i82) ss
where q2 = q2;
- QUERY PLAN
-----------------------------------------------------------
- Unique
- -> Merge Append
- Sort Key: "*SELECT* 1".q1
+ QUERY PLAN
+----------------------------------------------------
+ HashAggregate
+ Group Key: "*SELECT* 1".q1
+ -> Append
-> Subquery Scan on "*SELECT* 1"
- -> Unique
- -> Sort
- Sort Key: i81.q1, i81.q2
- -> Seq Scan on int8_tbl i81
- Filter: (q2 IS NOT NULL)
+ -> HashAggregate
+ Group Key: i81.q1, i81.q2
+ -> Seq Scan on int8_tbl i81
+ Filter: (q2 IS NOT NULL)
-> Subquery Scan on "*SELECT* 2"
- -> Unique
- -> Sort
- Sort Key: i82.q1, i82.q2
- -> Seq Scan on int8_tbl i82
- Filter: (q2 IS NOT NULL)
-(15 rows)
+ -> HashAggregate
+ Group Key: i82.q1, i82.q2
+ -> Seq Scan on int8_tbl i82
+ Filter: (q2 IS NOT NULL)
+(13 rows)
select distinct q1 from
(select distinct * from int8_tbl i81
@@ -1339,24 +1337,22 @@ select distinct q1 from
union all
select distinct * from int8_tbl i82) ss
where -q1 = q2;
- QUERY PLAN
---------------------------------------------------------
- Unique
- -> Merge Append
- Sort Key: "*SELECT* 1".q1
+ QUERY PLAN
+--------------------------------------------------
+ HashAggregate
+ Group Key: "*SELECT* 1".q1
+ -> Append
-> Subquery Scan on "*SELECT* 1"
- -> Unique
- -> Sort
- Sort Key: i81.q1, i81.q2
- -> Seq Scan on int8_tbl i81
- Filter: ((- q1) = q2)
+ -> HashAggregate
+ Group Key: i81.q1, i81.q2
+ -> Seq Scan on int8_tbl i81
+ Filter: ((- q1) = q2)
-> Subquery Scan on "*SELECT* 2"
- -> Unique
- -> Sort
- Sort Key: i82.q1, i82.q2
- -> Seq Scan on int8_tbl i82
- Filter: ((- q1) = q2)
-(15 rows)
+ -> HashAggregate
+ Group Key: i82.q1, i82.q2
+ -> Seq Scan on int8_tbl i82
+ Filter: ((- q1) = q2)
+(13 rows)
select distinct q1 from
(select distinct * from int8_tbl i81
diff --git a/src/test/regress/sql/aggregates.sql b/src/test/regress/sql/aggregates.sql
index 2f5d0e00f3d..c6e0d7ba2b7 100644
--- a/src/test/regress/sql/aggregates.sql
+++ b/src/test/regress/sql/aggregates.sql
@@ -1025,6 +1025,105 @@ SELECT balk(hundred) FROM tenk1;
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+
+SELECT
+ i AS id,
+ i/2 AS p,
+ format('%60s', i%2) AS v,
+ i/4 AS c,
+ i/8 AS d,
+ (random() * (10000/8))::int as e --the same as d but no correlation with p
+ INTO btg
+FROM
+ generate_series(1, 10000) i;
+
+VACUUM btg;
+ANALYZE btg;
+
+-- GROUP BY optimization by reorder columns by frequency
+
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+
+CREATE STATISTICS btg_dep ON d, e, p FROM btg;
+ANALYZE btg;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+
+
+-- GROUP BY optimization by reorder columns by index scan
+
+CREATE INDEX ON btg(p, v);
+SET enable_seqscan=off;
+SET enable_bitmapscan=off;
+VACUUM btg;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d ORDER BY p, v;
+
+DROP TABLE btg;
+
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+RESET enable_seqscan;
+RESET enable_bitmapscan;
+
+
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/sql/incremental_sort.sql b/src/test/regress/sql/incremental_sort.sql
index d8768a6b54d..1de163e0395 100644
--- a/src/test/regress/sql/incremental_sort.sql
+++ b/src/test/regress/sql/incremental_sort.sql
@@ -213,7 +213,7 @@ set parallel_tuple_cost = 0;
set max_parallel_workers_per_gather = 2;
create table t (a int, b int, c int);
-insert into t select mod(i,10),mod(i,10),i from generate_series(1,10000) s(i);
+insert into t select mod(i,10),mod(i,10),i from generate_series(1,60000) s(i);
create index on t (a);
analyze t;
--
2.34.1
[text/x-patch] 0002-Use-default-restriction-for-number-of-group-20220315.patch (2.1K, ../../[email protected]/3-0002-Use-default-restriction-for-number-of-group-20220315.patch)
download | inline diff:
From cfd3515a0024515546afe89fc9b2595990eaf3b2 Mon Sep 17 00:00:00 2001
From: Andrey Lepikhov <[email protected]>
Date: Thu, 10 Feb 2022 13:51:51 +0500
Subject: [PATCH 2/3] Use default restriction for number of groups.
---
src/backend/optimizer/path/costsize.c | 22 +++++++---------------
1 file changed, 7 insertions(+), 15 deletions(-)
diff --git a/src/backend/optimizer/path/costsize.c b/src/backend/optimizer/path/costsize.c
index 68a32740d7e..b9e975df10c 100644
--- a/src/backend/optimizer/path/costsize.c
+++ b/src/backend/optimizer/path/costsize.c
@@ -1756,8 +1756,8 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
/*
* is_fake_var
- * Workaround for generate_append_tlist() which generates fake Vars with
- * varno == 0, that will cause a fail of estimate_num_group() call
+ * Workaround for generate_append_tlist() which generates fake Vars for the
+ * case of "varno 0", that will cause a fail of estimate_num_group() call
*
* XXX Ummm, why would estimate_num_group fail with this?
*/
@@ -1978,21 +1978,13 @@ compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
tuplesPerPrevGroup, NULL, NULL,
&cache_varinfos,
list_length(pathkeyExprs) - 1);
- else if (tuples > 4.0)
+ else
/*
- * Use geometric mean as estimation if there is no any stats.
- * Don't use DEFAULT_NUM_DISTINCT because it used for only one
- * column while here we try to estimate number of groups over
- * set of columns.
- *
- * XXX Perhaps this should use DEFAULT_NUM_DISTINCT at least to
- * limit the calculated values, somehow?
- *
- * XXX What's the logic of the following formula?
+ * In case of full uncertainity use default defensive approach. It
+ * means that any permutations of such vars are equivalent.
+ * Also, see comments for the cost_incremental_sort routine.
*/
- nGroups = ceil(2.0 + sqrt(tuples) * (i + 1) / list_length(pathkeys));
- else
- nGroups = tuples;
+ nGroups = Min(tuplesPerPrevGroup, DEFAULT_NUM_DISTINCT);
/*
* Presorted keys aren't participated in comparison but still checked
--
2.34.1
[text/x-patch] 0003-fixup-tweak-costing-20220315.patch (3.4K, ../../[email protected]/4-0003-fixup-tweak-costing-20220315.patch)
download | inline diff:
From 7432a917c30ab5e5f1db9d2569a38e59e4f37c3a Mon Sep 17 00:00:00 2001
From: Tomas Vondra <[email protected]>
Date: Mon, 14 Mar 2022 23:28:55 +0100
Subject: [PATCH 3/3] fixup: tweak costing
---
src/backend/optimizer/path/costsize.c | 42 +++++++++++++--------------
1 file changed, 21 insertions(+), 21 deletions(-)
diff --git a/src/backend/optimizer/path/costsize.c b/src/backend/optimizer/path/costsize.c
index b9e975df10c..7224efaba5e 100644
--- a/src/backend/optimizer/path/costsize.c
+++ b/src/backend/optimizer/path/costsize.c
@@ -1872,6 +1872,15 @@ get_width_cost_multiplier(PlannerInfo *root, Expr *expr)
*
* N * sum( Fk * log(Gk) )
*
+ * For bounded heap sort we haven't such a duplicates-related research. So invent
+ * it based on a simple logic (M - number of output tuples):
+ * For one column we can estimate number of comparisons as:
+ * N * log(min{N,M})
+ * For the case of many columns we can natively estimate number of comparisons by
+ * the formula:
+ * sum(max{n_i,M} * log(min{N,M})),
+ * where n_0 == N, n_i - number of tuples per group, estimated on previous step.
+ *
* Note: We also consider column width, not just the comparator cost.
*
* NOTE: some callers currently pass NIL for pathkeys because they
@@ -1881,7 +1890,7 @@ get_width_cost_multiplier(PlannerInfo *root, Expr *expr)
static Cost
compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
Cost comparison_cost, double tuples, double output_tuples,
- bool heapSort)
+ bool bounded_sort)
{
Cost per_tuple_cost = 0.0;
ListCell *lc;
@@ -1907,7 +1916,7 @@ compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
* of this function, but I'm not sure. I suggest we introduce some simple
* constants for that, instead of magic values.
*/
- output_tuples = (heapSort) ? 2.0 * output_tuples : tuples;
+ output_tuples = (bounded_sort) ? 2.0 * output_tuples : tuples;
per_tuple_cost += 2.0 * cpu_operator_cost * LOG2(output_tuples);
/* add cost provided by caller */
@@ -1925,8 +1934,7 @@ compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
{
PathKey *pathkey = (PathKey*) lfirst(lc);
EquivalenceMember *em;
- double nGroups,
- correctedNGroups;
+ double nGroups;
/*
* We believe that equivalence members aren't very different, so, to
@@ -1992,24 +2000,16 @@ compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
*/
if (i >= nPresortedKeys)
{
- if (heapSort)
- {
- double heap_tuples;
-
- /* have to keep at least one group, and a multiple of group size */
- heap_tuples = Max(ceil(output_tuples / tuplesPerPrevGroup) * tuplesPerPrevGroup,
- tuplesPerPrevGroup);
-
- /* so how many (whole) groups is that? */
- correctedNGroups = ceil(heap_tuples / tuplesPerPrevGroup);
- }
+ /*
+ * Quick sort and 'top-N' sorting (bounded heap sort) algorithms
+ * have different formulas for time complexity estimation.
+ */
+ if (bounded_sort)
+ per_tuple_cost += totalFuncCost *
+ (Max(tuplesPerPrevGroup, output_tuples) / tuples) *
+ LOG2(2.0 * Min(tuplesPerPrevGroup, output_tuples));
else
- /* all groups in the input */
- correctedNGroups = nGroups;
-
- correctedNGroups = Max(1.0, ceil(correctedNGroups));
-
- per_tuple_cost += totalFuncCost * LOG2(correctedNGroups);
+ per_tuple_cost += totalFuncCost * LOG2(nGroups);
}
i++;
--
2.34.1
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-03-18 07:06 Andrey V. Lepikhov <[email protected]>
parent: Tomas Vondra <[email protected]>
0 siblings, 1 reply; 38+ messages in thread
From: Andrey V. Lepikhov @ 2022-03-18 07:06 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; Teodor Sigaev <[email protected]>; +Cc: Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On 3/15/22 13:26, Tomas Vondra wrote:
> Thanks for the rebase. The two proposed changes (tweaked costing and
> simplified fake_var handling) seem fine to me. I think the last thing
> that needs to be done is cleanup of the debug GUCs, which I added to
> allow easier experimentation with the patch.
Thanks, I'm waiting for the last step.
>
> I probably won't remove the GUCs entirely, though. I plan to add a
> single GUC that would enable/disable this optimization. I'm not a huge
> fan of adding more and more GUCs, but in this case it's probably the
> right thing to do given the complexity of estimating cost with
> correlated columns etc.
Agree. Because it is a kind of automation we should allow user to switch
it off in the case of problems or manual tuning.
Also, I looked through this patch. It has some minor problems:
1. Multiple typos in the patch comment.
2. The term 'cardinality of a key' - may be replace with 'number of
duplicates'?
--
regards,
Andrey Lepikhov
Postgres Professional
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-03-29 00:48 Tomas Vondra <[email protected]>
parent: Andrey V. Lepikhov <[email protected]>
0 siblings, 1 reply; 38+ messages in thread
From: Tomas Vondra @ 2022-03-29 00:48 UTC (permalink / raw)
To: Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; +Cc: Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
Hi,
Here's a rebased/improved version of the patch, with smaller parts
addressing various issues. There are seven parts:
0001 - main part, just rebased
0002 - replace the debug GUC options with a single GUC to disable the
optimization if needed
0003 - minor code cleanup, removal of unnecessary variable
0004 - various comment fixes (rewordings, typos, ...)
0005 - a minor code simplification, addressing FIXMEs from 0004
0006 - adds the new GUC to the docs
0007 - demonstrates plan changes with a disabled optimization
The first 6 parts should be squashed and committed at one, I only kept
them separate for clarity. The 0007 is merely a demonstration of the new
GUC and that it disables the optimization.
> Agree. Because it is a kind of automation we should allow user to switch
> it off in the case of problems or manual tuning.
> > Also, I looked through this patch. It has some minor problems:
> 1. Multiple typos in the patch comment.
I went through the comments and checked all of them for grammar mistakes
and typos using a word processor, so hopefully that should be OK. But
maybe there's still something wrong.
> 2. The term 'cardinality of a key' - may be replace with 'number of
> duplicates'?
No, cardinality means "number of distinct values", so "duplicates" would
be wrong. And I think "cardinality" is well established term, so I think
it's fine.
BTW I named the GUC enable_group_by_reordering, I wonder if it should be
named differently, e.g. enable_groupby_reordering? Opinions?
regards
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
Attachments:
[text/x-patch] 0001-Optimize-order-of-GROUP-BY-keys-20220328.patch (102.1K, ../../[email protected]/2-0001-Optimize-order-of-GROUP-BY-keys-20220328.patch)
download | inline diff:
From 61369b499770918edb929d0ca837a123bea3fcdb Mon Sep 17 00:00:00 2001
From: Tomas Vondra <[email protected]>
Date: Fri, 21 Jan 2022 20:22:14 +0100
Subject: [PATCH 1/7] Optimize order of GROUP BY keys
When evaluating a query with a multi-column GROUP BY clause using sort,
the cost may be heavily dependent on the order in which the keys are
compared when building the groups. Grouping does not imply any ordering,
so we're allowed to compare the keys in arbitrary order, and a Hash Agg
leverages this. But for Group Agg, we simply compared keys in the order
as specified in the query. This commit explores alternative ordering of
the keys, trying to find a cheaper one.
To pick the optimal order of comparisons, multiple parametes need to be
considered, both local and global ones. Intuitively, it's better to
leave more expensive comparions (for complex data types etc.) at the
very end, because maybe one of the earlier comparisons will be enough to
make a decision, and the expensive comparison will be unnecessary. The
other local parameter is cardinality - the higher the cardinality of a
key, the lower the probability we'll need to compare additional keys. We
explore possible orderings by a combination of exhaustive and greedy
approach, and pick the cheapest ordering.
But there are global parameters too, in the sense that we can't fully
evaluate them at the point when creating the sort. For example, there
might be an ORDER BY clause, or another ordering-dependent operation,
higher up in the query plan. The path may be already sorted by (some
of) the keys, in which case we can either use incremental sort or even
skip the sort entirely. So we consider various additional pathkeys in
addition to the "cheapers comparion" order.
The original patch was proposed by Teodor Sigaev, and later reworked and
improved by Dmitry Dolgov. Reviews by a number of people, including me,
Andrey Lepikhov, Claudio Freire and Ibrar Ahmed.
Author: Dmitry Dolgov, Teodor Sigaev
Reviewed-by: Tomas Vondra, Andrey Lepikhov, Claudio Freire, Ibrar Ahmed
Discussion: https://postgr.es/m/7c79e6a5-8597-74e8-0671-1c39d124c9d6%40sigaev.ru
Discussion: https://postgr.es/m/CA%2Bq6zcW_4o2NC0zutLkOJPsFt80megSpX_dVRo6GK9PC-Jx_Ag%40mail.gmail.com
---
.../postgres_fdw/expected/postgres_fdw.out | 15 +-
src/backend/optimizer/path/costsize.c | 369 +++++++++-
src/backend/optimizer/path/equivclass.c | 13 +-
src/backend/optimizer/path/pathkeys.c | 580 ++++++++++++++++
src/backend/optimizer/plan/planner.c | 653 ++++++++++--------
src/backend/optimizer/util/pathnode.c | 2 +-
src/backend/utils/adt/selfuncs.c | 37 +-
src/backend/utils/misc/guc.c | 32 +
src/include/nodes/nodes.h | 1 +
src/include/nodes/pathnodes.h | 10 +
src/include/optimizer/cost.h | 4 +-
src/include/optimizer/paths.h | 11 +
src/include/utils/selfuncs.h | 5 +
src/test/regress/expected/aggregates.out | 244 ++++++-
src/test/regress/expected/guc.out | 9 +-
.../regress/expected/incremental_sort.out | 2 +-
src/test/regress/expected/join.out | 51 +-
.../regress/expected/partition_aggregate.out | 136 ++--
src/test/regress/expected/partition_join.out | 75 +-
src/test/regress/expected/union.out | 60 +-
src/test/regress/sql/aggregates.sql | 99 +++
src/test/regress/sql/incremental_sort.sql | 2 +-
22 files changed, 1913 insertions(+), 497 deletions(-)
diff --git a/contrib/postgres_fdw/expected/postgres_fdw.out b/contrib/postgres_fdw/expected/postgres_fdw.out
index f210f911880..2ffc836824a 100644
--- a/contrib/postgres_fdw/expected/postgres_fdw.out
+++ b/contrib/postgres_fdw/expected/postgres_fdw.out
@@ -2741,16 +2741,13 @@ select c2 * (random() <= 1)::int as c2 from ft2 group by c2 * (random() <= 1)::i
-- GROUP BY clause in various forms, cardinal, alias and constant expression
explain (verbose, costs off)
select count(c2) w, c2 x, 5 y, 7.0 z from ft1 group by 2, y, 9.0::int order by 2;
- QUERY PLAN
----------------------------------------------------------------------------------------
- Sort
+ QUERY PLAN
+------------------------------------------------------------------------------------------------------------
+ Foreign Scan
Output: (count(c2)), c2, 5, 7.0, 9
- Sort Key: ft1.c2
- -> Foreign Scan
- Output: (count(c2)), c2, 5, 7.0, 9
- Relations: Aggregate on (public.ft1)
- Remote SQL: SELECT count(c2), c2, 5, 7.0, 9 FROM "S 1"."T 1" GROUP BY 2, 3, 5
-(7 rows)
+ Relations: Aggregate on (public.ft1)
+ Remote SQL: SELECT count(c2), c2, 5, 7.0, 9 FROM "S 1"."T 1" GROUP BY 2, 3, 5 ORDER BY c2 ASC NULLS LAST
+(4 rows)
select count(c2) w, c2 x, 5 y, 7.0 z from ft1 group by 2, y, 9.0::int order by 2;
w | x | y | z
diff --git a/src/backend/optimizer/path/costsize.c b/src/backend/optimizer/path/costsize.c
index 1b07ea392d9..f4f7ce01065 100644
--- a/src/backend/optimizer/path/costsize.c
+++ b/src/backend/optimizer/path/costsize.c
@@ -1755,6 +1755,325 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
rterm->pathtarget->width);
}
+/*
+ * is_fake_var
+ * Workaround for generate_append_tlist() which generates fake Vars with
+ * varno == 0, that will cause a fail of estimate_num_group() call
+ *
+ * XXX Ummm, why would estimate_num_group fail with this?
+ */
+static bool
+is_fake_var(Expr *expr)
+{
+ if (IsA(expr, RelabelType))
+ expr = (Expr *) ((RelabelType *) expr)->arg;
+
+ return (IsA(expr, Var) && ((Var *) expr)->varno == 0);
+}
+
+/*
+ * get_width_cost_multiplier
+ * Returns relative complexity of comparing two values based on it's width.
+ * The idea behind - long values have more expensive comparison. Return value is
+ * in cpu_operator_cost unit.
+ */
+static double
+get_width_cost_multiplier(PlannerInfo *root, Expr *expr)
+{
+ double width = -1.0; /* fake value */
+
+ if (IsA(expr, RelabelType))
+ expr = (Expr *) ((RelabelType *) expr)->arg;
+
+ /* Try to find actual stat in corresponding relation */
+ if (IsA(expr, Var))
+ {
+ Var *var = (Var *) expr;
+
+ if (var->varno > 0 && var->varno < root->simple_rel_array_size)
+ {
+ RelOptInfo *rel = root->simple_rel_array[var->varno];
+
+ if (rel != NULL &&
+ var->varattno >= rel->min_attr &&
+ var->varattno <= rel->max_attr)
+ {
+ int ndx = var->varattno - rel->min_attr;
+
+ if (rel->attr_widths[ndx] > 0)
+ width = rel->attr_widths[ndx];
+ }
+ }
+ }
+
+ /* Didn't find any actual stats, use estimation by type */
+ if (width < 0.0)
+ {
+ Node *node = (Node*) expr;
+
+ width = get_typavgwidth(exprType(node), exprTypmod(node));
+ }
+
+ /*
+ * Any value in pgsql is passed by Datum type, so any operation with value
+ * could not be cheaper than operation with Datum type
+ */
+ if (width <= sizeof(Datum))
+ return 1.0;
+
+ /*
+ * Seems, cost of comparision is not directly proportional to args width,
+ * because comparing args could be differ width (we known only average over
+ * column) and difference often could be defined only by looking on first
+ * bytes. So, use log16(width) as estimation.
+ */
+ return 1.0 + 0.125 * LOG2(width / sizeof(Datum));
+}
+
+/*
+ * compute_cpu_sort_cost
+ * compute CPU cost of sort (i.e. in-memory)
+ *
+ * The main thing we need to calculate to estimate sort CPU costs is the number
+ * of calls to the comparator functions. The difficulty is that for multi-column
+ * sorts there may be different data types involved (for some of which the calls
+ * may be much more expensive). Furthermore, the columns may have very different
+ * number of distinct values - the higher the number, the fewer comparisons will
+ * be needed for the following columns.
+ *
+ * The algoritm is incremental - we add pathkeys one by one, and at each step we
+ * estimate the number of necessary comparisons (based on the number of distinct
+ * groups in the current pathkey prefix and the new pathkey), and the comparison
+ * costs (which is data type specific).
+ *
+ * Estimation of the number of comparisons is based on ideas from:
+ *
+ * "Quicksort Is Optimal", Robert Sedgewick, Jon Bentley, 2002
+ * [https://www.cs.princeton.edu/~rs/talks/QuicksortIsOptimal.pdf]
+ *
+ * In term of that paper, let N - number of tuples, Xi - number of identical
+ * tuples with value Ki, then the estimate of number of comparisons is:
+ *
+ * log(N! / (X1! * X2! * ..)) ~ sum(Xi * log(N/Xi))
+ *
+ * We assume all Xi the same because now we don't have any estimation of
+ * group sizes, we have only know the estimate of number of groups (distinct
+ * values). In that case, formula becomes:
+ *
+ * N * log(NumberOfGroups)
+ *
+ * For multi-column sorts we need to estimate the number of comparisons for
+ * each individual column - for example with columns (c1, c2, ..., ck) we
+ * can estimate that number of comparisons on ck is roughly
+ *
+ * ncomparisons(c1, c2, ..., ck) / ncomparisons(c1, c2, ..., c(k-1))
+ *
+ * Let k be a column number, Gk - number of groups defined by k columns, and Fk
+ * the cost of the comparison is
+ *
+ * N * sum( Fk * log(Gk) )
+ *
+ * Note: We also consider column width, not just the comparator cost.
+ *
+ * NOTE: some callers currently pass NIL for pathkeys because they
+ * can't conveniently supply the sort keys. In this case, it will fallback to
+ * simple comparison cost estimate.
+ */
+static Cost
+compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
+ Cost comparison_cost, double tuples, double output_tuples,
+ bool heapSort)
+{
+ Cost per_tuple_cost = 0.0;
+ ListCell *lc;
+ List *pathkeyExprs = NIL;
+ double tuplesPerPrevGroup = tuples;
+ double totalFuncCost = 1.0;
+ bool has_fake_var = false;
+ int i = 0;
+ Oid prev_datatype = InvalidOid;
+ Cost funcCost = 0.0;
+ List *cache_varinfos = NIL;
+
+ /* fallback if pathkeys is unknown */
+ if (list_length(pathkeys) == 0)
+ {
+ /*
+ * If we'll use a bounded heap-sort keeping just K tuples in memory, for
+ * a total number of tuple comparisons of N log2 K; but the constant
+ * factor is a bit higher than for quicksort. Tweak it so that the
+ * cost curve is continuous at the crossover point.
+ *
+ * XXX I suppose the "quicksort factor" references to 1.5 at the end
+ * of this function, but I'm not sure. I suggest we introduce some simple
+ * constants for that, instead of magic values.
+ */
+ output_tuples = (heapSort) ? 2.0 * output_tuples : tuples;
+ per_tuple_cost += 2.0 * cpu_operator_cost * LOG2(output_tuples);
+
+ /* add cost provided by caller */
+ per_tuple_cost += comparison_cost;
+
+ return per_tuple_cost * tuples;
+ }
+
+ /*
+ * Computing total cost of sorting takes into account:
+ * - per column comparison function cost
+ * - we try to compute needed number of comparison per column
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey*) lfirst(lc);
+ EquivalenceMember *em;
+ double nGroups,
+ correctedNGroups;
+
+ /*
+ * We believe that equivalence members aren't very different, so, to
+ * estimate cost we take just first member
+ */
+ em = (EquivalenceMember *) linitial(pathkey->pk_eclass->ec_members);
+
+ if (em->em_datatype != InvalidOid)
+ {
+ /* do not lookup funcCost if data type is the same as previous */
+ if (prev_datatype != em->em_datatype)
+ {
+ Oid sortop;
+ QualCost cost;
+
+ sortop = get_opfamily_member(pathkey->pk_opfamily,
+ em->em_datatype, em->em_datatype,
+ pathkey->pk_strategy);
+
+ cost.startup = 0;
+ cost.per_tuple = 0;
+ add_function_cost(root, get_opcode(sortop), NULL, &cost);
+ /* we need procost, not product of procost and cpu_operator_cost */
+ funcCost = cost.per_tuple / cpu_operator_cost;
+ prev_datatype = em->em_datatype;
+ }
+ }
+ else
+ funcCost = 1.0; /* fallback */
+
+ /* Try to take into account actual width fee */
+ funcCost *= get_width_cost_multiplier(root, em->em_expr);
+
+ totalFuncCost += funcCost;
+
+ /* Remember if we have a fake var in pathkeys */
+ has_fake_var |= is_fake_var(em->em_expr);
+ pathkeyExprs = lappend(pathkeyExprs, em->em_expr);
+
+ /*
+ * Prevent call estimate_num_groups() with fake Var. Note,
+ * pathkeyExprs contains only previous columns
+ */
+ if (has_fake_var == false)
+ /*
+ * Recursively compute number of groups in a group from previous step
+ */
+ nGroups = estimate_num_groups_incremental(root, pathkeyExprs,
+ tuplesPerPrevGroup, NULL, NULL,
+ &cache_varinfos,
+ list_length(pathkeyExprs) - 1);
+ else if (tuples > 4.0)
+ /*
+ * Use geometric mean as estimation if there is no any stats.
+ * Don't use DEFAULT_NUM_DISTINCT because it used for only one
+ * column while here we try to estimate number of groups over
+ * set of columns.
+ *
+ * XXX Perhaps this should use DEFAULT_NUM_DISTINCT at least to
+ * limit the calculated values, somehow?
+ *
+ * XXX What's the logic of the following formula?
+ */
+ nGroups = ceil(2.0 + sqrt(tuples) * (i + 1) / list_length(pathkeys));
+ else
+ nGroups = tuples;
+
+ /*
+ * Presorted keys aren't participated in comparison but still checked
+ * by qsort comparator.
+ */
+ if (i >= nPresortedKeys)
+ {
+ if (heapSort)
+ {
+ double heap_tuples;
+
+ /* have to keep at least one group, and a multiple of group size */
+ heap_tuples = Max(ceil(output_tuples / tuplesPerPrevGroup) * tuplesPerPrevGroup,
+ tuplesPerPrevGroup);
+
+ /* so how many (whole) groups is that? */
+ correctedNGroups = ceil(heap_tuples / tuplesPerPrevGroup);
+ }
+ else
+ /* all groups in the input */
+ correctedNGroups = nGroups;
+
+ correctedNGroups = Max(1.0, ceil(correctedNGroups));
+
+ per_tuple_cost += totalFuncCost * LOG2(correctedNGroups);
+ }
+
+ i++;
+
+ /*
+ * Real-world distribution isn't uniform but now we don't have a way to
+ * determine that, so, add multiplier to get closer to worst case.
+ * But ensure the number of tuples does not exceed the group size in the
+ * preceding step.
+ */
+ tuplesPerPrevGroup = Min(tuplesPerPrevGroup,
+ ceil(1.5 * tuplesPerPrevGroup / nGroups));
+
+ /*
+ * We could skip all following columns for cost estimation, because we
+ * believe that tuples are unique by the set of previous columns
+ */
+ if (tuplesPerPrevGroup <= 1.0)
+ break;
+ }
+
+ list_free(pathkeyExprs);
+
+ /* per_tuple_cost is in cpu_operator_cost units */
+ per_tuple_cost *= cpu_operator_cost;
+
+ /*
+ * Accordingly to "Introduction to algorithms", Thomas H. Cormen, Charles E.
+ * Leiserson, Ronald L. Rivest, ISBN 0-07-013143-0, quicksort estimation
+ * formula has additional term proportional to number of tuples (See Chapter
+ * 8.2 and Theorem 4.1). It has meaning with low number of tuples,
+ * approximately less that 1e4. Of course, it could be implemented as
+ * additional multiplier under logarithm, but use more complicated formula
+ * which takes into account number of unique tuples and it isn't clear how
+ * to combine multiplier with groups. Estimate it as 10 in cpu_operator_cost
+ * unit.
+ */
+ per_tuple_cost += 10 * cpu_operator_cost;
+
+ per_tuple_cost += comparison_cost;
+
+ return tuples * per_tuple_cost;
+}
+
+/*
+ * simple wrapper just to estimate best sort path
+ */
+Cost
+cost_sort_estimate(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
+ double tuples)
+{
+ return compute_cpu_sort_cost(root, pathkeys, nPresortedKeys,
+ 0, tuples, tuples, false);
+}
+
/*
* cost_tuplesort
* Determines and returns the cost of sorting a relation using tuplesort,
@@ -1771,7 +2090,7 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
* number of initial runs formed and M is the merge order used by tuplesort.c.
* Since the average initial run should be about sort_mem, we have
* disk traffic = 2 * relsize * ceil(logM(p / sort_mem))
- * cpu = comparison_cost * t * log2(t)
+ * and cpu cost (computed by compute_cpu_sort_cost()).
*
* If the sort is bounded (i.e., only the first k result tuples are needed)
* and k tuples can fit into sort_mem, we use a heap method that keeps only
@@ -1790,9 +2109,11 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
* 'comparison_cost' is the extra cost per comparison, if any
* 'sort_mem' is the number of kilobytes of work memory allowed for the sort
* 'limit_tuples' is the bound on the number of output tuples; -1 if no bound
+ * 'startup_cost' is expected to be 0 at input. If there is "input cost" it should
+ * be added by caller later.
*/
static void
-cost_tuplesort(Cost *startup_cost, Cost *run_cost,
+cost_tuplesort(PlannerInfo *root, List *pathkeys, Cost *startup_cost, Cost *run_cost,
double tuples, int width,
Cost comparison_cost, int sort_mem,
double limit_tuples)
@@ -1809,9 +2130,6 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
if (tuples < 2.0)
tuples = 2.0;
- /* Include the default cost-per-comparison */
- comparison_cost += 2.0 * cpu_operator_cost;
-
/* Do we have a useful LIMIT? */
if (limit_tuples > 0 && limit_tuples < tuples)
{
@@ -1835,12 +2153,10 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
double log_runs;
double npageaccesses;
- /*
- * CPU costs
- *
- * Assume about N log2 N comparisons
- */
- *startup_cost = comparison_cost * tuples * LOG2(tuples);
+ /* CPU costs */
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ tuples, false);
/* Disk costs */
@@ -1856,18 +2172,17 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
}
else if (tuples > 2 * output_tuples || input_bytes > sort_mem_bytes)
{
- /*
- * We'll use a bounded heap-sort keeping just K tuples in memory, for
- * a total number of tuple comparisons of N log2 K; but the constant
- * factor is a bit higher than for quicksort. Tweak it so that the
- * cost curve is continuous at the crossover point.
- */
- *startup_cost = comparison_cost * tuples * LOG2(2.0 * output_tuples);
+ /* We'll use a bounded heap-sort keeping just K tuples in memory. */
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ output_tuples, true);
}
else
{
/* We'll use plain quicksort on all the input tuples */
- *startup_cost = comparison_cost * tuples * LOG2(tuples);
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ tuples, false);
}
/*
@@ -1900,8 +2215,8 @@ cost_incremental_sort(Path *path,
double input_tuples, int width, Cost comparison_cost, int sort_mem,
double limit_tuples)
{
- Cost startup_cost = 0,
- run_cost = 0,
+ Cost startup_cost,
+ run_cost,
input_run_cost = input_total_cost - input_startup_cost;
double group_tuples,
input_groups;
@@ -1986,7 +2301,7 @@ cost_incremental_sort(Path *path,
* pessimistic about incremental sort performance and increase its average
* group size by half.
*/
- cost_tuplesort(&group_startup_cost, &group_run_cost,
+ cost_tuplesort(root, pathkeys, &group_startup_cost, &group_run_cost,
1.5 * group_tuples, width, comparison_cost, sort_mem,
limit_tuples);
@@ -1994,7 +2309,7 @@ cost_incremental_sort(Path *path,
* Startup cost of incremental sort is the startup cost of its first group
* plus the cost of its input.
*/
- startup_cost += group_startup_cost
+ startup_cost = group_startup_cost
+ input_startup_cost + group_input_run_cost;
/*
@@ -2003,7 +2318,7 @@ cost_incremental_sort(Path *path,
* group, plus the total cost to process the remaining groups, plus the
* remaining cost of input.
*/
- run_cost += group_run_cost
+ run_cost = group_run_cost
+ (group_run_cost + group_startup_cost) * (input_groups - 1)
+ group_input_run_cost * (input_groups - 1);
@@ -2043,7 +2358,7 @@ cost_sort(Path *path, PlannerInfo *root,
Cost startup_cost;
Cost run_cost;
- cost_tuplesort(&startup_cost, &run_cost,
+ cost_tuplesort(root, pathkeys, &startup_cost, &run_cost,
tuples, width,
comparison_cost, sort_mem,
limit_tuples);
@@ -2141,7 +2456,7 @@ append_nonpartial_cost(List *subpaths, int numpaths, int parallel_workers)
* Determines and returns the cost of an Append node.
*/
void
-cost_append(AppendPath *apath)
+cost_append(AppendPath *apath, PlannerInfo *root)
{
ListCell *l;
@@ -2209,7 +2524,7 @@ cost_append(AppendPath *apath)
* any child.
*/
cost_sort(&sort_path,
- NULL, /* doesn't currently need root */
+ root,
pathkeys,
subpath->total_cost,
subpath->rows,
diff --git a/src/backend/optimizer/path/equivclass.c b/src/backend/optimizer/path/equivclass.c
index 8c6770de972..5487ae2ee4c 100644
--- a/src/backend/optimizer/path/equivclass.c
+++ b/src/backend/optimizer/path/equivclass.c
@@ -681,7 +681,18 @@ get_eclass_for_sort_expr(PlannerInfo *root,
if (opcintype == cur_em->em_datatype &&
equal(expr, cur_em->em_expr))
- return cur_ec; /* Match! */
+ {
+ /*
+ * Match!
+ *
+ * Copy sortref if it wasn't set yet, it's possible if ec was
+ * constructed from WHERE clause, ie it doesn't have target
+ * reference at all
+ */
+ if (cur_ec->ec_sortref == 0 && sortref > 0)
+ cur_ec->ec_sortref = sortref;
+ return cur_ec;
+ }
}
}
diff --git a/src/backend/optimizer/path/pathkeys.c b/src/backend/optimizer/path/pathkeys.c
index 86a35cdef17..f18d4c8e4e7 100644
--- a/src/backend/optimizer/path/pathkeys.c
+++ b/src/backend/optimizer/path/pathkeys.c
@@ -17,16 +17,19 @@
*/
#include "postgres.h"
+#include "miscadmin.h"
#include "access/stratnum.h"
#include "catalog/pg_opfamily.h"
#include "nodes/makefuncs.h"
#include "nodes/nodeFuncs.h"
#include "nodes/plannodes.h"
+#include "optimizer/cost.h"
#include "optimizer/optimizer.h"
#include "optimizer/pathnode.h"
#include "optimizer/paths.h"
#include "partitioning/partbounds.h"
#include "utils/lsyscache.h"
+#include "utils/selfuncs.h"
static bool pathkey_is_redundant(PathKey *new_pathkey, List *pathkeys);
@@ -334,6 +337,530 @@ pathkeys_contained_in(List *keys1, List *keys2)
return false;
}
+/************************<DEBUG PART>*************************************/
+bool debug_group_by_reorder_by_pathkeys = true;
+bool debug_group_by_match_order_by = true;
+bool debug_cheapest_group_by = true;
+/************************</DEBUG PART>************************************/
+
+/*
+ * group_keys_reorder_by_pathkeys
+ * Reorder GROUP BY keys to match pathkeys of input path.
+ *
+ * Function returns new lists (pathkeys and clauses), original GROUP BY lists
+ * stay untouched.
+ *
+ * Returns the number of GROUP BY keys with a matching pathkey.
+ */
+int
+group_keys_reorder_by_pathkeys(List *pathkeys, List **group_pathkeys,
+ List **group_clauses)
+{
+ List *new_group_pathkeys= NIL,
+ *new_group_clauses = NIL;
+ ListCell *lc;
+ int n;
+
+ if (debug_group_by_reorder_by_pathkeys == false)
+ return 0;
+
+ if (pathkeys == NIL || *group_pathkeys == NIL)
+ return 0;
+
+ /*
+ * Walk the pathkeys (determining ordering of the input path) and see if
+ * there's a matching GROUP BY key. If we find one, we append if to the
+ * list, and do the same for the clauses.
+ *
+ * Once we find first pathkey without a matching GROUP BY key, the rest of
+ * the pathkeys is useless and can't be used to evaluate the grouping, so
+ * we abort the loop and ignore the remaining pathkeys.
+ *
+ * XXX Pathkeys are built in a way to allow simply comparing pointers.
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(lc);
+ SortGroupClause *sgc;
+
+ /* abort on first mismatch */
+ if (!list_member_ptr(*group_pathkeys, pathkey))
+ break;
+
+ new_group_pathkeys = lappend(new_group_pathkeys, pathkey);
+
+ sgc = get_sortgroupref_clause(pathkey->pk_eclass->ec_sortref,
+ *group_clauses);
+
+ new_group_clauses = lappend(new_group_clauses, sgc);
+ }
+
+ /* remember the number of pathkeys with a matching GROUP BY key */
+ n = list_length(new_group_pathkeys);
+
+ /* XXX maybe return when (n == 0) */
+
+ /* append the remaining group pathkeys (will be treated as not sorted) */
+ *group_pathkeys = list_concat_unique_ptr(new_group_pathkeys,
+ *group_pathkeys);
+ *group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+
+ return n;
+}
+
+/*
+ * Used to generate all permutations of a pathkey list.
+ */
+typedef struct PathkeyMutatorState {
+ List *elemsList;
+ ListCell **elemCells;
+ void **elems;
+ int *positions;
+ int mutatorNColumns;
+ int count;
+} PathkeyMutatorState;
+
+
+/*
+ * PathkeyMutatorInit
+ * Initialize state of the permutation generator.
+ *
+ * We want to generate permutations of elements in the "elems" list. We may want
+ * to skip some number of elements at the beginning (when treating as presorted)
+ * or at the end (we only permute a limited number of group keys).
+ *
+ * The list is decomposed into elements, and we also keep pointers to individual
+ * cells. This allows us to build the permuted list quickly and cheaply, without
+ * creating any copies.
+ */
+static void
+PathkeyMutatorInit(PathkeyMutatorState *state, List *elems, int start, int end)
+{
+ int i;
+ int n = end - start;
+ ListCell *lc;
+
+ memset(state, 0, sizeof(*state));
+
+ state->mutatorNColumns = n;
+
+ state->elemsList = list_copy(elems);
+
+ state->elems = palloc(sizeof(void*) * n);
+ state->elemCells = palloc(sizeof(ListCell*) * n);
+ state->positions = palloc(sizeof(int) * n);
+
+ i = 0;
+ for_each_cell(lc, state->elemsList, list_nth_cell(state->elemsList, start))
+ {
+ state->elemCells[i] = lc;
+ state->elems[i] = lfirst(lc);
+ state->positions[i] = i + 1;
+ i++;
+
+ if (i >= n)
+ break;
+ }
+}
+
+/* Swap two elements of an array. */
+static void
+PathkeyMutatorSwap(int *a, int i, int j)
+{
+ int s = a[i];
+
+ a[i] = a[j];
+ a[j] = s;
+}
+
+/*
+ * Generate the next permutation of elements.
+ */
+static bool
+PathkeyMutatorNextSet(int *a, int n)
+{
+ int j, k, l, r;
+
+ j = n - 2;
+
+ while (j >= 0 && a[j] >= a[j + 1])
+ j--;
+
+ if (j < 0)
+ return false;
+
+ k = n - 1;
+
+ while (k >= 0 && a[j] >= a[k])
+ k--;
+
+ PathkeyMutatorSwap(a, j, k);
+
+ l = j + 1;
+ r = n - 1;
+
+ while (l < r)
+ PathkeyMutatorSwap(a, l++, r--);
+
+ return true;
+}
+
+/*
+ * PathkeyMutatorNext
+ * Generate the next permutation of list of elements.
+ *
+ * Returns the next permutation (as a list of elements) or NIL if there are no
+ * more permutations.
+ */
+static List *
+PathkeyMutatorNext(PathkeyMutatorState *state)
+{
+ int i;
+
+ state->count++;
+
+ /* first permutation is original list */
+ if (state->count == 1)
+ return state->elemsList;
+
+ /* when there are no more permutations, return NIL */
+ if (!PathkeyMutatorNextSet(state->positions, state->mutatorNColumns))
+ {
+ pfree(state->elems);
+ pfree(state->elemCells);
+ pfree(state->positions);
+
+ list_free(state->elemsList);
+
+ return NIL;
+ }
+
+ /* update the list cells to point to the right elements */
+ for(i=0; i<state->mutatorNColumns; i++)
+ lfirst(state->elemCells[i]) =
+ (void *) state->elems[ state->positions[i] - 1 ];
+
+ return state->elemsList;
+}
+
+/*
+ * Cost of comparing pathkeys.
+ */
+typedef struct PathkeySortCost
+{
+ Cost cost;
+ PathKey *pathkey;
+} PathkeySortCost;
+
+static int
+pathkey_sort_cost_comparator(const void *_a, const void *_b)
+{
+ const PathkeySortCost *a = (PathkeySortCost *) _a;
+ const PathkeySortCost *b = (PathkeySortCost *) _b;
+
+ if (a->cost < b->cost)
+ return -1;
+ else if (a->cost == b->cost)
+ return 0;
+ return 1;
+}
+
+/*
+ * get_cheapest_group_keys_order
+ * Returns the pathkeys in an order cheapest to evaluate.
+ *
+ * Given a list of pathkeys, we try to reorder them in a way that minimizes
+ * the CPU cost of sorting. This depends mainly on the cost of comparator
+ * function (the pathkeys may use different data types) and the number of
+ * distinct values in each column (which affects the number of comparator
+ * calls for the following pathkeys).
+ *
+ * In case the input is partially sorted, only the remaining pathkeys are
+ * considered.
+ *
+ * Returns newly allocated lists. If no reordering is possible (or needed),
+ * the lists are set to NIL.
+ */
+static bool
+get_cheapest_group_keys_order(PlannerInfo *root, double nrows,
+ List **group_pathkeys, List **group_clauses,
+ int n_preordered)
+{
+ List *new_group_pathkeys = NIL,
+ *new_group_clauses = NIL,
+ *var_group_pathkeys;
+
+ ListCell *cell;
+ PathkeyMutatorState mstate;
+ double cheapest_sort_cost = -1.0;
+
+ int nFreeKeys;
+ int nToPermute;
+
+ /* If this optimization is disabled, we're done. */
+ if (!debug_cheapest_group_by)
+ return false;
+
+ /* If there are less than 2 unsorted pathkeys, we're done. */
+ if (list_length(*group_pathkeys) - n_preordered < 2)
+ return false;
+
+ /*
+ * We could exhaustively cost all possible orderings of the pathkeys, but for
+ * large number of pathkeys that might be prohibitively expensive. So we try
+ * to apply a simple cheap heuristics first - we sort the pathkeys by sort
+ * cost (as if the pathkey was sorted independently) and then check only the
+ * four cheapest pathkeys. The remaining pathkeys are kept ordered by cost.
+ *
+ * XXX This is a very simple heuristics, and likely to work fine for most
+ * cases (because number of GROUP BY clauses tends to be lower than 4). But
+ * it ignores how the number of distinct values in each pathkey affects the
+ * following sorts. It may be better to use "more expensive" pathkey first
+ * if it has many distinct values, because it then limits the number of
+ * comparisons for the remaining pathkeys. But evaluating that is kinda the
+ * expensive bit we're trying to not do.
+ */
+ nFreeKeys = list_length(*group_pathkeys) - n_preordered;
+ nToPermute = 4;
+ if (nFreeKeys > nToPermute)
+ {
+ int i;
+ PathkeySortCost *costs = palloc(sizeof(PathkeySortCost) * nFreeKeys);
+
+ /* skip the pre-ordered pathkeys */
+ cell = list_nth_cell(*group_pathkeys, n_preordered);
+
+ /* estimate cost for sorting individual pathkeys */
+ for (i = 0; cell != NULL; i++, (cell = lnext(*group_pathkeys, cell)))
+ {
+ List *to_cost = list_make1(lfirst(cell));
+
+ Assert(i < nFreeKeys);
+
+ costs[i].pathkey = lfirst(cell);
+ costs[i].cost = cost_sort_estimate(root, to_cost, 0, nrows);
+
+ pfree(to_cost);
+ }
+
+ /* sort the pathkeys by sort cost in ascending order */
+ qsort(costs, nFreeKeys, sizeof(*costs), pathkey_sort_cost_comparator);
+
+ /*
+ * Rebuild the list of pathkeys - first the preordered ones, then the
+ * rest ordered by cost.
+ */
+ new_group_pathkeys = list_truncate(list_copy(*group_pathkeys), n_preordered);
+
+ for (i = 0; i < nFreeKeys; i++)
+ new_group_pathkeys = lappend(new_group_pathkeys, costs[i].pathkey);
+
+ pfree(costs);
+ }
+ else
+ {
+ /*
+ * Since v13 list_free() can clean list elements so for original list
+ * not to be modified it should be copied to a new one which can then
+ * be cleaned safely if needed.
+ */
+ new_group_pathkeys = list_copy(*group_pathkeys);
+ nToPermute = nFreeKeys;
+ }
+
+ Assert(list_length(new_group_pathkeys) == list_length(*group_pathkeys));
+
+ /*
+ * Generate pathkey lists with permutations of the first nToPermute pathkeys.
+ *
+ * XXX We simply calculate sort cost for each individual pathkey list, but
+ * there's room for two dynamic programming optimizations here. Firstly, we
+ * may pass the current "best" cost to cost_sort_estimate so that it can
+ * "abort" if the estimated pathkeys list exceeds it. Secondly, it could pass
+ * return information about the position when it exceeded the cost, and we
+ * could skip all permutations with the same prefix.
+ *
+ * Imagine we've already found ordering with cost C1, and we're evaluating
+ * another ordering - cost_sort_estimate() calculates cost by adding the
+ * pathkeys one by one (more or less), and the cost only grows. If at any
+ * point it exceeds C1, it can't possibly be "better" so we can discard it.
+ * But we also know that we can discard all ordering with the same prefix,
+ * because if we're estimating (a,b,c,d) and we exceeded C1 at (a,b) then
+ * the same thing will happen for any ordering with this prefix.
+ *
+ *
+ */
+ PathkeyMutatorInit(&mstate, new_group_pathkeys, n_preordered, n_preordered + nToPermute);
+
+ while((var_group_pathkeys = PathkeyMutatorNext(&mstate)) != NIL)
+ {
+ Cost cost;
+
+ cost = cost_sort_estimate(root, var_group_pathkeys, n_preordered, nrows);
+
+ if (cost < cheapest_sort_cost || cheapest_sort_cost < 0)
+ {
+ list_free(new_group_pathkeys);
+ new_group_pathkeys = list_copy(var_group_pathkeys);
+ cheapest_sort_cost = cost;
+ }
+ }
+
+ /* Reorder the group clauses according to the reordered pathkeys. */
+ foreach(cell, new_group_pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(cell);
+
+ new_group_clauses = lappend(new_group_clauses,
+ get_sortgroupref_clause(pathkey->pk_eclass->ec_sortref,
+ *group_clauses));
+ }
+
+ /* Just append the rest GROUP BY clauses */
+ new_group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+
+ *group_pathkeys = new_group_pathkeys;
+ *group_clauses = new_group_clauses;
+
+ return true;
+}
+
+/*
+ * get_useful_group_keys_orderings
+ * Determine which orderings of GROUP BY keys are potentially interesting.
+ *
+ * Returns list of PathKeyInfo items, each representing an interesting ordering
+ * of GROUP BY keys. Each items stores pathkeys and clauses in matching order.
+ *
+ * The function considers (and keeps) multiple group by orderings:
+ *
+ * - the original ordering, as specified by the GROUP BY clause
+ *
+ * - GROUP BY keys reordered to minimize the sort cost
+ *
+ * - GROUP BY keys reordered to match path ordering (as much as possible), with
+ * the tail reoredered to minimize the sort cost
+ *
+ * - GROUP BY keys to match target ORDER BY clause (as much as possible), with
+ * the tail reoredered to minimize the sort cost
+ *
+ * There are other potentially interesting orderings (e.g. it might be best to
+ * match the first ORDER BY key, order the remaining keys differently and then
+ * rely on incremental sort to fix this), but we ignore those for now. To make
+ * this work we'd have to pretty much generate all possible permutations.
+ */
+List *
+get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys,
+ List *group_pathkeys, List *group_clauses)
+{
+ Query *parse = root->parse;
+ List *infos = NIL;
+ PathKeyInfo *info;
+ int n_preordered = 0;
+
+ List *pathkeys = group_pathkeys;
+ List *clauses = group_clauses;
+
+ /* always return at least the original pathkeys/clauses */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+
+ /* for grouping sets we can't do any reordering */
+ if (parse->groupingSets)
+ return infos;
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost, ignoring both the
+ * target ordering (ORDER BY) and ordering of the input path.
+ */
+ if (get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered))
+ {
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * If the path is sorted in some way, try reordering the group keys to match
+ * as much of the ordering as possible - we get this sort for free (mostly).
+ *
+ * We must not do this when there are no grouping sets, because those use
+ * more complex logic to decide the ordering.
+ *
+ * XXX Isn't this somewhat redundant with presorted_keys? Actually, it's
+ * more a complement, because it allows benefiting from incremental sort
+ * as much as possible.
+ *
+ * XXX This does nothing if (n_preordered == 0). We shouldn't create the
+ * info in this case.
+ */
+ if (path_pathkeys)
+ {
+ n_preordered = group_keys_reorder_by_pathkeys(path_pathkeys,
+ &pathkeys,
+ &clauses);
+
+ /* reorder the tail to minimize sort cost */
+ get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, in
+ * which case get_cheapest_group_keys_order returns false. But we
+ * still want to keep the keys reordered to path_pathkeys.
+ */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost (this time consider
+ * the ORDER BY clause, but only if set debug_group_by_match_order_by).
+ *
+ * XXX This does nothing if (n_preordered == 0). We shouldn't create the
+ * info in this case.
+ */
+ if (root->sort_pathkeys && debug_group_by_match_order_by)
+ {
+ n_preordered = group_keys_reorder_by_pathkeys(root->sort_pathkeys,
+ &pathkeys,
+ &clauses);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, in
+ * which case get_cheapest_group_keys_order returns false. But we
+ * still want to keep the keys reordered to sort_pathkeys.
+ */
+ get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered);
+
+ /* keep the group keys reordered to match ordering of input path */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ return infos;
+}
+
/*
* pathkeys_count_contained_in
* Same as pathkeys_contained_in, but also sets length of longest
@@ -1862,6 +2389,54 @@ pathkeys_useful_for_ordering(PlannerInfo *root, List *pathkeys)
return n_common_pathkeys;
}
+/*
+ * pathkeys_useful_for_grouping
+ * Count the number of pathkeys that are useful for grouping (instead of
+ * explicit sort)
+ *
+ * Group pathkeys could be reordered to benefit from the odering. The ordering
+ * may not be "complete" and may require incremental sort, but that's fine. So
+ * we simply count prefix pathkeys with a matching group key, and stop once we
+ * find the first pathkey without a match.
+ *
+ * So e.g. with pathkeys (a,b,c) and group keys (a,b,e) this determines (a,b)
+ * pathkeys are useful for grouping, and we might do incremental sort to get
+ * path ordered by (a,b,e).
+ *
+ * This logic is necessary to retain paths with ordeding not matching grouping
+ * keys directly, without the reordering.
+ *
+ * Returns the length of pathkey prefix with matching group keys.
+ */
+static int
+pathkeys_useful_for_grouping(PlannerInfo *root, List *pathkeys)
+{
+ ListCell *key;
+ int n = 0;
+
+ /* no special ordering requested for grouping */
+ if (root->group_pathkeys == NIL)
+ return 0;
+
+ /* unordered path */
+ if (pathkeys == NIL)
+ return 0;
+
+ /* walk the pathkeys and search for matching group key */
+ foreach(key, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(key);
+
+ /* no matching group key, we're done */
+ if (!list_member_ptr(root->group_pathkeys, pathkey))
+ break;
+
+ n++;
+ }
+
+ return n;
+}
+
/*
* truncate_useless_pathkeys
* Shorten the given pathkey list to just the useful pathkeys.
@@ -1876,6 +2451,9 @@ truncate_useless_pathkeys(PlannerInfo *root,
nuseful = pathkeys_useful_for_merging(root, rel, pathkeys);
nuseful2 = pathkeys_useful_for_ordering(root, pathkeys);
+ if (nuseful2 > nuseful)
+ nuseful = nuseful2;
+ nuseful2 = pathkeys_useful_for_grouping(root, pathkeys);
if (nuseful2 > nuseful)
nuseful = nuseful2;
@@ -1911,6 +2489,8 @@ has_useful_pathkeys(PlannerInfo *root, RelOptInfo *rel)
{
if (rel->joininfo != NIL || rel->has_eclass_joins)
return true; /* might be able to use pathkeys for merging */
+ if (root->group_pathkeys != NIL)
+ return true; /* might be able to use pathkeys for grouping */
if (root->query_pathkeys != NIL)
return true; /* might be able to use them for ordering */
return false; /* definitely useless */
diff --git a/src/backend/optimizer/plan/planner.c b/src/backend/optimizer/plan/planner.c
index 547fda20a23..bb957d50d23 100644
--- a/src/backend/optimizer/plan/planner.c
+++ b/src/backend/optimizer/plan/planner.c
@@ -6218,24 +6218,124 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
+ Path *path_save = path;
Path *path_original = path;
- bool is_sorted;
- int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ List *pathkey_orderings = NIL;
+
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
+
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
- if (path == cheapest_path || is_sorted)
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- /* Sort the cheapest-total path if it isn't already sorted */
- if (!is_sorted)
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
+
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (path == cheapest_path || is_sorted)
+ {
+ /* Sort the cheapest-total path if it isn't already sorted */
+ if (!is_sorted)
+ {
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ }
+
+ /* Now decide what to stick atop it */
+ if (parse->groupingSets)
+ {
+ consider_groupingsets_paths(root, grouped_rel,
+ path, true, can_hash,
+ gd, agg_costs, dNumGroups);
+ }
+ else if (parse->hasAggs)
+ {
+ /*
+ * We have aggregation, possibly with plain GROUP BY. Make
+ * an AggPath.
+ */
+ add_path(grouped_rel, (Path *)
+ create_agg_path(root,
+ grouped_rel,
+ path,
+ grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_SIMPLE,
+ info->clauses,
+ havingQual,
+ agg_costs,
+ dNumGroups));
+ }
+ else if (group_clauses)
+ {
+ /*
+ * We have GROUP BY without aggregation or grouping sets.
+ * Make a GroupPath.
+ */
+ add_path(grouped_rel, (Path *)
+ create_group_path(root,
+ grouped_rel,
+ path,
+ info->clauses,
+ havingQual,
+ dNumGroups));
+ }
+ else
+ {
+ /* Other cases should have been handled above */
+ Assert(false);
+ }
+ }
+
+ /*
+ * Now we may consider incremental sort on this path, but only
+ * when the path is not already sorted and when incremental sort
+ * is enabled.
+ */
+ if (is_sorted || !enable_incremental_sort)
+ continue;
+
+ /* Restore the input path (we might have added Sort on top). */
+ path = path_original;
+
+ /* no shared prefix, no point in building incremental sort */
+ if (presorted_keys == 0)
+ continue;
+
+ /*
+ * We should have already excluded pathkeys of length 1 because
+ * then presorted_keys > 0 would imply is_sorted was true.
+ */
+ Assert(list_length(root->group_pathkeys) != 1);
+
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
/* Now decide what to stick atop it */
if (parse->groupingSets)
@@ -6247,17 +6347,17 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
else if (parse->hasAggs)
{
/*
- * We have aggregation, possibly with plain GROUP BY. Make
- * an AggPath.
+ * We have aggregation, possibly with plain GROUP BY. Make an
+ * AggPath.
*/
add_path(grouped_rel, (Path *)
create_agg_path(root,
grouped_rel,
path,
grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_SIMPLE,
- parse->groupClause,
+ info->clauses,
havingQual,
agg_costs,
dNumGroups));
@@ -6265,14 +6365,14 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
else if (parse->groupClause)
{
/*
- * We have GROUP BY without aggregation or grouping sets.
- * Make a GroupPath.
+ * We have GROUP BY without aggregation or grouping sets. Make
+ * a GroupPath.
*/
add_path(grouped_rel, (Path *)
create_group_path(root,
grouped_rel,
path,
- parse->groupClause,
+ info->clauses,
havingQual,
dNumGroups));
}
@@ -6282,79 +6382,6 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
Assert(false);
}
}
-
- /*
- * Now we may consider incremental sort on this path, but only
- * when the path is not already sorted and when incremental sort
- * is enabled.
- */
- if (is_sorted || !enable_incremental_sort)
- continue;
-
- /* Restore the input path (we might have added Sort on top). */
- path = path_original;
-
- /* no shared prefix, no point in building incremental sort */
- if (presorted_keys == 0)
- continue;
-
- /*
- * We should have already excluded pathkeys of length 1 because
- * then presorted_keys > 0 would imply is_sorted was true.
- */
- Assert(list_length(root->group_pathkeys) != 1);
-
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
-
- /* Now decide what to stick atop it */
- if (parse->groupingSets)
- {
- consider_groupingsets_paths(root, grouped_rel,
- path, true, can_hash,
- gd, agg_costs, dNumGroups);
- }
- else if (parse->hasAggs)
- {
- /*
- * We have aggregation, possibly with plain GROUP BY. Make an
- * AggPath.
- */
- add_path(grouped_rel, (Path *)
- create_agg_path(root,
- grouped_rel,
- path,
- grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_SIMPLE,
- parse->groupClause,
- havingQual,
- agg_costs,
- dNumGroups));
- }
- else if (parse->groupClause)
- {
- /*
- * We have GROUP BY without aggregation or grouping sets. Make
- * a GroupPath.
- */
- add_path(grouped_rel, (Path *)
- create_group_path(root,
- grouped_rel,
- path,
- parse->groupClause,
- havingQual,
- dNumGroups));
- }
- else
- {
- /* Other cases should have been handled above */
- Assert(false);
- }
}
/*
@@ -6365,100 +6392,125 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
{
foreach(lc, partially_grouped_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
Path *path_original = path;
- bool is_sorted;
- int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ List *pathkey_orderings = NIL;
- /*
- * Insert a Sort node, if required. But there's no point in
- * sorting anything but the cheapest path.
- */
- if (!is_sorted)
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
+
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- if (path != partially_grouped_rel->cheapest_total_path)
- continue;
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- }
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- if (parse->hasAggs)
- add_path(grouped_rel, (Path *)
- create_agg_path(root,
- grouped_rel,
- path,
- grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_FINAL_DESERIAL,
- parse->groupClause,
- havingQual,
- agg_final_costs,
- dNumGroups));
- else
- add_path(grouped_rel, (Path *)
- create_group_path(root,
- grouped_rel,
- path,
- parse->groupClause,
- havingQual,
- dNumGroups));
+ /* restore the path (we replace it in the loop) */
+ path = path_original;
- /*
- * Now we may consider incremental sort on this path, but only
- * when the path is not already sorted and when incremental
- * sort is enabled.
- */
- if (is_sorted || !enable_incremental_sort)
- continue;
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
- /* Restore the input path (we might have added Sort on top). */
- path = path_original;
+ /*
+ * Insert a Sort node, if required. But there's no point in
+ * sorting anything but the cheapest path.
+ */
+ if (!is_sorted)
+ {
+ if (path != partially_grouped_rel->cheapest_total_path)
+ continue;
- /* no shared prefix, not point in building incremental sort */
- if (presorted_keys == 0)
- continue;
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ }
- /*
- * We should have already excluded pathkeys of length 1
- * because then presorted_keys > 0 would imply is_sorted was
- * true.
- */
- Assert(list_length(root->group_pathkeys) != 1);
+ if (parse->hasAggs)
+ add_path(grouped_rel, (Path *)
+ create_agg_path(root,
+ grouped_rel,
+ path,
+ grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_FINAL_DESERIAL,
+ info->clauses,
+ havingQual,
+ agg_final_costs,
+ dNumGroups));
+ else
+ add_path(grouped_rel, (Path *)
+ create_group_path(root,
+ grouped_rel,
+ path,
+ info->clauses,
+ havingQual,
+ dNumGroups));
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
+ /*
+ * Now we may consider incremental sort on this path, but only
+ * when the path is not already sorted and when incremental
+ * sort is enabled.
+ */
+ if (is_sorted || !enable_incremental_sort)
+ continue;
- if (parse->hasAggs)
- add_path(grouped_rel, (Path *)
- create_agg_path(root,
- grouped_rel,
- path,
- grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_FINAL_DESERIAL,
- parse->groupClause,
- havingQual,
- agg_final_costs,
- dNumGroups));
- else
- add_path(grouped_rel, (Path *)
- create_group_path(root,
- grouped_rel,
- path,
- parse->groupClause,
- havingQual,
- dNumGroups));
+ /* Restore the input path (we might have added Sort on top). */
+ path = path_original;
+
+ /* no shared prefix, not point in building incremental sort */
+ if (presorted_keys == 0)
+ continue;
+
+ /*
+ * We should have already excluded pathkeys of length 1
+ * because then presorted_keys > 0 would imply is_sorted was
+ * true.
+ */
+ Assert(list_length(root->group_pathkeys) != 1);
+
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+
+ if (parse->hasAggs)
+ add_path(grouped_rel, (Path *)
+ create_agg_path(root,
+ grouped_rel,
+ path,
+ grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_FINAL_DESERIAL,
+ info->clauses,
+ havingQual,
+ agg_final_costs,
+ dNumGroups));
+ else
+ add_path(grouped_rel, (Path *)
+ create_group_path(root,
+ grouped_rel,
+ path,
+ info->clauses,
+ havingQual,
+ dNumGroups));
+ }
}
}
}
@@ -6661,41 +6713,71 @@ create_partial_grouping_paths(PlannerInfo *root,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
+ Path *path_save = path;
+
+ List *pathkey_orderings = NIL;
+
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
+
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
+
+ Assert(list_length(pathkey_orderings) > 0);
- is_sorted = pathkeys_contained_in(root->group_pathkeys,
- path->pathkeys);
- if (path == cheapest_total_path || is_sorted)
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- /* Sort the cheapest partial path, if it isn't already */
- if (!is_sorted)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- if (parse->hasAggs)
- add_path(partially_grouped_rel, (Path *)
- create_agg_path(root,
- partially_grouped_rel,
- path,
- partially_grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_INITIAL_SERIAL,
- parse->groupClause,
- NIL,
- agg_partial_costs,
- dNumPartialGroups));
- else
- add_path(partially_grouped_rel, (Path *)
- create_group_path(root,
- partially_grouped_rel,
- path,
- parse->groupClause,
- NIL,
- dNumPartialGroups));
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
+
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (path == cheapest_total_path || is_sorted)
+ {
+ /* Sort the cheapest partial path, if it isn't already */
+ if (!is_sorted)
+ {
+ path = (Path *) create_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_path(partially_grouped_rel, (Path *)
+ create_agg_path(root,
+ partially_grouped_rel,
+ path,
+ partially_grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_INITIAL_SERIAL,
+ info->clauses,
+ NIL,
+ agg_partial_costs,
+ dNumPartialGroups));
+ else
+ add_path(partially_grouped_rel, (Path *)
+ create_group_path(root,
+ partially_grouped_rel,
+ path,
+ info->clauses,
+ NIL,
+ dNumPartialGroups));
+ }
}
}
@@ -6705,6 +6787,8 @@ create_partial_grouping_paths(PlannerInfo *root,
* We can also skip the entire loop when we only have a single-item
* group_pathkeys because then we can't possibly have a presorted
* prefix of the list without having the list be fully sorted.
+ *
+ * XXX Shouldn't this also consider the group-key-reordering?
*/
if (enable_incremental_sort && list_length(root->group_pathkeys) > 1)
{
@@ -6763,24 +6847,100 @@ create_partial_grouping_paths(PlannerInfo *root,
/* Similar to above logic, but for partial paths. */
foreach(lc, input_rel->partial_pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
Path *path_original = path;
- bool is_sorted;
- int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ List *pathkey_orderings = NIL;
+
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
- if (path == cheapest_partial_path || is_sorted)
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- /* Sort the cheapest partial path, if it isn't already */
- if (!is_sorted)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+ /* restore the path (we replace it in the loop) */
+ path = path_original;
+
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (path == cheapest_partial_path || is_sorted)
+ {
+
+ /* Sort the cheapest partial path, if it isn't already */
+ if (!is_sorted)
+ {
+ path = (Path *) create_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_partial_path(partially_grouped_rel, (Path *)
+ create_agg_path(root,
+ partially_grouped_rel,
+ path,
+ partially_grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_INITIAL_SERIAL,
+ info->clauses,
+ NIL,
+ agg_partial_costs,
+ dNumPartialPartialGroups));
+ else
+ add_partial_path(partially_grouped_rel, (Path *)
+ create_group_path(root,
+ partially_grouped_rel,
+ path,
+ info->clauses,
+ NIL,
+ dNumPartialPartialGroups));
+ }
+
+ /*
+ * Now we may consider incremental sort on this path, but only
+ * when the path is not already sorted and when incremental sort
+ * is enabled.
+ */
+ if (is_sorted || !enable_incremental_sort)
+ continue;
+
+ /* Restore the input path (we might have added Sort on top). */
+ path = path_original;
+
+ /* no shared prefix, not point in building incremental sort */
+ if (presorted_keys == 0)
+ continue;
+
+ /*
+ * We should have already excluded pathkeys of length 1 because
+ * then presorted_keys > 0 would imply is_sorted was true.
+ */
+ Assert(list_length(root->group_pathkeys) != 1);
+
+ path = (Path *) create_incremental_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
if (parse->hasAggs)
add_partial_path(partially_grouped_rel, (Path *)
@@ -6788,9 +6948,9 @@ create_partial_grouping_paths(PlannerInfo *root,
partially_grouped_rel,
path,
partially_grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_INITIAL_SERIAL,
- parse->groupClause,
+ info->clauses,
NIL,
agg_partial_costs,
dNumPartialPartialGroups));
@@ -6799,59 +6959,10 @@ create_partial_grouping_paths(PlannerInfo *root,
create_group_path(root,
partially_grouped_rel,
path,
- parse->groupClause,
+ info->clauses,
NIL,
dNumPartialPartialGroups));
}
-
- /*
- * Now we may consider incremental sort on this path, but only
- * when the path is not already sorted and when incremental sort
- * is enabled.
- */
- if (is_sorted || !enable_incremental_sort)
- continue;
-
- /* Restore the input path (we might have added Sort on top). */
- path = path_original;
-
- /* no shared prefix, not point in building incremental sort */
- if (presorted_keys == 0)
- continue;
-
- /*
- * We should have already excluded pathkeys of length 1 because
- * then presorted_keys > 0 would imply is_sorted was true.
- */
- Assert(list_length(root->group_pathkeys) != 1);
-
- path = (Path *) create_incremental_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
-
- if (parse->hasAggs)
- add_partial_path(partially_grouped_rel, (Path *)
- create_agg_path(root,
- partially_grouped_rel,
- path,
- partially_grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_INITIAL_SERIAL,
- parse->groupClause,
- NIL,
- agg_partial_costs,
- dNumPartialPartialGroups));
- else
- add_partial_path(partially_grouped_rel, (Path *)
- create_group_path(root,
- partially_grouped_rel,
- path,
- parse->groupClause,
- NIL,
- dNumPartialPartialGroups));
}
}
diff --git a/src/backend/optimizer/util/pathnode.c b/src/backend/optimizer/util/pathnode.c
index 99df76b6b71..1670e546440 100644
--- a/src/backend/optimizer/util/pathnode.c
+++ b/src/backend/optimizer/util/pathnode.c
@@ -1342,7 +1342,7 @@ create_append_path(PlannerInfo *root,
pathnode->path.pathkeys = child->pathkeys;
}
else
- cost_append(pathnode);
+ cost_append(pathnode, root);
/* If the caller provided a row estimate, override the computed value. */
if (rows >= 0)
diff --git a/src/backend/utils/adt/selfuncs.c b/src/backend/utils/adt/selfuncs.c
index 1fbb0b28c3b..36c303e45c5 100644
--- a/src/backend/utils/adt/selfuncs.c
+++ b/src/backend/utils/adt/selfuncs.c
@@ -3294,7 +3294,10 @@ add_unique_group_var(PlannerInfo *root, List *varinfos,
}
/*
- * estimate_num_groups - Estimate number of groups in a grouped query
+ * estimate_num_groups/estimate_num_groups_incremental
+ * - Estimate number of groups in a grouped query.
+ * _incremental variant is performance optimization for
+ * case of adding one-by-one column
*
* Given a query having a GROUP BY clause, estimate how many groups there
* will be --- ie, the number of distinct combinations of the GROUP BY
@@ -3368,11 +3371,22 @@ double
estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
List **pgset, EstimationInfo *estinfo)
{
- List *varinfos = NIL;
+ return estimate_num_groups_incremental(root, groupExprs,
+ input_rows, pgset, estinfo,
+ NULL, 0);
+}
+
+double
+estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
+ double input_rows,
+ List **pgset, EstimationInfo *estinfo,
+ List **cache_varinfos, int prevNExprs)
+{
+ List *varinfos = (cache_varinfos) ? *cache_varinfos : NIL;
double srf_multiplier = 1.0;
double numdistinct;
ListCell *l;
- int i;
+ int i, j;
/* Zero the estinfo output parameter, if non-NULL */
if (estinfo != NULL)
@@ -3403,7 +3417,7 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
*/
numdistinct = 1.0;
- i = 0;
+ i = j = 0;
foreach(l, groupExprs)
{
Node *groupexpr = (Node *) lfirst(l);
@@ -3412,6 +3426,14 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
List *varshere;
ListCell *l2;
+ /* was done on previous call */
+ if (cache_varinfos && j++ < prevNExprs)
+ {
+ if (pgset)
+ i++; /* to keep in sync with lines below */
+ continue;
+ }
+
/* is expression in this grouping set? */
if (pgset && !list_member_int(*pgset, i++))
continue;
@@ -3481,7 +3503,11 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
if (varshere == NIL)
{
if (contain_volatile_functions(groupexpr))
+ {
+ if (cache_varinfos)
+ *cache_varinfos = varinfos;
return input_rows;
+ }
continue;
}
@@ -3498,6 +3524,9 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
}
}
+ if (cache_varinfos)
+ *cache_varinfos = varinfos;
+
/*
* If now no Vars, we must have an all-constant or all-boolean GROUP BY
* list.
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index eb3a03b9762..31f347d469d 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2140,6 +2140,38 @@ static struct config_bool ConfigureNamesBool[] =
NULL, NULL, NULL
},
+/************************<DEBUG OPT GROUP BY>*********************************/
+ {
+ {"debug_group_by_reorder_by_pathkeys", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("enable reorder GROUP BY by pathkeys"),
+ NULL,
+ GUC_NOT_IN_SAMPLE
+ },
+ &debug_group_by_reorder_by_pathkeys,
+ true,
+ NULL, NULL, NULL
+ },
+ {
+ {"debug_enable_group_by_match_order_by", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("enable matching GROUP BY by ORDER BY."),
+ NULL,
+ GUC_NOT_IN_SAMPLE
+ },
+ &debug_group_by_match_order_by,
+ true,
+ NULL, NULL, NULL
+ },
+ {
+ {"debug_enable_cheapest_group_by", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("find a cheapest order of columns in GROUP BY."),
+ NULL,
+ GUC_NOT_IN_SAMPLE
+ },
+ &debug_cheapest_group_by,
+ true,
+ NULL, NULL, NULL
+ },
+/************************</DEBUG OPT GROUP BY>********************************/
/* End-of-list marker */
{
{NULL, 0, 0, NULL, NULL}, NULL, false, NULL, NULL, NULL
diff --git a/src/include/nodes/nodes.h b/src/include/nodes/nodes.h
index e8f30367a48..e74a7c72ea6 100644
--- a/src/include/nodes/nodes.h
+++ b/src/include/nodes/nodes.h
@@ -271,6 +271,7 @@ typedef enum NodeTag
T_EquivalenceClass,
T_EquivalenceMember,
T_PathKey,
+ T_PathKeyInfo,
T_PathTarget,
T_RestrictInfo,
T_IndexClause,
diff --git a/src/include/nodes/pathnodes.h b/src/include/nodes/pathnodes.h
index 365000bdcdb..6cbcb67bdf1 100644
--- a/src/include/nodes/pathnodes.h
+++ b/src/include/nodes/pathnodes.h
@@ -1070,6 +1070,16 @@ typedef struct PathKey
bool pk_nulls_first; /* do NULLs come before normal values? */
} PathKey;
+/*
+ * Combines information about pathkeys and the associated clauses.
+ */
+typedef struct PathKeyInfo
+{
+ NodeTag type;
+ List *pathkeys;
+ List *clauses;
+} PathKeyInfo;
+
/*
* VolatileFunctionStatus -- allows nodes to cache their
* contain_volatile_functions properties. VOLATILITY_UNKNOWN means not yet
diff --git a/src/include/optimizer/cost.h b/src/include/optimizer/cost.h
index bc12071af6e..dc7fc174114 100644
--- a/src/include/optimizer/cost.h
+++ b/src/include/optimizer/cost.h
@@ -114,7 +114,9 @@ extern void cost_incremental_sort(Path *path,
Cost input_startup_cost, Cost input_total_cost,
double input_tuples, int width, Cost comparison_cost, int sort_mem,
double limit_tuples);
-extern void cost_append(AppendPath *path);
+extern Cost cost_sort_estimate(PlannerInfo *root, List *pathkeys,
+ int nPresortedKeys, double tuples);
+extern void cost_append(AppendPath *path, PlannerInfo *root);
extern void cost_merge_append(Path *path, PlannerInfo *root,
List *pathkeys, int n_streams,
Cost input_startup_cost, Cost input_total_cost,
diff --git a/src/include/optimizer/paths.h b/src/include/optimizer/paths.h
index 0c3a0b90c85..525b594cbd0 100644
--- a/src/include/optimizer/paths.h
+++ b/src/include/optimizer/paths.h
@@ -203,6 +203,17 @@ typedef enum
extern PathKeysComparison compare_pathkeys(List *keys1, List *keys2);
extern bool pathkeys_contained_in(List *keys1, List *keys2);
extern bool pathkeys_count_contained_in(List *keys1, List *keys2, int *n_common);
+extern int group_keys_reorder_by_pathkeys(List *pathkeys,
+ List **group_pathkeys,
+ List **group_clauses);
+/************************<DEBUG OPT GROUP BY>*********************************/
+extern bool debug_group_by_reorder_by_pathkeys;
+extern bool debug_group_by_match_order_by;
+extern bool debug_cheapest_group_by;
+/************************</DEBUG OPT GROUP BY>********************************/
+extern List *get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys,
+ List *pathkeys, List *clauses);
extern Path *get_cheapest_path_for_pathkeys(List *paths, List *pathkeys,
Relids required_outer,
CostSelector cost_criterion,
diff --git a/src/include/utils/selfuncs.h b/src/include/utils/selfuncs.h
index 8f3d73edfb2..c313a08d541 100644
--- a/src/include/utils/selfuncs.h
+++ b/src/include/utils/selfuncs.h
@@ -214,6 +214,11 @@ extern double estimate_num_groups(PlannerInfo *root, List *groupExprs,
double input_rows, List **pgset,
EstimationInfo *estinfo);
+extern double estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
+ double input_rows, List **pgset,
+ EstimationInfo *estinfo,
+ List **cache_varinfos, int prevNExprs);
+
extern void estimate_hash_bucket_stats(PlannerInfo *root,
Node *hashkey, double nbuckets,
Selectivity *mcv_freq,
diff --git a/src/test/regress/expected/aggregates.out b/src/test/regress/expected/aggregates.out
index 0a23a39aa29..601047fa3dd 100644
--- a/src/test/regress/expected/aggregates.out
+++ b/src/test/regress/expected/aggregates.out
@@ -1210,7 +1210,8 @@ explain (costs off)
select distinct min(f1), max(f1) from minmaxtest;
QUERY PLAN
---------------------------------------------------------------------------------------------
- Unique
+ HashAggregate
+ Group Key: $0, $1
InitPlan 1 (returns $0)
-> Limit
-> Merge Append
@@ -1233,10 +1234,8 @@ explain (costs off)
-> Index Only Scan using minmaxtest2i on minmaxtest2 minmaxtest_8
Index Cond: (f1 IS NOT NULL)
-> Index Only Scan Backward using minmaxtest3i on minmaxtest3 minmaxtest_9
- -> Sort
- Sort Key: ($0), ($1)
- -> Result
-(26 rows)
+ -> Result
+(25 rows)
select distinct min(f1), max(f1) from minmaxtest;
min | max
@@ -2448,6 +2447,241 @@ SELECT balk(hundred) FROM tenk1;
(1 row)
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+SELECT
+ i AS id,
+ i/2 AS p,
+ format('%60s', i%2) AS v,
+ i/4 AS c,
+ i/8 AS d,
+ (random() * (10000/8))::int as e --the same as d but no correlation with p
+ INTO btg
+FROM
+ generate_series(1, 10000) i;
+VACUUM btg;
+ANALYZE btg;
+-- GROUP BY optimization by reorder columns by frequency
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Sort
+ Sort Key: p, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Sort
+ Sort Key: p, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, c, v
+ -> Sort
+ Sort Key: p, c, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: v, p, c
+ -> Sort
+ Sort Key: v, p, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: p, d, c, v
+ -> Sort
+ Sort Key: p, d, c, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: v, p, d, c
+ -> Sort
+ Sort Key: v, p, d, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: p, v, d, c
+ -> Sort
+ Sort Key: p, v, d, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, d, e
+ -> Sort
+ Sort Key: p, d, e
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, e, d
+ -> Sort
+ Sort Key: p, e, d
+ -> Seq Scan on btg
+(5 rows)
+
+CREATE STATISTICS btg_dep ON d, e, p FROM btg;
+ANALYZE btg;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, d, e
+ -> Sort
+ Sort Key: p, d, e
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, e, d
+ -> Sort
+ Sort Key: p, e, d
+ -> Seq Scan on btg
+(5 rows)
+
+-- GROUP BY optimization by reorder columns by index scan
+CREATE INDEX ON btg(p, v);
+SET enable_seqscan=off;
+SET enable_bitmapscan=off;
+VACUUM btg;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v ORDER BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p ORDER BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, c, v
+ -> Incremental Sort
+ Sort Key: p, c, v
+ Presorted Key: p
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY p, v;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, v, c
+ -> Incremental Sort
+ Sort Key: p, v, c
+ Presorted Key: p, v
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, c, d, v
+ -> Incremental Sort
+ Sort Key: p, c, d, v
+ Presorted Key: p
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d ORDER BY p, v;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, v, c, d
+ -> Incremental Sort
+ Sort Key: p, v, c, d
+ Presorted Key: p, v
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+DROP TABLE btg;
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+RESET enable_seqscan;
+RESET enable_bitmapscan;
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/expected/guc.out b/src/test/regress/expected/guc.out
index 3de6404ba5b..98a16c118f2 100644
--- a/src/test/regress/expected/guc.out
+++ b/src/test/regress/expected/guc.out
@@ -858,10 +858,13 @@ SELECT name FROM tab_settings_flags
SELECT name FROM tab_settings_flags
WHERE category ~ '^Query Tuning' AND NOT explain
ORDER BY 1;
- name
----------------------------
+ name
+--------------------------------------
+ debug_enable_cheapest_group_by
+ debug_enable_group_by_match_order_by
+ debug_group_by_reorder_by_pathkeys
default_statistics_target
-(1 row)
+(4 rows)
-- Runtime-computed GUCs should be part of the preset category.
SELECT name FROM tab_settings_flags
diff --git a/src/test/regress/expected/incremental_sort.out b/src/test/regress/expected/incremental_sort.out
index 545e301e482..21c429226f7 100644
--- a/src/test/regress/expected/incremental_sort.out
+++ b/src/test/regress/expected/incremental_sort.out
@@ -1439,7 +1439,7 @@ set parallel_setup_cost = 0;
set parallel_tuple_cost = 0;
set max_parallel_workers_per_gather = 2;
create table t (a int, b int, c int);
-insert into t select mod(i,10),mod(i,10),i from generate_series(1,10000) s(i);
+insert into t select mod(i,10),mod(i,10),i from generate_series(1,60000) s(i);
create index on t (a);
analyze t;
set enable_incremental_sort = off;
diff --git a/src/test/regress/expected/join.out b/src/test/regress/expected/join.out
index 19caebabd01..bf1a2db2cf0 100644
--- a/src/test/regress/expected/join.out
+++ b/src/test/regress/expected/join.out
@@ -1984,8 +1984,8 @@ USING (name);
------+----+----
bb | 12 | 13
cc | 22 | 23
- dd | | 33
ee | 42 |
+ dd | | 33
(4 rows)
-- Cases with non-nullable expressions in subquery results;
@@ -2019,8 +2019,8 @@ NATURAL FULL JOIN
------+------+------+------+------
bb | 12 | 2 | 13 | 3
cc | 22 | 2 | 23 | 3
- dd | | | 33 | 3
ee | 42 | 2 | |
+ dd | | | 33 | 3
(4 rows)
SELECT * FROM
@@ -4618,18 +4618,20 @@ select d.* from d left join (select * from b group by b.id, b.c_id) s
explain (costs off)
select d.* from d left join (select distinct * from b) s
on d.a = s.id;
- QUERY PLAN
---------------------------------------
- Merge Right Join
- Merge Cond: (b.id = d.a)
- -> Unique
- -> Sort
- Sort Key: b.id, b.c_id
- -> Seq Scan on b
+ QUERY PLAN
+---------------------------------------------
+ Merge Left Join
+ Merge Cond: (d.a = s.id)
-> Sort
Sort Key: d.a
-> Seq Scan on d
-(9 rows)
+ -> Sort
+ Sort Key: s.id
+ -> Subquery Scan on s
+ -> HashAggregate
+ Group Key: b.id, b.c_id
+ -> Seq Scan on b
+(11 rows)
-- check join removal works when uniqueness of the join condition is enforced
-- by a UNION
@@ -6336,44 +6338,39 @@ select * from j1 natural join j2;
explain (verbose, costs off)
select * from j1
inner join (select distinct id from j3) j3 on j1.id = j3.id;
- QUERY PLAN
------------------------------------------
+ QUERY PLAN
+-----------------------------------
Nested Loop
Output: j1.id, j3.id
Inner Unique: true
Join Filter: (j1.id = j3.id)
- -> Unique
+ -> HashAggregate
Output: j3.id
- -> Sort
+ Group Key: j3.id
+ -> Seq Scan on public.j3
Output: j3.id
- Sort Key: j3.id
- -> Seq Scan on public.j3
- Output: j3.id
-> Seq Scan on public.j1
Output: j1.id
-(13 rows)
+(11 rows)
-- ensure group by clause allows the inner to become unique
explain (verbose, costs off)
select * from j1
inner join (select id from j3 group by id) j3 on j1.id = j3.id;
- QUERY PLAN
------------------------------------------
+ QUERY PLAN
+-----------------------------------
Nested Loop
Output: j1.id, j3.id
Inner Unique: true
Join Filter: (j1.id = j3.id)
- -> Group
+ -> HashAggregate
Output: j3.id
Group Key: j3.id
- -> Sort
+ -> Seq Scan on public.j3
Output: j3.id
- Sort Key: j3.id
- -> Seq Scan on public.j3
- Output: j3.id
-> Seq Scan on public.j1
Output: j1.id
-(14 rows)
+(11 rows)
drop table j1;
drop table j2;
diff --git a/src/test/regress/expected/partition_aggregate.out b/src/test/regress/expected/partition_aggregate.out
index dfa4b036b52..a08a3825ff6 100644
--- a/src/test/regress/expected/partition_aggregate.out
+++ b/src/test/regress/expected/partition_aggregate.out
@@ -952,32 +952,30 @@ SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HA
--------------------------------------------------------------------------------------
Sort
Sort Key: pagg_tab_ml.a, (sum(pagg_tab_ml.b)), (array_agg(DISTINCT pagg_tab_ml.c))
- -> Gather
- Workers Planned: 2
- -> Parallel Append
- -> GroupAggregate
- Group Key: pagg_tab_ml.a
- Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml.a
- -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
- -> GroupAggregate
- Group Key: pagg_tab_ml_5.a
- Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_5.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
- -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
- -> GroupAggregate
- Group Key: pagg_tab_ml_2.a
- Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_2.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
- -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
-(27 rows)
+ -> Append
+ -> GroupAggregate
+ Group Key: pagg_tab_ml.a
+ Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml.a
+ -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_2.a
+ Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_2.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
+ -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_5.a
+ Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_5.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
+ -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
+(25 rows)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3 ORDER BY 1, 2, 3;
a | sum | array_agg | count
@@ -996,34 +994,32 @@ SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HA
-- Without ORDER BY clause, to test Gather at top-most path
EXPLAIN (COSTS OFF)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3;
- QUERY PLAN
----------------------------------------------------------------------------
- Gather
- Workers Planned: 2
- -> Parallel Append
- -> GroupAggregate
- Group Key: pagg_tab_ml.a
- Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml.a
- -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
- -> GroupAggregate
- Group Key: pagg_tab_ml_5.a
- Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_5.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
- -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
- -> GroupAggregate
- Group Key: pagg_tab_ml_2.a
- Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_2.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
- -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
-(25 rows)
+ QUERY PLAN
+---------------------------------------------------------------------
+ Append
+ -> GroupAggregate
+ Group Key: pagg_tab_ml.a
+ Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml.a
+ -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_2.a
+ Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_2.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
+ -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_5.a
+ Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_5.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
+ -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
+(23 rows)
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
@@ -1379,28 +1375,26 @@ SELECT x, sum(y), avg(y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) <
-- When GROUP BY clause does not match; partial aggregation is performed for each partition.
EXPLAIN (COSTS OFF)
SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) < 12 ORDER BY 1, 2, 3;
- QUERY PLAN
--------------------------------------------------------------------------------------------
+ QUERY PLAN
+-------------------------------------------------------------------------------------
Sort
Sort Key: pagg_tab_para.y, (sum(pagg_tab_para.x)), (avg(pagg_tab_para.x))
- -> Finalize GroupAggregate
+ -> Finalize HashAggregate
Group Key: pagg_tab_para.y
Filter: (avg(pagg_tab_para.x) < '12'::numeric)
- -> Gather Merge
+ -> Gather
Workers Planned: 2
- -> Sort
- Sort Key: pagg_tab_para.y
- -> Parallel Append
- -> Partial HashAggregate
- Group Key: pagg_tab_para.y
- -> Parallel Seq Scan on pagg_tab_para_p1 pagg_tab_para
- -> Partial HashAggregate
- Group Key: pagg_tab_para_1.y
- -> Parallel Seq Scan on pagg_tab_para_p2 pagg_tab_para_1
- -> Partial HashAggregate
- Group Key: pagg_tab_para_2.y
- -> Parallel Seq Scan on pagg_tab_para_p3 pagg_tab_para_2
-(19 rows)
+ -> Parallel Append
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para.y
+ -> Parallel Seq Scan on pagg_tab_para_p1 pagg_tab_para
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para_1.y
+ -> Parallel Seq Scan on pagg_tab_para_p2 pagg_tab_para_1
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para_2.y
+ -> Parallel Seq Scan on pagg_tab_para_p3 pagg_tab_para_2
+(17 rows)
SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) < 12 ORDER BY 1, 2, 3;
y | sum | avg | count
diff --git a/src/test/regress/expected/partition_join.out b/src/test/regress/expected/partition_join.out
index bb5b7c47a45..03926a84138 100644
--- a/src/test/regress/expected/partition_join.out
+++ b/src/test/regress/expected/partition_join.out
@@ -466,52 +466,41 @@ EXPLAIN (COSTS OFF)
SELECT a, b FROM prt1 FULL JOIN prt2 p2(b,a,c) USING(a,b)
WHERE a BETWEEN 490 AND 510
GROUP BY 1, 2 ORDER BY 1, 2;
- QUERY PLAN
------------------------------------------------------------------------------------------------------------------
+ QUERY PLAN
+-----------------------------------------------------------------------------------------------------------
Group
Group Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Merge Append
+ -> Sort
Sort Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Group
- Group Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Sort
- Sort Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Merge Full Join
- Merge Cond: ((prt1.a = p2.a) AND (prt1.b = p2.b))
- Filter: ((COALESCE(prt1.a, p2.a) >= 490) AND (COALESCE(prt1.a, p2.a) <= 510))
- -> Sort
- Sort Key: prt1.a, prt1.b
- -> Seq Scan on prt1_p1 prt1
- -> Sort
- Sort Key: p2.a, p2.b
- -> Seq Scan on prt2_p1 p2
- -> Group
- Group Key: (COALESCE(prt1_1.a, p2_1.a)), (COALESCE(prt1_1.b, p2_1.b))
- -> Sort
- Sort Key: (COALESCE(prt1_1.a, p2_1.a)), (COALESCE(prt1_1.b, p2_1.b))
- -> Merge Full Join
- Merge Cond: ((prt1_1.a = p2_1.a) AND (prt1_1.b = p2_1.b))
- Filter: ((COALESCE(prt1_1.a, p2_1.a) >= 490) AND (COALESCE(prt1_1.a, p2_1.a) <= 510))
- -> Sort
- Sort Key: prt1_1.a, prt1_1.b
- -> Seq Scan on prt1_p2 prt1_1
- -> Sort
- Sort Key: p2_1.a, p2_1.b
- -> Seq Scan on prt2_p2 p2_1
- -> Group
- Group Key: (COALESCE(prt1_2.a, p2_2.a)), (COALESCE(prt1_2.b, p2_2.b))
- -> Sort
- Sort Key: (COALESCE(prt1_2.a, p2_2.a)), (COALESCE(prt1_2.b, p2_2.b))
- -> Merge Full Join
- Merge Cond: ((prt1_2.a = p2_2.a) AND (prt1_2.b = p2_2.b))
- Filter: ((COALESCE(prt1_2.a, p2_2.a) >= 490) AND (COALESCE(prt1_2.a, p2_2.a) <= 510))
- -> Sort
- Sort Key: prt1_2.a, prt1_2.b
- -> Seq Scan on prt1_p3 prt1_2
- -> Sort
- Sort Key: p2_2.a, p2_2.b
- -> Seq Scan on prt2_p3 p2_2
-(43 rows)
+ -> Append
+ -> Merge Full Join
+ Merge Cond: ((prt1_1.a = p2_1.a) AND (prt1_1.b = p2_1.b))
+ Filter: ((COALESCE(prt1_1.a, p2_1.a) >= 490) AND (COALESCE(prt1_1.a, p2_1.a) <= 510))
+ -> Sort
+ Sort Key: prt1_1.a, prt1_1.b
+ -> Seq Scan on prt1_p1 prt1_1
+ -> Sort
+ Sort Key: p2_1.a, p2_1.b
+ -> Seq Scan on prt2_p1 p2_1
+ -> Merge Full Join
+ Merge Cond: ((prt1_2.a = p2_2.a) AND (prt1_2.b = p2_2.b))
+ Filter: ((COALESCE(prt1_2.a, p2_2.a) >= 490) AND (COALESCE(prt1_2.a, p2_2.a) <= 510))
+ -> Sort
+ Sort Key: prt1_2.a, prt1_2.b
+ -> Seq Scan on prt1_p2 prt1_2
+ -> Sort
+ Sort Key: p2_2.a, p2_2.b
+ -> Seq Scan on prt2_p2 p2_2
+ -> Merge Full Join
+ Merge Cond: ((prt1_3.b = p2_3.b) AND (prt1_3.a = p2_3.a))
+ Filter: ((COALESCE(prt1_3.a, p2_3.a) >= 490) AND (COALESCE(prt1_3.a, p2_3.a) <= 510))
+ -> Sort
+ Sort Key: prt1_3.b, prt1_3.a
+ -> Seq Scan on prt1_p3 prt1_3
+ -> Sort
+ Sort Key: p2_3.b, p2_3.a
+ -> Seq Scan on prt2_p3 p2_3
+(32 rows)
SELECT a, b FROM prt1 FULL JOIN prt2 p2(b,a,c) USING(a,b)
WHERE a BETWEEN 490 AND 510
diff --git a/src/test/regress/expected/union.out b/src/test/regress/expected/union.out
index dece7310cfe..7ac4a9380e2 100644
--- a/src/test/regress/expected/union.out
+++ b/src/test/regress/expected/union.out
@@ -1303,24 +1303,22 @@ select distinct q1 from
union all
select distinct * from int8_tbl i82) ss
where q2 = q2;
- QUERY PLAN
-----------------------------------------------------------
- Unique
- -> Merge Append
- Sort Key: "*SELECT* 1".q1
+ QUERY PLAN
+----------------------------------------------------
+ HashAggregate
+ Group Key: "*SELECT* 1".q1
+ -> Append
-> Subquery Scan on "*SELECT* 1"
- -> Unique
- -> Sort
- Sort Key: i81.q1, i81.q2
- -> Seq Scan on int8_tbl i81
- Filter: (q2 IS NOT NULL)
+ -> HashAggregate
+ Group Key: i81.q1, i81.q2
+ -> Seq Scan on int8_tbl i81
+ Filter: (q2 IS NOT NULL)
-> Subquery Scan on "*SELECT* 2"
- -> Unique
- -> Sort
- Sort Key: i82.q1, i82.q2
- -> Seq Scan on int8_tbl i82
- Filter: (q2 IS NOT NULL)
-(15 rows)
+ -> HashAggregate
+ Group Key: i82.q1, i82.q2
+ -> Seq Scan on int8_tbl i82
+ Filter: (q2 IS NOT NULL)
+(13 rows)
select distinct q1 from
(select distinct * from int8_tbl i81
@@ -1339,24 +1337,22 @@ select distinct q1 from
union all
select distinct * from int8_tbl i82) ss
where -q1 = q2;
- QUERY PLAN
---------------------------------------------------------
- Unique
- -> Merge Append
- Sort Key: "*SELECT* 1".q1
+ QUERY PLAN
+--------------------------------------------------
+ HashAggregate
+ Group Key: "*SELECT* 1".q1
+ -> Append
-> Subquery Scan on "*SELECT* 1"
- -> Unique
- -> Sort
- Sort Key: i81.q1, i81.q2
- -> Seq Scan on int8_tbl i81
- Filter: ((- q1) = q2)
+ -> HashAggregate
+ Group Key: i81.q1, i81.q2
+ -> Seq Scan on int8_tbl i81
+ Filter: ((- q1) = q2)
-> Subquery Scan on "*SELECT* 2"
- -> Unique
- -> Sort
- Sort Key: i82.q1, i82.q2
- -> Seq Scan on int8_tbl i82
- Filter: ((- q1) = q2)
-(15 rows)
+ -> HashAggregate
+ Group Key: i82.q1, i82.q2
+ -> Seq Scan on int8_tbl i82
+ Filter: ((- q1) = q2)
+(13 rows)
select distinct q1 from
(select distinct * from int8_tbl i81
diff --git a/src/test/regress/sql/aggregates.sql b/src/test/regress/sql/aggregates.sql
index 2f5d0e00f3d..c6e0d7ba2b7 100644
--- a/src/test/regress/sql/aggregates.sql
+++ b/src/test/regress/sql/aggregates.sql
@@ -1025,6 +1025,105 @@ SELECT balk(hundred) FROM tenk1;
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+
+SELECT
+ i AS id,
+ i/2 AS p,
+ format('%60s', i%2) AS v,
+ i/4 AS c,
+ i/8 AS d,
+ (random() * (10000/8))::int as e --the same as d but no correlation with p
+ INTO btg
+FROM
+ generate_series(1, 10000) i;
+
+VACUUM btg;
+ANALYZE btg;
+
+-- GROUP BY optimization by reorder columns by frequency
+
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+
+CREATE STATISTICS btg_dep ON d, e, p FROM btg;
+ANALYZE btg;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+
+
+-- GROUP BY optimization by reorder columns by index scan
+
+CREATE INDEX ON btg(p, v);
+SET enable_seqscan=off;
+SET enable_bitmapscan=off;
+VACUUM btg;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d ORDER BY p, v;
+
+DROP TABLE btg;
+
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+RESET enable_seqscan;
+RESET enable_bitmapscan;
+
+
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/sql/incremental_sort.sql b/src/test/regress/sql/incremental_sort.sql
index d8768a6b54d..1de163e0395 100644
--- a/src/test/regress/sql/incremental_sort.sql
+++ b/src/test/regress/sql/incremental_sort.sql
@@ -213,7 +213,7 @@ set parallel_tuple_cost = 0;
set max_parallel_workers_per_gather = 2;
create table t (a int, b int, c int);
-insert into t select mod(i,10),mod(i,10),i from generate_series(1,10000) s(i);
+insert into t select mod(i,10),mod(i,10),i from generate_series(1,60000) s(i);
create index on t (a);
analyze t;
--
2.34.1
[text/x-patch] 0002-fixup-rework-GUC-options-20220328.patch (7.6K, ../../[email protected]/3-0002-fixup-rework-GUC-options-20220328.patch)
download | inline diff:
From 03e21ff5c7f6a0c4e2d43a0be647f0cd75a04786 Mon Sep 17 00:00:00 2001
From: Tomas Vondra <[email protected]>
Date: Mon, 28 Mar 2022 17:14:54 +0200
Subject: [PATCH 2/7] fixup: rework GUC options
Replace the three GUC debug options with a single GUC that just disables
the optimization and reverts back to the order specified in the query.
---
src/backend/optimizer/path/pathkeys.c | 24 ++++++---------
src/backend/utils/misc/guc.c | 42 ++++++--------------------
src/include/optimizer/paths.h | 6 +---
src/test/regress/expected/guc.out | 9 ++----
src/test/regress/expected/sysviews.out | 3 +-
5 files changed, 26 insertions(+), 58 deletions(-)
diff --git a/src/backend/optimizer/path/pathkeys.c b/src/backend/optimizer/path/pathkeys.c
index f18d4c8e4e7..fcc99f0fb1c 100644
--- a/src/backend/optimizer/path/pathkeys.c
+++ b/src/backend/optimizer/path/pathkeys.c
@@ -31,6 +31,8 @@
#include "utils/lsyscache.h"
#include "utils/selfuncs.h"
+/* Consider reordering of GROUP BY keys? */
+bool enable_group_by_reordering = true;
static bool pathkey_is_redundant(PathKey *new_pathkey, List *pathkeys);
static bool matches_boolean_partition_clause(RestrictInfo *rinfo,
@@ -337,12 +339,6 @@ pathkeys_contained_in(List *keys1, List *keys2)
return false;
}
-/************************<DEBUG PART>*************************************/
-bool debug_group_by_reorder_by_pathkeys = true;
-bool debug_group_by_match_order_by = true;
-bool debug_cheapest_group_by = true;
-/************************</DEBUG PART>************************************/
-
/*
* group_keys_reorder_by_pathkeys
* Reorder GROUP BY keys to match pathkeys of input path.
@@ -361,9 +357,6 @@ group_keys_reorder_by_pathkeys(List *pathkeys, List **group_pathkeys,
ListCell *lc;
int n;
- if (debug_group_by_reorder_by_pathkeys == false)
- return 0;
-
if (pathkeys == NIL || *group_pathkeys == NIL)
return 0;
@@ -598,10 +591,6 @@ get_cheapest_group_keys_order(PlannerInfo *root, double nrows,
int nFreeKeys;
int nToPermute;
- /* If this optimization is disabled, we're done. */
- if (!debug_cheapest_group_by)
- return false;
-
/* If there are less than 2 unsorted pathkeys, we're done. */
if (list_length(*group_pathkeys) - n_preordered < 2)
return false;
@@ -771,6 +760,13 @@ get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
infos = lappend(infos, info);
+ /*
+ * If the optimization is disabled, we consider only the pathkey order as
+ * specified in the query. We don't do any reorderings.
+ */
+ if (!enable_group_by_reordering)
+ return infos;
+
/* for grouping sets we can't do any reordering */
if (parse->groupingSets)
return infos;
@@ -834,7 +830,7 @@ get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
* XXX This does nothing if (n_preordered == 0). We shouldn't create the
* info in this case.
*/
- if (root->sort_pathkeys && debug_group_by_match_order_by)
+ if (root->sort_pathkeys)
{
n_preordered = group_keys_reorder_by_pathkeys(root->sort_pathkeys,
&pathkeys,
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index 31f347d469d..9e8ab1420d9 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -1182,6 +1182,16 @@ static struct config_bool ConfigureNamesBool[] =
true,
NULL, NULL, NULL
},
+ {
+ {"enable_group_by_reordering", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("enable reordering of GROUP BY key"),
+ NULL,
+ GUC_EXPLAIN
+ },
+ &enable_group_by_reordering,
+ true,
+ NULL, NULL, NULL
+ },
{
{"geqo", PGC_USERSET, QUERY_TUNING_GEQO,
gettext_noop("Enables genetic query optimization."),
@@ -2140,38 +2150,6 @@ static struct config_bool ConfigureNamesBool[] =
NULL, NULL, NULL
},
-/************************<DEBUG OPT GROUP BY>*********************************/
- {
- {"debug_group_by_reorder_by_pathkeys", PGC_USERSET, QUERY_TUNING_METHOD,
- gettext_noop("enable reorder GROUP BY by pathkeys"),
- NULL,
- GUC_NOT_IN_SAMPLE
- },
- &debug_group_by_reorder_by_pathkeys,
- true,
- NULL, NULL, NULL
- },
- {
- {"debug_enable_group_by_match_order_by", PGC_USERSET, QUERY_TUNING_METHOD,
- gettext_noop("enable matching GROUP BY by ORDER BY."),
- NULL,
- GUC_NOT_IN_SAMPLE
- },
- &debug_group_by_match_order_by,
- true,
- NULL, NULL, NULL
- },
- {
- {"debug_enable_cheapest_group_by", PGC_USERSET, QUERY_TUNING_METHOD,
- gettext_noop("find a cheapest order of columns in GROUP BY."),
- NULL,
- GUC_NOT_IN_SAMPLE
- },
- &debug_cheapest_group_by,
- true,
- NULL, NULL, NULL
- },
-/************************</DEBUG OPT GROUP BY>********************************/
/* End-of-list marker */
{
{NULL, 0, 0, NULL, NULL}, NULL, false, NULL, NULL, NULL
diff --git a/src/include/optimizer/paths.h b/src/include/optimizer/paths.h
index 525b594cbd0..2ffa24aa899 100644
--- a/src/include/optimizer/paths.h
+++ b/src/include/optimizer/paths.h
@@ -24,6 +24,7 @@ extern PGDLLIMPORT bool enable_geqo;
extern PGDLLIMPORT int geqo_threshold;
extern PGDLLIMPORT int min_parallel_table_scan_size;
extern PGDLLIMPORT int min_parallel_index_scan_size;
+extern PGDLLIMPORT bool enable_group_by_reordering;
/* Hook for plugins to get control in set_rel_pathlist() */
typedef void (*set_rel_pathlist_hook_type) (PlannerInfo *root,
@@ -206,11 +207,6 @@ extern bool pathkeys_count_contained_in(List *keys1, List *keys2, int *n_common)
extern int group_keys_reorder_by_pathkeys(List *pathkeys,
List **group_pathkeys,
List **group_clauses);
-/************************<DEBUG OPT GROUP BY>*********************************/
-extern bool debug_group_by_reorder_by_pathkeys;
-extern bool debug_group_by_match_order_by;
-extern bool debug_cheapest_group_by;
-/************************</DEBUG OPT GROUP BY>********************************/
extern List *get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
List *path_pathkeys,
List *pathkeys, List *clauses);
diff --git a/src/test/regress/expected/guc.out b/src/test/regress/expected/guc.out
index 98a16c118f2..3de6404ba5b 100644
--- a/src/test/regress/expected/guc.out
+++ b/src/test/regress/expected/guc.out
@@ -858,13 +858,10 @@ SELECT name FROM tab_settings_flags
SELECT name FROM tab_settings_flags
WHERE category ~ '^Query Tuning' AND NOT explain
ORDER BY 1;
- name
---------------------------------------
- debug_enable_cheapest_group_by
- debug_enable_group_by_match_order_by
- debug_group_by_reorder_by_pathkeys
+ name
+---------------------------
default_statistics_target
-(4 rows)
+(1 row)
-- Runtime-computed GUCs should be part of the preset category.
SELECT name FROM tab_settings_flags
diff --git a/src/test/regress/expected/sysviews.out b/src/test/regress/expected/sysviews.out
index 442eeb1e3fe..72891690399 100644
--- a/src/test/regress/expected/sysviews.out
+++ b/src/test/regress/expected/sysviews.out
@@ -105,6 +105,7 @@ select name, setting from pg_settings where name like 'enable%';
enable_async_append | on
enable_bitmapscan | on
enable_gathermerge | on
+ enable_group_by_reordering | on
enable_hashagg | on
enable_hashjoin | on
enable_incremental_sort | on
@@ -122,7 +123,7 @@ select name, setting from pg_settings where name like 'enable%';
enable_seqscan | on
enable_sort | on
enable_tidscan | on
-(20 rows)
+(21 rows)
-- Test that the pg_timezone_names and pg_timezone_abbrevs views are
-- more-or-less working. We can't test their contents in any great detail
--
2.34.1
[text/x-patch] 0003-fixup-minor-code-cleanup-20220328.patch (1.8K, ../../[email protected]/4-0003-fixup-minor-code-cleanup-20220328.patch)
download | inline diff:
From 3beb6e4669f47f78f6356a2ba52ebe013262bda6 Mon Sep 17 00:00:00 2001
From: Tomas Vondra <[email protected]>
Date: Mon, 28 Mar 2022 18:40:22 +0200
Subject: [PATCH 3/7] fixup: minor code cleanup
---
src/backend/optimizer/plan/planner.c | 6 +-----
1 file changed, 1 insertion(+), 5 deletions(-)
diff --git a/src/backend/optimizer/plan/planner.c b/src/backend/optimizer/plan/planner.c
index bb957d50d23..b2569c5d0c7 100644
--- a/src/backend/optimizer/plan/planner.c
+++ b/src/backend/optimizer/plan/planner.c
@@ -6220,7 +6220,6 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
{
ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- Path *path_save = path;
Path *path_original = path;
List *pathkey_orderings = NIL;
@@ -6245,7 +6244,7 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
/* restore the path (we replace it in the loop) */
- path = path_save;
+ path = path_original;
is_sorted = pathkeys_count_contained_in(info->pathkeys,
path->pathkeys,
@@ -6255,13 +6254,11 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
{
/* Sort the cheapest-total path if it isn't already sorted */
if (!is_sorted)
- {
path = (Path *) create_sort_path(root,
grouped_rel,
path,
info->pathkeys,
-1.0);
- }
/* Now decide what to stick atop it */
if (parse->groupingSets)
@@ -6432,7 +6429,6 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
{
if (path != partially_grouped_rel->cheapest_total_path)
continue;
-
path = (Path *) create_sort_path(root,
grouped_rel,
path,
--
2.34.1
[text/x-patch] 0004-fixup-comments-cleanup-type-fixes-and-rewor-20220328.patch (21.9K, ../../[email protected]/5-0004-fixup-comments-cleanup-type-fixes-and-rewor-20220328.patch)
download | inline diff:
From 7a3a7facd15d2c0034eeff7041df56a4e0d456f4 Mon Sep 17 00:00:00 2001
From: Tomas Vondra <[email protected]>
Date: Tue, 29 Mar 2022 01:12:15 +0200
Subject: [PATCH 4/7] fixup: comments cleanup, type fixes and rewording
---
src/backend/optimizer/path/costsize.c | 124 +++++++++++++-----------
src/backend/optimizer/path/equivclass.c | 6 +-
src/backend/optimizer/path/pathkeys.c | 68 ++++++-------
src/backend/utils/adt/selfuncs.c | 13 ++-
4 files changed, 105 insertions(+), 106 deletions(-)
diff --git a/src/backend/optimizer/path/costsize.c b/src/backend/optimizer/path/costsize.c
index f4f7ce01065..bb0fe0c22ee 100644
--- a/src/backend/optimizer/path/costsize.c
+++ b/src/backend/optimizer/path/costsize.c
@@ -1773,9 +1773,9 @@ is_fake_var(Expr *expr)
/*
* get_width_cost_multiplier
- * Returns relative complexity of comparing two values based on it's width.
- * The idea behind - long values have more expensive comparison. Return value is
- * in cpu_operator_cost unit.
+ * Returns relative complexity of comparing two values based on its width.
+ * The idea behind is that the comparison becomes more expensive the longer the
+ * value is. Return value is in cpu_operator_cost units.
*/
static double
get_width_cost_multiplier(PlannerInfo *root, Expr *expr)
@@ -1806,7 +1806,7 @@ get_width_cost_multiplier(PlannerInfo *root, Expr *expr)
}
}
- /* Didn't find any actual stats, use estimation by type */
+ /* Didn't find any actual stats, try using type width instead. */
if (width < 0.0)
{
Node *node = (Node*) expr;
@@ -1815,17 +1815,20 @@ get_width_cost_multiplier(PlannerInfo *root, Expr *expr)
}
/*
- * Any value in pgsql is passed by Datum type, so any operation with value
- * could not be cheaper than operation with Datum type
+ * Values are passed as Datum type, so comparisons can't be cheaper than
+ * comparing a Datum value.
+ *
+ * FIXME I find this reasoning questionable. We may pass int2, and comparing
+ * it is probably a bit cheaper than comparing a bigint.
*/
if (width <= sizeof(Datum))
return 1.0;
/*
- * Seems, cost of comparision is not directly proportional to args width,
- * because comparing args could be differ width (we known only average over
- * column) and difference often could be defined only by looking on first
- * bytes. So, use log16(width) as estimation.
+ * We consider the cost of a comparison not to be directly proportional to
+ * width of the argument, because widths of the arguments could be slightly
+ * different (we only know the average width for the whole column). So we
+ * use log16(width) as an estimate.
*/
return 1.0 + 0.125 * LOG2(width / sizeof(Datum));
}
@@ -1837,11 +1840,11 @@ get_width_cost_multiplier(PlannerInfo *root, Expr *expr)
* The main thing we need to calculate to estimate sort CPU costs is the number
* of calls to the comparator functions. The difficulty is that for multi-column
* sorts there may be different data types involved (for some of which the calls
- * may be much more expensive). Furthermore, the columns may have very different
+ * may be much more expensive). Furthermore, columns may have a very different
* number of distinct values - the higher the number, the fewer comparisons will
* be needed for the following columns.
*
- * The algoritm is incremental - we add pathkeys one by one, and at each step we
+ * The algorithm is incremental - we add pathkeys one by one, and at each step we
* estimate the number of necessary comparisons (based on the number of distinct
* groups in the current pathkey prefix and the new pathkey), and the comparison
* costs (which is data type specific).
@@ -1892,7 +1895,6 @@ compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
bool has_fake_var = false;
int i = 0;
Oid prev_datatype = InvalidOid;
- Cost funcCost = 0.0;
List *cache_varinfos = NIL;
/* fallback if pathkeys is unknown */
@@ -1901,12 +1903,8 @@ compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
/*
* If we'll use a bounded heap-sort keeping just K tuples in memory, for
* a total number of tuple comparisons of N log2 K; but the constant
- * factor is a bit higher than for quicksort. Tweak it so that the
- * cost curve is continuous at the crossover point.
- *
- * XXX I suppose the "quicksort factor" references to 1.5 at the end
- * of this function, but I'm not sure. I suggest we introduce some simple
- * constants for that, instead of magic values.
+ * factor is a bit higher than for quicksort. Tweak it so that the cost
+ * curve is continuous at the crossover point.
*/
output_tuples = (heapSort) ? 2.0 * output_tuples : tuples;
per_tuple_cost += 2.0 * cpu_operator_cost * LOG2(output_tuples);
@@ -1924,20 +1922,21 @@ compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
*/
foreach(lc, pathkeys)
{
- PathKey *pathkey = (PathKey*) lfirst(lc);
- EquivalenceMember *em;
- double nGroups,
- correctedNGroups;
+ PathKey *pathkey = (PathKey*) lfirst(lc);
+ EquivalenceMember *em;
+ double nGroups,
+ correctedNGroups;
+ Cost funcCost = 1.0;
/*
* We believe that equivalence members aren't very different, so, to
- * estimate cost we take just first member
+ * estimate cost we consider just the first member.
*/
em = (EquivalenceMember *) linitial(pathkey->pk_eclass->ec_members);
if (em->em_datatype != InvalidOid)
{
- /* do not lookup funcCost if data type is the same as previous */
+ /* do not lookup funcCost if the data type is the same */
if (prev_datatype != em->em_datatype)
{
Oid sortop;
@@ -1950,54 +1949,56 @@ compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
cost.startup = 0;
cost.per_tuple = 0;
add_function_cost(root, get_opcode(sortop), NULL, &cost);
- /* we need procost, not product of procost and cpu_operator_cost */
+
+ /*
+ * add_function_cost returns the product of cpu_operator_cost
+ * and procost, but we need just procost, co undo that.
+ */
funcCost = cost.per_tuple / cpu_operator_cost;
+
prev_datatype = em->em_datatype;
}
}
- else
- funcCost = 1.0; /* fallback */
- /* Try to take into account actual width fee */
+ /* factor in the width of the values in this column */
funcCost *= get_width_cost_multiplier(root, em->em_expr);
+ /* now we have per-key cost, so add to the running total */
totalFuncCost += funcCost;
- /* Remember if we have a fake var in pathkeys */
+ /* remember if we have found a fake Var in pathkeys */
has_fake_var |= is_fake_var(em->em_expr);
pathkeyExprs = lappend(pathkeyExprs, em->em_expr);
/*
- * Prevent call estimate_num_groups() with fake Var. Note,
- * pathkeyExprs contains only previous columns
+ * We need to calculate the number of comparisons for this column, which
+ * requires knowing the group size. So we estimate the number of groups
+ * by calling estimate_num_groups_incremental(), which estimates the
+ * group size for "new" pathkeys.
+ *
+ * Note: estimate_num_groups_incremntal does not handle fake Vars, so use
+ * a default estimate otherwise.
*/
- if (has_fake_var == false)
- /*
- * Recursively compute number of groups in a group from previous step
- */
+ if (!has_fake_var)
nGroups = estimate_num_groups_incremental(root, pathkeyExprs,
tuplesPerPrevGroup, NULL, NULL,
&cache_varinfos,
list_length(pathkeyExprs) - 1);
else if (tuples > 4.0)
/*
- * Use geometric mean as estimation if there is no any stats.
- * Don't use DEFAULT_NUM_DISTINCT because it used for only one
- * column while here we try to estimate number of groups over
- * set of columns.
- *
- * XXX Perhaps this should use DEFAULT_NUM_DISTINCT at least to
- * limit the calculated values, somehow?
+ * Use geometric mean as estimation if there are no stats.
*
- * XXX What's the logic of the following formula?
+ * We don't use DEFAULT_NUM_DISTINCT here, because that’s used for
+ * a single column, but here we’re dealing with multiple columns.
*/
nGroups = ceil(2.0 + sqrt(tuples) * (i + 1) / list_length(pathkeys));
else
nGroups = tuples;
/*
- * Presorted keys aren't participated in comparison but still checked
- * by qsort comparator.
+ * Presorted keys are not considered in the cost above, but we still do
+ * have to compare them in the qsort comparator. So make sure to factor
+ * in the cost in that case.
*/
if (i >= nPresortedKeys)
{
@@ -2006,10 +2007,13 @@ compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
double heap_tuples;
/* have to keep at least one group, and a multiple of group size */
+ /* FIXME Do we actually need the Max, here? Surely the ceil() returns
+ * at least 1.0 here, no? */
heap_tuples = Max(ceil(output_tuples / tuplesPerPrevGroup) * tuplesPerPrevGroup,
tuplesPerPrevGroup);
/* so how many (whole) groups is that? */
+ /* FIXME so why not just ceil(putput_tuples / tuplesPerPrevGroup)? */
correctedNGroups = ceil(heap_tuples / tuplesPerPrevGroup);
}
else
@@ -2024,17 +2028,19 @@ compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
i++;
/*
- * Real-world distribution isn't uniform but now we don't have a way to
- * determine that, so, add multiplier to get closer to worst case.
- * But ensure the number of tuples does not exceed the group size in the
- * preceding step.
+ * Uniform distributions with all groups being of the same size are the
+ * best case, with nice smooth behavior. Real-world distributions tend
+ * not to be uniform, though, and we don’t have any reliable easy-to-use
+ * information. As a basic defense against skewed distributions, we use
+ * a 1.5 factor to make the expected group a bit larger, but we need to
+ * be careful not to make the group larger than in the preceding step.
*/
tuplesPerPrevGroup = Min(tuplesPerPrevGroup,
ceil(1.5 * tuplesPerPrevGroup / nGroups));
/*
- * We could skip all following columns for cost estimation, because we
- * believe that tuples are unique by the set of previous columns
+ * Once we get single-row group, it means tuples in the group are unique
+ * and we can skip all remaining columns.
*/
if (tuplesPerPrevGroup <= 1.0)
break;
@@ -2049,12 +2055,12 @@ compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
* Accordingly to "Introduction to algorithms", Thomas H. Cormen, Charles E.
* Leiserson, Ronald L. Rivest, ISBN 0-07-013143-0, quicksort estimation
* formula has additional term proportional to number of tuples (See Chapter
- * 8.2 and Theorem 4.1). It has meaning with low number of tuples,
- * approximately less that 1e4. Of course, it could be implemented as
- * additional multiplier under logarithm, but use more complicated formula
- * which takes into account number of unique tuples and it isn't clear how
- * to combine multiplier with groups. Estimate it as 10 in cpu_operator_cost
- * unit.
+ * 8.2 and Theorem 4.1). That affects cases with a low number of tuples,
+ * approximately less than 1e4. We could implement it as an additional
+ * multiplier under the logarithm, but we use a bit more complex formula
+ * which takes into account the number of unique tuples and it’s not clear
+ * how to combine the multiplier with the number of groups. Estimate it as
+ * 10 in cpu_operator_cost unit.
*/
per_tuple_cost += 10 * cpu_operator_cost;
@@ -2110,7 +2116,7 @@ cost_sort_estimate(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
* 'sort_mem' is the number of kilobytes of work memory allowed for the sort
* 'limit_tuples' is the bound on the number of output tuples; -1 if no bound
* 'startup_cost' is expected to be 0 at input. If there is "input cost" it should
- * be added by caller later.
+ * be added by caller later
*/
static void
cost_tuplesort(PlannerInfo *root, List *pathkeys, Cost *startup_cost, Cost *run_cost,
diff --git a/src/backend/optimizer/path/equivclass.c b/src/backend/optimizer/path/equivclass.c
index 5487ae2ee4c..258302840f8 100644
--- a/src/backend/optimizer/path/equivclass.c
+++ b/src/backend/optimizer/path/equivclass.c
@@ -685,9 +685,9 @@ get_eclass_for_sort_expr(PlannerInfo *root,
/*
* Match!
*
- * Copy sortref if it wasn't set yet, it's possible if ec was
- * constructed from WHERE clause, ie it doesn't have target
- * reference at all
+ * Copy the sortref if it wasn't set yet. That may happen if the
+ * ec was constructed from WHERE clause, i.e. it doesn't have a
+ * target reference at all.
*/
if (cur_ec->ec_sortref == 0 && sortref > 0)
cur_ec->ec_sortref = sortref;
diff --git a/src/backend/optimizer/path/pathkeys.c b/src/backend/optimizer/path/pathkeys.c
index fcc99f0fb1c..7913d62625d 100644
--- a/src/backend/optimizer/path/pathkeys.c
+++ b/src/backend/optimizer/path/pathkeys.c
@@ -362,12 +362,12 @@ group_keys_reorder_by_pathkeys(List *pathkeys, List **group_pathkeys,
/*
* Walk the pathkeys (determining ordering of the input path) and see if
- * there's a matching GROUP BY key. If we find one, we append if to the
+ * there's a matching GROUP BY key. If we find one, we append it to the
* list, and do the same for the clauses.
*
- * Once we find first pathkey without a matching GROUP BY key, the rest of
- * the pathkeys is useless and can't be used to evaluate the grouping, so
- * we abort the loop and ignore the remaining pathkeys.
+ * Once we find the first pathkey without a matching GROUP BY key, the rest
+ * of the pathkeys are useless and can't be used to evaluate the grouping,
+ * so we abort the loop and ignore the remaining pathkeys.
*
* XXX Pathkeys are built in a way to allow simply comparing pointers.
*/
@@ -391,8 +391,6 @@ group_keys_reorder_by_pathkeys(List *pathkeys, List **group_pathkeys,
/* remember the number of pathkeys with a matching GROUP BY key */
n = list_length(new_group_pathkeys);
- /* XXX maybe return when (n == 0) */
-
/* append the remaining group pathkeys (will be treated as not sorted) */
*group_pathkeys = list_concat_unique_ptr(new_group_pathkeys,
*group_pathkeys);
@@ -530,7 +528,7 @@ PathkeyMutatorNext(PathkeyMutatorState *state)
}
/* update the list cells to point to the right elements */
- for(i=0; i<state->mutatorNColumns; i++)
+ for(i = 0; i < state->mutatorNColumns; i++)
lfirst(state->elemCells[i]) =
(void *) state->elems[ state->positions[i] - 1 ];
@@ -597,18 +595,18 @@ get_cheapest_group_keys_order(PlannerInfo *root, double nrows,
/*
* We could exhaustively cost all possible orderings of the pathkeys, but for
- * large number of pathkeys that might be prohibitively expensive. So we try
- * to apply a simple cheap heuristics first - we sort the pathkeys by sort
- * cost (as if the pathkey was sorted independently) and then check only the
- * four cheapest pathkeys. The remaining pathkeys are kept ordered by cost.
+ * a large number of pathkeys it might be prohibitively expensive. So we try
+ * to apply simple cheap heuristics first - we sort the pathkeys by sort cost
+ * (as if the pathkey was sorted independently) and then check only the four
+ * cheapest pathkeys. The remaining pathkeys are kept ordered by cost.
*
- * XXX This is a very simple heuristics, and likely to work fine for most
- * cases (because number of GROUP BY clauses tends to be lower than 4). But
- * it ignores how the number of distinct values in each pathkey affects the
- * following sorts. It may be better to use "more expensive" pathkey first
- * if it has many distinct values, because it then limits the number of
- * comparisons for the remaining pathkeys. But evaluating that is kinda the
- * expensive bit we're trying to not do.
+ * XXX This is a very simple heuristics, but likely to work fine for most
+ * cases (because the number of GROUP BY clauses tends to be lower than 4).
+ * But it ignores how the number of distinct values in each pathkey affects
+ * the following steps. It might be better to use "more expensive" pathkey
+ * first if it has many distinct values, because it then limits the number
+ * of comparisons for the remaining pathkeys. But evaluating that is likely
+ * quite the expensive.
*/
nFreeKeys = list_length(*group_pathkeys) - n_preordered;
nToPermute = 4;
@@ -649,11 +647,7 @@ get_cheapest_group_keys_order(PlannerInfo *root, double nrows,
}
else
{
- /*
- * Since v13 list_free() can clean list elements so for original list
- * not to be modified it should be copied to a new one which can then
- * be cleaned safely if needed.
- */
+ /* Copy the list, so that we can free the new list by list_free. */
new_group_pathkeys = list_copy(*group_pathkeys);
nToPermute = nFreeKeys;
}
@@ -667,18 +661,16 @@ get_cheapest_group_keys_order(PlannerInfo *root, double nrows,
* there's room for two dynamic programming optimizations here. Firstly, we
* may pass the current "best" cost to cost_sort_estimate so that it can
* "abort" if the estimated pathkeys list exceeds it. Secondly, it could pass
- * return information about the position when it exceeded the cost, and we
- * could skip all permutations with the same prefix.
+ * the return information about the position when it exceeded the cost, and
+ * we could skip all permutations with the same prefix.
*
* Imagine we've already found ordering with cost C1, and we're evaluating
* another ordering - cost_sort_estimate() calculates cost by adding the
* pathkeys one by one (more or less), and the cost only grows. If at any
* point it exceeds C1, it can't possibly be "better" so we can discard it.
* But we also know that we can discard all ordering with the same prefix,
- * because if we're estimating (a,b,c,d) and we exceeded C1 at (a,b) then
- * the same thing will happen for any ordering with this prefix.
- *
- *
+ * because if we're estimating (a,b,c,d) and we exceed C1 at (a,b) then the
+ * same thing will happen for any ordering with this prefix.
*/
PathkeyMutatorInit(&mstate, new_group_pathkeys, n_preordered, n_preordered + nToPermute);
@@ -721,7 +713,7 @@ get_cheapest_group_keys_order(PlannerInfo *root, double nrows,
* Determine which orderings of GROUP BY keys are potentially interesting.
*
* Returns list of PathKeyInfo items, each representing an interesting ordering
- * of GROUP BY keys. Each items stores pathkeys and clauses in matching order.
+ * of GROUP BY keys. Each item stores pathkeys and clauses in matching order.
*
* The function considers (and keeps) multiple group by orderings:
*
@@ -730,15 +722,15 @@ get_cheapest_group_keys_order(PlannerInfo *root, double nrows,
* - GROUP BY keys reordered to minimize the sort cost
*
* - GROUP BY keys reordered to match path ordering (as much as possible), with
- * the tail reoredered to minimize the sort cost
+ * the tail reordered to minimize the sort cost
*
* - GROUP BY keys to match target ORDER BY clause (as much as possible), with
- * the tail reoredered to minimize the sort cost
+ * the tail reordered to minimize the sort cost
*
* There are other potentially interesting orderings (e.g. it might be best to
* match the first ORDER BY key, order the remaining keys differently and then
- * rely on incremental sort to fix this), but we ignore those for now. To make
- * this work we'd have to pretty much generate all possible permutations.
+ * rely on the incremental sort to fix this), but we ignore those for now. To
+ * make this work we'd have to pretty much generate all possible permutations.
*/
List *
get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
@@ -761,8 +753,9 @@ get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
infos = lappend(infos, info);
/*
- * If the optimization is disabled, we consider only the pathkey order as
- * specified in the query. We don't do any reorderings.
+ * Should we try generating alternative orderings of the group keys? If not,
+ * we produce only the order specified in the query, i.e. the optimization
+ * is effectively disabled.
*/
if (!enable_group_by_reordering)
return infos;
@@ -826,9 +819,6 @@ get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
/*
* Try reordering pathkeys to minimize the sort cost (this time consider
* the ORDER BY clause, but only if set debug_group_by_match_order_by).
- *
- * XXX This does nothing if (n_preordered == 0). We shouldn't create the
- * info in this case.
*/
if (root->sort_pathkeys)
{
diff --git a/src/backend/utils/adt/selfuncs.c b/src/backend/utils/adt/selfuncs.c
index 36c303e45c5..fb4fb987e7f 100644
--- a/src/backend/utils/adt/selfuncs.c
+++ b/src/backend/utils/adt/selfuncs.c
@@ -3294,10 +3294,7 @@ add_unique_group_var(PlannerInfo *root, List *varinfos,
}
/*
- * estimate_num_groups/estimate_num_groups_incremental
- * - Estimate number of groups in a grouped query.
- * _incremental variant is performance optimization for
- * case of adding one-by-one column
+ * estimate_num_groups - Estimate number of groups in a grouped query
*
* Given a query having a GROUP BY clause, estimate how many groups there
* will be --- ie, the number of distinct combinations of the GROUP BY
@@ -3376,6 +3373,11 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
NULL, 0);
}
+/*
+ * estimate_num_groups_incremental
+ * An estimate_num_groups variant, optimized for cases that are adding the
+ * expressions incrementally (e.g. one by one).
+ */
double
estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
double input_rows,
@@ -3386,7 +3388,8 @@ estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
double srf_multiplier = 1.0;
double numdistinct;
ListCell *l;
- int i, j;
+ int i,
+ j;
/* Zero the estinfo output parameter, if non-NULL */
if (estinfo != NULL)
--
2.34.1
[text/x-patch] 0005-fixup-minor-simplification-20220328.patch (1.3K, ../../[email protected]/6-0005-fixup-minor-simplification-20220328.patch)
download | inline diff:
From 23c94d59b410930c1a0187c1c8eb131995ba3ccc Mon Sep 17 00:00:00 2001
From: Tomas Vondra <[email protected]>
Date: Tue, 29 Mar 2022 02:26:13 +0200
Subject: [PATCH 5/7] fixup: minor simplification
---
src/backend/optimizer/path/costsize.c | 11 +----------
1 file changed, 1 insertion(+), 10 deletions(-)
diff --git a/src/backend/optimizer/path/costsize.c b/src/backend/optimizer/path/costsize.c
index bb0fe0c22ee..3e1daf8b893 100644
--- a/src/backend/optimizer/path/costsize.c
+++ b/src/backend/optimizer/path/costsize.c
@@ -2004,17 +2004,8 @@ compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
{
if (heapSort)
{
- double heap_tuples;
-
/* have to keep at least one group, and a multiple of group size */
- /* FIXME Do we actually need the Max, here? Surely the ceil() returns
- * at least 1.0 here, no? */
- heap_tuples = Max(ceil(output_tuples / tuplesPerPrevGroup) * tuplesPerPrevGroup,
- tuplesPerPrevGroup);
-
- /* so how many (whole) groups is that? */
- /* FIXME so why not just ceil(putput_tuples / tuplesPerPrevGroup)? */
- correctedNGroups = ceil(heap_tuples / tuplesPerPrevGroup);
+ correctedNGroups = ceil(output_tuples / tuplesPerPrevGroup);
}
else
/* all groups in the input */
--
2.34.1
[text/x-patch] 0006-fixup-add-GUC-to-docs-20220328.patch (1.3K, ../../[email protected]/7-0006-fixup-add-GUC-to-docs-20220328.patch)
download | inline diff:
From 31a138efbb203caebcfe8636fd7b873c3362208e Mon Sep 17 00:00:00 2001
From: Tomas Vondra <[email protected]>
Date: Tue, 29 Mar 2022 02:45:52 +0200
Subject: [PATCH 6/7] fixup: add GUC to docs
---
doc/src/sgml/config.sgml | 14 ++++++++++++++
1 file changed, 14 insertions(+)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index 05df48131d7..0c0de61b855 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -4961,6 +4961,20 @@ ANY <replaceable class="parameter">num_sync</replaceable> ( <replaceable class="
</listitem>
</varlistentry>
+ <varlistentry id="guc-enable-groupby-reordering" xreflabel="enable_group_by_reordering">
+ <term><varname>enable_group_by_reordering</varname> (<type>boolean</type>)
+ <indexterm>
+ <primary><varname>enable_group_by_reordering</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Enables or disables reodering of keys in <literal>GROUP BY</literal>
+ clause. The default is <literal>on</literal>.
+ </para>
+ </listitem>
+ </varlistentry>
+
<varlistentry id="guc-enable-hashagg" xreflabel="enable_hashagg">
<term><varname>enable_hashagg</varname> (<type>boolean</type>)
<indexterm>
--
2.34.1
[text/x-patch] 0007-demonstrate-disabled-optimization-20220328.patch (15.9K, ../../[email protected]/8-0007-demonstrate-disabled-optimization-20220328.patch)
download | inline diff:
From 4dd50166459d6fb1c131d215cf79237dfdfd4b05 Mon Sep 17 00:00:00 2001
From: Tomas Vondra <[email protected]>
Date: Mon, 28 Mar 2022 23:31:29 +0200
Subject: [PATCH 7/7] demonstrate disabled optimization
tweak tests to show costing, disable optimization to show what changes
---
.../postgres_fdw/expected/postgres_fdw.out | 4 +-
contrib/postgres_fdw/sql/postgres_fdw.sql | 2 +-
src/backend/optimizer/path/pathkeys.c | 2 +-
src/backend/utils/misc/guc.c | 2 +-
src/test/regress/expected/aggregates.out | 168 +++++++++---------
src/test/regress/sql/aggregates.sql | 26 +--
6 files changed, 102 insertions(+), 102 deletions(-)
diff --git a/contrib/postgres_fdw/expected/postgres_fdw.out b/contrib/postgres_fdw/expected/postgres_fdw.out
index 2ffc836824a..bd826eaedd5 100644
--- a/contrib/postgres_fdw/expected/postgres_fdw.out
+++ b/contrib/postgres_fdw/expected/postgres_fdw.out
@@ -2739,11 +2739,11 @@ select c2 * (random() <= 1)::int as c2 from ft2 group by c2 * (random() <= 1)::i
(9 rows)
-- GROUP BY clause in various forms, cardinal, alias and constant expression
-explain (verbose, costs off)
+explain (verbose, costs on)
select count(c2) w, c2 x, 5 y, 7.0 z from ft1 group by 2, y, 9.0::int order by 2;
QUERY PLAN
------------------------------------------------------------------------------------------------------------
- Foreign Scan
+ Foreign Scan (cost=110.50..132.85 rows=10 width=52)
Output: (count(c2)), c2, 5, 7.0, 9
Relations: Aggregate on (public.ft1)
Remote SQL: SELECT count(c2), c2, 5, 7.0, 9 FROM "S 1"."T 1" GROUP BY 2, 3, 5 ORDER BY c2 ASC NULLS LAST
diff --git a/contrib/postgres_fdw/sql/postgres_fdw.sql b/contrib/postgres_fdw/sql/postgres_fdw.sql
index 95b6b7192e6..61b04d6f9a0 100644
--- a/contrib/postgres_fdw/sql/postgres_fdw.sql
+++ b/contrib/postgres_fdw/sql/postgres_fdw.sql
@@ -732,7 +732,7 @@ explain (verbose, costs off)
select c2 * (random() <= 1)::int as c2 from ft2 group by c2 * (random() <= 1)::int order by 1;
-- GROUP BY clause in various forms, cardinal, alias and constant expression
-explain (verbose, costs off)
+explain (verbose, costs on)
select count(c2) w, c2 x, 5 y, 7.0 z from ft1 group by 2, y, 9.0::int order by 2;
select count(c2) w, c2 x, 5 y, 7.0 z from ft1 group by 2, y, 9.0::int order by 2;
diff --git a/src/backend/optimizer/path/pathkeys.c b/src/backend/optimizer/path/pathkeys.c
index 7913d62625d..9488798d445 100644
--- a/src/backend/optimizer/path/pathkeys.c
+++ b/src/backend/optimizer/path/pathkeys.c
@@ -32,7 +32,7 @@
#include "utils/selfuncs.h"
/* Consider reordering of GROUP BY keys? */
-bool enable_group_by_reordering = true;
+bool enable_group_by_reordering = false;
static bool pathkey_is_redundant(PathKey *new_pathkey, List *pathkeys);
static bool matches_boolean_partition_clause(RestrictInfo *rinfo,
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index 9e8ab1420d9..ebc9904487f 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -1189,7 +1189,7 @@ static struct config_bool ConfigureNamesBool[] =
GUC_EXPLAIN
},
&enable_group_by_reordering,
- true,
+ false,
NULL, NULL, NULL
},
{
diff --git a/src/test/regress/expected/aggregates.out b/src/test/regress/expected/aggregates.out
index 601047fa3dd..26ea59f015e 100644
--- a/src/test/regress/expected/aggregates.out
+++ b/src/test/regress/expected/aggregates.out
@@ -1206,35 +1206,35 @@ select min(f1), max(f1) from minmaxtest;
(1 row)
-- DISTINCT doesn't do anything useful here, but it shouldn't fail
-explain (costs off)
+explain (costs on)
select distinct min(f1), max(f1) from minmaxtest;
- QUERY PLAN
----------------------------------------------------------------------------------------------
- HashAggregate
+ QUERY PLAN
+-----------------------------------------------------------------------------------------------------------------------------------
+ HashAggregate (cost=1.40..1.41 rows=1 width=8)
Group Key: $0, $1
InitPlan 1 (returns $0)
- -> Limit
- -> Merge Append
+ -> Limit (cost=0.64..0.69 rows=1 width=4)
+ -> Merge Append (cost=0.64..394.50 rows=7865 width=4)
Sort Key: minmaxtest.f1
- -> Index Only Scan using minmaxtesti on minmaxtest minmaxtest_1
+ -> Index Only Scan using minmaxtesti on minmaxtest minmaxtest_1 (cost=0.14..16.59 rows=254 width=4)
Index Cond: (f1 IS NOT NULL)
- -> Index Only Scan using minmaxtest1i on minmaxtest1 minmaxtest_2
+ -> Index Only Scan using minmaxtest1i on minmaxtest1 minmaxtest_2 (cost=0.15..92.55 rows=2537 width=4)
Index Cond: (f1 IS NOT NULL)
- -> Index Only Scan Backward using minmaxtest2i on minmaxtest2 minmaxtest_3
+ -> Index Only Scan Backward using minmaxtest2i on minmaxtest2 minmaxtest_3 (cost=0.15..92.55 rows=2537 width=4)
Index Cond: (f1 IS NOT NULL)
- -> Index Only Scan using minmaxtest3i on minmaxtest3 minmaxtest_4
+ -> Index Only Scan using minmaxtest3i on minmaxtest3 minmaxtest_4 (cost=0.14..74.79 rows=2537 width=4)
InitPlan 2 (returns $1)
- -> Limit
- -> Merge Append
+ -> Limit (cost=0.64..0.69 rows=1 width=4)
+ -> Merge Append (cost=0.64..394.50 rows=7865 width=4)
Sort Key: minmaxtest_5.f1 DESC
- -> Index Only Scan Backward using minmaxtesti on minmaxtest minmaxtest_6
+ -> Index Only Scan Backward using minmaxtesti on minmaxtest minmaxtest_6 (cost=0.14..16.59 rows=254 width=4)
Index Cond: (f1 IS NOT NULL)
- -> Index Only Scan Backward using minmaxtest1i on minmaxtest1 minmaxtest_7
+ -> Index Only Scan Backward using minmaxtest1i on minmaxtest1 minmaxtest_7 (cost=0.15..92.55 rows=2537 width=4)
Index Cond: (f1 IS NOT NULL)
- -> Index Only Scan using minmaxtest2i on minmaxtest2 minmaxtest_8
+ -> Index Only Scan using minmaxtest2i on minmaxtest2 minmaxtest_8 (cost=0.15..92.55 rows=2537 width=4)
Index Cond: (f1 IS NOT NULL)
- -> Index Only Scan Backward using minmaxtest3i on minmaxtest3 minmaxtest_9
- -> Result
+ -> Index Only Scan Backward using minmaxtest3i on minmaxtest3 minmaxtest_9 (cost=0.14..74.79 rows=2537 width=4)
+ -> Result (cost=1.38..1.39 rows=1 width=8)
(25 rows)
select distinct min(f1), max(f1) from minmaxtest;
@@ -2454,7 +2454,7 @@ SELECT
format('%60s', i%2) AS v,
i/4 AS c,
i/8 AS d,
- (random() * (10000/8))::int as e --the same as d but no correlation with p
+ mod(i * 7703, 10000/8)::int as e --the same as d but no correlation with p
INTO btg
FROM
generate_series(1, 10000) i;
@@ -2464,127 +2464,127 @@ ANALYZE btg;
SET enable_hashagg=off;
SET max_parallel_workers= 0;
SET max_parallel_workers_per_gather = 0;
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY p, v;
- QUERY PLAN
------------------------------
- GroupAggregate
+ QUERY PLAN
+----------------------------------------------------------------------
+ GroupAggregate (cost=1240.79..1390.80 rows=5001 width=73)
Group Key: p, v
- -> Sort
+ -> Sort (cost=1240.79..1265.79 rows=10000 width=65)
Sort Key: p, v
- -> Seq Scan on btg
+ -> Seq Scan on btg (cost=0.00..243.00 rows=10000 width=65)
(5 rows)
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY v, p;
- QUERY PLAN
------------------------------
- GroupAggregate
+ QUERY PLAN
+----------------------------------------------------------------------
+ GroupAggregate (cost=1240.79..1390.80 rows=5001 width=73)
Group Key: p, v
- -> Sort
+ -> Sort (cost=1240.79..1265.79 rows=10000 width=65)
Sort Key: p, v
- -> Seq Scan on btg
+ -> Seq Scan on btg (cost=0.00..243.00 rows=10000 width=65)
(5 rows)
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY v, p, c;
- QUERY PLAN
------------------------------
- GroupAggregate
+ QUERY PLAN
+----------------------------------------------------------------------
+ GroupAggregate (cost=1335.43..1510.44 rows=5001 width=77)
Group Key: p, c, v
- -> Sort
+ -> Sort (cost=1335.43..1360.43 rows=10000 width=69)
Sort Key: p, c, v
- -> Seq Scan on btg
+ -> Seq Scan on btg (cost=0.00..243.00 rows=10000 width=69)
(5 rows)
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
- QUERY PLAN
------------------------------
- GroupAggregate
+ QUERY PLAN
+----------------------------------------------------------------------
+ GroupAggregate (cost=1759.31..1934.32 rows=5001 width=77)
Group Key: v, p, c
- -> Sort
+ -> Sort (cost=1759.31..1784.31 rows=10000 width=69)
Sort Key: v, p, c
- -> Seq Scan on btg
+ -> Seq Scan on btg (cost=0.00..243.00 rows=10000 width=69)
(5 rows)
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY v, p, d, c;
- QUERY PLAN
-------------------------------
- GroupAggregate
+ QUERY PLAN
+----------------------------------------------------------------------
+ GroupAggregate (cost=1460.43..1660.44 rows=5001 width=81)
Group Key: p, d, c, v
- -> Sort
+ -> Sort (cost=1460.43..1485.43 rows=10000 width=73)
Sort Key: p, d, c, v
- -> Seq Scan on btg
+ -> Seq Scan on btg (cost=0.00..243.00 rows=10000 width=73)
(5 rows)
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
- QUERY PLAN
-------------------------------
- GroupAggregate
+ QUERY PLAN
+----------------------------------------------------------------------
+ GroupAggregate (cost=1893.47..2093.48 rows=5001 width=81)
Group Key: v, p, d, c
- -> Sort
+ -> Sort (cost=1893.47..1918.47 rows=10000 width=73)
Sort Key: v, p, d, c
- -> Seq Scan on btg
+ -> Seq Scan on btg (cost=0.00..243.00 rows=10000 width=73)
(5 rows)
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
- QUERY PLAN
-------------------------------
- GroupAggregate
+ QUERY PLAN
+----------------------------------------------------------------------
+ GroupAggregate (cost=1484.11..1684.12 rows=5001 width=81)
Group Key: p, v, d, c
- -> Sort
+ -> Sort (cost=1484.11..1509.11 rows=10000 width=73)
Sort Key: p, v, d, c
- -> Seq Scan on btg
+ -> Seq Scan on btg (cost=0.00..243.00 rows=10000 width=73)
(5 rows)
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY p, d, e;
- QUERY PLAN
------------------------------
- GroupAggregate
+ QUERY PLAN
+----------------------------------------------------------------------
+ GroupAggregate (cost=1326.27..1501.28 rows=5001 width=20)
Group Key: p, d, e
- -> Sort
+ -> Sort (cost=1326.27..1351.27 rows=10000 width=12)
Sort Key: p, d, e
- -> Seq Scan on btg
+ -> Seq Scan on btg (cost=0.00..243.00 rows=10000 width=12)
(5 rows)
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY p, e, d;
- QUERY PLAN
------------------------------
- GroupAggregate
+ QUERY PLAN
+----------------------------------------------------------------------
+ GroupAggregate (cost=1326.27..1501.28 rows=5001 width=20)
Group Key: p, e, d
- -> Sort
+ -> Sort (cost=1326.27..1351.27 rows=10000 width=12)
Sort Key: p, e, d
- -> Seq Scan on btg
+ -> Seq Scan on btg (cost=0.00..243.00 rows=10000 width=12)
(5 rows)
CREATE STATISTICS btg_dep ON d, e, p FROM btg;
ANALYZE btg;
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY p, d, e;
- QUERY PLAN
------------------------------
- GroupAggregate
+ QUERY PLAN
+----------------------------------------------------------------------
+ GroupAggregate (cost=1326.27..1551.27 rows=10000 width=20)
Group Key: p, d, e
- -> Sort
+ -> Sort (cost=1326.27..1351.27 rows=10000 width=12)
Sort Key: p, d, e
- -> Seq Scan on btg
+ -> Seq Scan on btg (cost=0.00..243.00 rows=10000 width=12)
(5 rows)
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY p, e, d;
- QUERY PLAN
------------------------------
- GroupAggregate
+ QUERY PLAN
+----------------------------------------------------------------------
+ GroupAggregate (cost=1326.27..1551.27 rows=10000 width=20)
Group Key: p, e, d
- -> Sort
+ -> Sort (cost=1326.27..1351.27 rows=10000 width=12)
Sort Key: p, e, d
- -> Seq Scan on btg
+ -> Seq Scan on btg (cost=0.00..243.00 rows=10000 width=12)
(5 rows)
-- GROUP BY optimization by reorder columns by index scan
diff --git a/src/test/regress/sql/aggregates.sql b/src/test/regress/sql/aggregates.sql
index c6e0d7ba2b7..671a9daa696 100644
--- a/src/test/regress/sql/aggregates.sql
+++ b/src/test/regress/sql/aggregates.sql
@@ -425,7 +425,7 @@ explain (costs off)
select min(f1), max(f1) from minmaxtest;
-- DISTINCT doesn't do anything useful here, but it shouldn't fail
-explain (costs off)
+explain (costs on)
select distinct min(f1), max(f1) from minmaxtest;
select distinct min(f1), max(f1) from minmaxtest;
@@ -1033,7 +1033,7 @@ SELECT
format('%60s', i%2) AS v,
i/4 AS c,
i/8 AS d,
- (random() * (10000/8))::int as e --the same as d but no correlation with p
+ mod(i * 7703, 10000/8)::int as e --the same as d but no correlation with p
INTO btg
FROM
generate_series(1, 10000) i;
@@ -1047,40 +1047,40 @@ SET enable_hashagg=off;
SET max_parallel_workers= 0;
SET max_parallel_workers_per_gather = 0;
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY p, v;
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY v, p;
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY v, p, c;
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY v, p, d, c;
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY p, d, e;
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY p, e, d;
CREATE STATISTICS btg_dep ON d, e, p FROM btg;
ANALYZE btg;
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY p, d, e;
-EXPLAIN (COSTS off)
+EXPLAIN (COSTS on)
SELECT count(*) FROM btg GROUP BY p, e, d;
--
2.34.1
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-03-29 02:02 Zhihong Yu <[email protected]>
parent: Tomas Vondra <[email protected]>
0 siblings, 1 reply; 38+ messages in thread
From: Zhihong Yu @ 2022-03-29 02:02 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; +Cc: Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On Mon, Mar 28, 2022 at 5:49 PM Tomas Vondra <[email protected]>
wrote:
> Hi,
>
> Here's a rebased/improved version of the patch, with smaller parts
> addressing various issues. There are seven parts:
>
> 0001 - main part, just rebased
>
> 0002 - replace the debug GUC options with a single GUC to disable the
> optimization if needed
>
> 0003 - minor code cleanup, removal of unnecessary variable
>
> 0004 - various comment fixes (rewordings, typos, ...)
>
> 0005 - a minor code simplification, addressing FIXMEs from 0004
>
> 0006 - adds the new GUC to the docs
>
> 0007 - demonstrates plan changes with a disabled optimization
>
> The first 6 parts should be squashed and committed at one, I only kept
> them separate for clarity. The 0007 is merely a demonstration of the new
> GUC and that it disables the optimization.
>
> > Agree. Because it is a kind of automation we should allow user to switch
> > it off in the case of problems or manual tuning.
> > > Also, I looked through this patch. It has some minor problems:
> > 1. Multiple typos in the patch comment.
>
> I went through the comments and checked all of them for grammar mistakes
> and typos using a word processor, so hopefully that should be OK. But
> maybe there's still something wrong.
>
> > 2. The term 'cardinality of a key' - may be replace with 'number of
> > duplicates'?
>
> No, cardinality means "number of distinct values", so "duplicates" would
> be wrong. And I think "cardinality" is well established term, so I think
> it's fine.
>
> BTW I named the GUC enable_group_by_reordering, I wonder if it should be
> named differently, e.g. enable_groupby_reordering? Opinions?
>
>
> regards
>
> --
> Tomas Vondra
> EnterpriseDB: http://www.enterprisedb.com
> The Enterprise PostgreSQL Company
Hi,
For 0001-Optimize-order-of-GROUP-BY-keys-20220328.patch:
multiple parametes need to be
parametes -> parameters
leave more expensive comparions
comparions -> comparisons
+ if (has_fake_var == false)
The above can be written as:
if (!has_fake_var)
+ nGroups = ceil(2.0 + sqrt(tuples) * (i + 1) /
list_length(pathkeys));
Looks like the value of tuples doesn't change inside the loop.
You can precompute sqrt(tuples) outside the loop and store the value in a
variable.
+ return -1;
+ else if (a->cost == b->cost)
+ return 0;
+ return 1;
the keyword 'else' is not needed.
+ * Returns newly allocated lists. If no reordering is possible (or needed),
+ * the lists are set to NIL.
+ */
+static bool
+get_cheapest_group_keys_order(PlannerInfo *root, double nrows,
It seems the comment for return value doesn't match the bool return type.
+ /* If this optimization is disabled, we're done. */
+ if (!debug_cheapest_group_by)
It seems enable_cheapest_group_by would be better name for the flag.
Cheers
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-03-29 18:00 Tomas Vondra <[email protected]>
parent: Zhihong Yu <[email protected]>
0 siblings, 1 reply; 38+ messages in thread
From: Tomas Vondra @ 2022-03-29 18:00 UTC (permalink / raw)
To: Zhihong Yu <[email protected]>; +Cc: Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On 3/29/22 04:02, Zhihong Yu wrote:
>
>
> On Mon, Mar 28, 2022 at 5:49 PM Tomas Vondra
> <[email protected] <mailto:[email protected]>>
> wrote:
>
> Hi,
>
> Here's a rebased/improved version of the patch, with smaller parts
> addressing various issues. There are seven parts:
>
> 0001 - main part, just rebased
>
> 0002 - replace the debug GUC options with a single GUC to disable the
> optimization if needed
>
> 0003 - minor code cleanup, removal of unnecessary variable
>
> 0004 - various comment fixes (rewordings, typos, ...)
>
> 0005 - a minor code simplification, addressing FIXMEs from 0004
>
> 0006 - adds the new GUC to the docs
>
> 0007 - demonstrates plan changes with a disabled optimization
>
> The first 6 parts should be squashed and committed at one, I only kept
> them separate for clarity. The 0007 is merely a demonstration of the new
> GUC and that it disables the optimization.
>
> > Agree. Because it is a kind of automation we should allow user to
> switch
> > it off in the case of problems or manual tuning.
> > > Also, I looked through this patch. It has some minor problems:
> > 1. Multiple typos in the patch comment.
>
> I went through the comments and checked all of them for grammar mistakes
> and typos using a word processor, so hopefully that should be OK. But
> maybe there's still something wrong.
>
> > 2. The term 'cardinality of a key' - may be replace with 'number of
> > duplicates'?
>
> No, cardinality means "number of distinct values", so "duplicates" would
> be wrong. And I think "cardinality" is well established term, so I think
> it's fine.
>
> BTW I named the GUC enable_group_by_reordering, I wonder if it should be
> named differently, e.g. enable_groupby_reordering? Opinions?
>
>
> regards
>
> --
> Tomas Vondra
> EnterpriseDB: http://www.enterprisedb.com <http://www.enterprisedb.com;
> The Enterprise PostgreSQL Company
>
> Hi,
>
> For 0001-Optimize-order-of-GROUP-BY-keys-20220328.patch:
>
> multiple parametes need to be
>
> parametes -> parameters
>
> leave more expensive comparions
>
> comparions -> comparisons
>
> + if (has_fake_var == false)
>
> The above can be written as:
>
> if (!has_fake_var)
>
All of this was already fixed in one of the subsequent "fixup" patches.
Attached is a patch merging 0001-0006, which is what I proposed to get
committed.
> + nGroups = ceil(2.0 + sqrt(tuples) * (i + 1) /
> list_length(pathkeys));
>
> Looks like the value of tuples doesn't change inside the loop.
> You can precompute sqrt(tuples) outside the loop and store the value in
> a variable.
>
IMHO it makes the formula harder to read, and the effect is not going to
be measurable - we're processing only a couple elements. If the loop was
~100 iterations or more, maybe it'd have impact.
> + return -1;
> + else if (a->cost == b->cost)
> + return 0;
> + return 1;
>
> the keyword 'else' is not needed.
>
True, but this is how comparators are usually implemented.
> + * Returns newly allocated lists. If no reordering is possible (or needed),
> + * the lists are set to NIL.
> + */
> +static bool
> +get_cheapest_group_keys_order(PlannerInfo *root, double nrows,
>
> It seems the comment for return value doesn't match the bool return type.
>
Yup, this is a valid point. I've fixed/reworded the comment a bit.
> + /* If this optimization is disabled, we're done. */
> + if (!debug_cheapest_group_by)
>
> It seems enable_cheapest_group_by would be better name for the flag.
>
This was renamed to enable_order_by_reordering in one of the fixup patches.
Attached is a patch merging 0001 and all the fixup patches, i.e. the
patch I plan to commit.
There was a minor test failure - the new GUC was not added to the sample
config file, so 003_check_guc.pl was failing.
I'm not including the 0007 part, because that's meant to demonstrate
plan changes with disabled optimization, which would confuse cfbot.
regards
regards
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
Attachments:
[text/x-patch] 0001-Optimize-order-of-GROUP-BY-keys-20220329.patch (102.8K, ../../[email protected]/2-0001-Optimize-order-of-GROUP-BY-keys-20220329.patch)
download | inline diff:
From 66835c4ae52dc61d580b84ab15004c8f46919550 Mon Sep 17 00:00:00 2001
From: Tomas Vondra <[email protected]>
Date: Fri, 21 Jan 2022 20:22:14 +0100
Subject: [PATCH 1/2] Optimize order of GROUP BY keys
When evaluating a query with a multi-column GROUP BY clause using sort,
the cost may be heavily dependent on the order in which the keys are
compared when building the groups. Grouping does not imply any ordering,
so we're allowed to compare the keys in arbitrary order, and a Hash Agg
leverages this. But for Group Agg, we simply compared keys in the order
as specified in the query. This commit explores alternative ordering of
the keys, trying to find a cheaper one.
To pick the optimal order of comparisons, multiple parametes need to be
considered, both local and global ones. Intuitively, it's better to
leave more expensive comparions (for complex data types etc.) at the
very end, because maybe one of the earlier comparisons will be enough to
make a decision, and the expensive comparison will be unnecessary. The
other local parameter is cardinality - the higher the cardinality of a
key, the lower the probability we'll need to compare additional keys. We
explore possible orderings by a combination of exhaustive and greedy
approach, and pick the cheapest ordering.
But there are global parameters too, in the sense that we can't fully
evaluate them at the point when creating the sort. For example, there
might be an ORDER BY clause, or another ordering-dependent operation,
higher up in the query plan. The path may be already sorted by (some
of) the keys, in which case we can either use incremental sort or even
skip the sort entirely. So we consider various additional pathkeys in
addition to the "cheapers comparion" order.
The original patch was proposed by Teodor Sigaev, and later reworked and
improved by Dmitry Dolgov. Reviews by a number of people, including me,
Andrey Lepikhov, Claudio Freire and Ibrar Ahmed.
Author: Dmitry Dolgov, Teodor Sigaev, Tomas Vondra
Reviewed-by: Tomas Vondra, Andrey Lepikhov, Claudio Freire, Ibrar Ahmed, Zhihong Yu
Discussion: https://postgr.es/m/7c79e6a5-8597-74e8-0671-1c39d124c9d6%40sigaev.ru
Discussion: https://postgr.es/m/CA%2Bq6zcW_4o2NC0zutLkOJPsFt80megSpX_dVRo6GK9PC-Jx_Ag%40mail.gmail.com
---
.../postgres_fdw/expected/postgres_fdw.out | 15 +-
doc/src/sgml/config.sgml | 14 +
src/backend/optimizer/path/costsize.c | 366 +++++++++-
src/backend/optimizer/path/equivclass.c | 13 +-
src/backend/optimizer/path/pathkeys.c | 569 +++++++++++++++
src/backend/optimizer/plan/planner.c | 649 ++++++++++--------
src/backend/optimizer/util/pathnode.c | 2 +-
src/backend/utils/adt/selfuncs.c | 38 +-
src/backend/utils/misc/guc.c | 10 +
src/backend/utils/misc/postgresql.conf.sample | 1 +
src/include/nodes/nodes.h | 1 +
src/include/nodes/pathnodes.h | 10 +
src/include/optimizer/cost.h | 4 +-
src/include/optimizer/paths.h | 7 +
src/include/utils/selfuncs.h | 5 +
src/test/regress/expected/aggregates.out | 244 ++++++-
.../regress/expected/incremental_sort.out | 2 +-
src/test/regress/expected/join.out | 51 +-
.../regress/expected/partition_aggregate.out | 136 ++--
src/test/regress/expected/partition_join.out | 75 +-
src/test/regress/expected/sysviews.out | 3 +-
src/test/regress/expected/union.out | 60 +-
src/test/regress/sql/aggregates.sql | 99 +++
src/test/regress/sql/incremental_sort.sql | 2 +-
24 files changed, 1882 insertions(+), 494 deletions(-)
diff --git a/contrib/postgres_fdw/expected/postgres_fdw.out b/contrib/postgres_fdw/expected/postgres_fdw.out
index f210f911880..2ffc836824a 100644
--- a/contrib/postgres_fdw/expected/postgres_fdw.out
+++ b/contrib/postgres_fdw/expected/postgres_fdw.out
@@ -2741,16 +2741,13 @@ select c2 * (random() <= 1)::int as c2 from ft2 group by c2 * (random() <= 1)::i
-- GROUP BY clause in various forms, cardinal, alias and constant expression
explain (verbose, costs off)
select count(c2) w, c2 x, 5 y, 7.0 z from ft1 group by 2, y, 9.0::int order by 2;
- QUERY PLAN
----------------------------------------------------------------------------------------
- Sort
+ QUERY PLAN
+------------------------------------------------------------------------------------------------------------
+ Foreign Scan
Output: (count(c2)), c2, 5, 7.0, 9
- Sort Key: ft1.c2
- -> Foreign Scan
- Output: (count(c2)), c2, 5, 7.0, 9
- Relations: Aggregate on (public.ft1)
- Remote SQL: SELECT count(c2), c2, 5, 7.0, 9 FROM "S 1"."T 1" GROUP BY 2, 3, 5
-(7 rows)
+ Relations: Aggregate on (public.ft1)
+ Remote SQL: SELECT count(c2), c2, 5, 7.0, 9 FROM "S 1"."T 1" GROUP BY 2, 3, 5 ORDER BY c2 ASC NULLS LAST
+(4 rows)
select count(c2) w, c2 x, 5 y, 7.0 z from ft1 group by 2, y, 9.0::int order by 2;
w | x | y | z
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index 05df48131d7..0c0de61b855 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -4961,6 +4961,20 @@ ANY <replaceable class="parameter">num_sync</replaceable> ( <replaceable class="
</listitem>
</varlistentry>
+ <varlistentry id="guc-enable-groupby-reordering" xreflabel="enable_group_by_reordering">
+ <term><varname>enable_group_by_reordering</varname> (<type>boolean</type>)
+ <indexterm>
+ <primary><varname>enable_group_by_reordering</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Enables or disables reodering of keys in <literal>GROUP BY</literal>
+ clause. The default is <literal>on</literal>.
+ </para>
+ </listitem>
+ </varlistentry>
+
<varlistentry id="guc-enable-hashagg" xreflabel="enable_hashagg">
<term><varname>enable_hashagg</varname> (<type>boolean</type>)
<indexterm>
diff --git a/src/backend/optimizer/path/costsize.c b/src/backend/optimizer/path/costsize.c
index 1b07ea392d9..3e1daf8b893 100644
--- a/src/backend/optimizer/path/costsize.c
+++ b/src/backend/optimizer/path/costsize.c
@@ -1755,6 +1755,322 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
rterm->pathtarget->width);
}
+/*
+ * is_fake_var
+ * Workaround for generate_append_tlist() which generates fake Vars with
+ * varno == 0, that will cause a fail of estimate_num_group() call
+ *
+ * XXX Ummm, why would estimate_num_group fail with this?
+ */
+static bool
+is_fake_var(Expr *expr)
+{
+ if (IsA(expr, RelabelType))
+ expr = (Expr *) ((RelabelType *) expr)->arg;
+
+ return (IsA(expr, Var) && ((Var *) expr)->varno == 0);
+}
+
+/*
+ * get_width_cost_multiplier
+ * Returns relative complexity of comparing two values based on its width.
+ * The idea behind is that the comparison becomes more expensive the longer the
+ * value is. Return value is in cpu_operator_cost units.
+ */
+static double
+get_width_cost_multiplier(PlannerInfo *root, Expr *expr)
+{
+ double width = -1.0; /* fake value */
+
+ if (IsA(expr, RelabelType))
+ expr = (Expr *) ((RelabelType *) expr)->arg;
+
+ /* Try to find actual stat in corresponding relation */
+ if (IsA(expr, Var))
+ {
+ Var *var = (Var *) expr;
+
+ if (var->varno > 0 && var->varno < root->simple_rel_array_size)
+ {
+ RelOptInfo *rel = root->simple_rel_array[var->varno];
+
+ if (rel != NULL &&
+ var->varattno >= rel->min_attr &&
+ var->varattno <= rel->max_attr)
+ {
+ int ndx = var->varattno - rel->min_attr;
+
+ if (rel->attr_widths[ndx] > 0)
+ width = rel->attr_widths[ndx];
+ }
+ }
+ }
+
+ /* Didn't find any actual stats, try using type width instead. */
+ if (width < 0.0)
+ {
+ Node *node = (Node*) expr;
+
+ width = get_typavgwidth(exprType(node), exprTypmod(node));
+ }
+
+ /*
+ * Values are passed as Datum type, so comparisons can't be cheaper than
+ * comparing a Datum value.
+ *
+ * FIXME I find this reasoning questionable. We may pass int2, and comparing
+ * it is probably a bit cheaper than comparing a bigint.
+ */
+ if (width <= sizeof(Datum))
+ return 1.0;
+
+ /*
+ * We consider the cost of a comparison not to be directly proportional to
+ * width of the argument, because widths of the arguments could be slightly
+ * different (we only know the average width for the whole column). So we
+ * use log16(width) as an estimate.
+ */
+ return 1.0 + 0.125 * LOG2(width / sizeof(Datum));
+}
+
+/*
+ * compute_cpu_sort_cost
+ * compute CPU cost of sort (i.e. in-memory)
+ *
+ * The main thing we need to calculate to estimate sort CPU costs is the number
+ * of calls to the comparator functions. The difficulty is that for multi-column
+ * sorts there may be different data types involved (for some of which the calls
+ * may be much more expensive). Furthermore, columns may have a very different
+ * number of distinct values - the higher the number, the fewer comparisons will
+ * be needed for the following columns.
+ *
+ * The algorithm is incremental - we add pathkeys one by one, and at each step we
+ * estimate the number of necessary comparisons (based on the number of distinct
+ * groups in the current pathkey prefix and the new pathkey), and the comparison
+ * costs (which is data type specific).
+ *
+ * Estimation of the number of comparisons is based on ideas from:
+ *
+ * "Quicksort Is Optimal", Robert Sedgewick, Jon Bentley, 2002
+ * [https://www.cs.princeton.edu/~rs/talks/QuicksortIsOptimal.pdf]
+ *
+ * In term of that paper, let N - number of tuples, Xi - number of identical
+ * tuples with value Ki, then the estimate of number of comparisons is:
+ *
+ * log(N! / (X1! * X2! * ..)) ~ sum(Xi * log(N/Xi))
+ *
+ * We assume all Xi the same because now we don't have any estimation of
+ * group sizes, we have only know the estimate of number of groups (distinct
+ * values). In that case, formula becomes:
+ *
+ * N * log(NumberOfGroups)
+ *
+ * For multi-column sorts we need to estimate the number of comparisons for
+ * each individual column - for example with columns (c1, c2, ..., ck) we
+ * can estimate that number of comparisons on ck is roughly
+ *
+ * ncomparisons(c1, c2, ..., ck) / ncomparisons(c1, c2, ..., c(k-1))
+ *
+ * Let k be a column number, Gk - number of groups defined by k columns, and Fk
+ * the cost of the comparison is
+ *
+ * N * sum( Fk * log(Gk) )
+ *
+ * Note: We also consider column width, not just the comparator cost.
+ *
+ * NOTE: some callers currently pass NIL for pathkeys because they
+ * can't conveniently supply the sort keys. In this case, it will fallback to
+ * simple comparison cost estimate.
+ */
+static Cost
+compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
+ Cost comparison_cost, double tuples, double output_tuples,
+ bool heapSort)
+{
+ Cost per_tuple_cost = 0.0;
+ ListCell *lc;
+ List *pathkeyExprs = NIL;
+ double tuplesPerPrevGroup = tuples;
+ double totalFuncCost = 1.0;
+ bool has_fake_var = false;
+ int i = 0;
+ Oid prev_datatype = InvalidOid;
+ List *cache_varinfos = NIL;
+
+ /* fallback if pathkeys is unknown */
+ if (list_length(pathkeys) == 0)
+ {
+ /*
+ * If we'll use a bounded heap-sort keeping just K tuples in memory, for
+ * a total number of tuple comparisons of N log2 K; but the constant
+ * factor is a bit higher than for quicksort. Tweak it so that the cost
+ * curve is continuous at the crossover point.
+ */
+ output_tuples = (heapSort) ? 2.0 * output_tuples : tuples;
+ per_tuple_cost += 2.0 * cpu_operator_cost * LOG2(output_tuples);
+
+ /* add cost provided by caller */
+ per_tuple_cost += comparison_cost;
+
+ return per_tuple_cost * tuples;
+ }
+
+ /*
+ * Computing total cost of sorting takes into account:
+ * - per column comparison function cost
+ * - we try to compute needed number of comparison per column
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey*) lfirst(lc);
+ EquivalenceMember *em;
+ double nGroups,
+ correctedNGroups;
+ Cost funcCost = 1.0;
+
+ /*
+ * We believe that equivalence members aren't very different, so, to
+ * estimate cost we consider just the first member.
+ */
+ em = (EquivalenceMember *) linitial(pathkey->pk_eclass->ec_members);
+
+ if (em->em_datatype != InvalidOid)
+ {
+ /* do not lookup funcCost if the data type is the same */
+ if (prev_datatype != em->em_datatype)
+ {
+ Oid sortop;
+ QualCost cost;
+
+ sortop = get_opfamily_member(pathkey->pk_opfamily,
+ em->em_datatype, em->em_datatype,
+ pathkey->pk_strategy);
+
+ cost.startup = 0;
+ cost.per_tuple = 0;
+ add_function_cost(root, get_opcode(sortop), NULL, &cost);
+
+ /*
+ * add_function_cost returns the product of cpu_operator_cost
+ * and procost, but we need just procost, co undo that.
+ */
+ funcCost = cost.per_tuple / cpu_operator_cost;
+
+ prev_datatype = em->em_datatype;
+ }
+ }
+
+ /* factor in the width of the values in this column */
+ funcCost *= get_width_cost_multiplier(root, em->em_expr);
+
+ /* now we have per-key cost, so add to the running total */
+ totalFuncCost += funcCost;
+
+ /* remember if we have found a fake Var in pathkeys */
+ has_fake_var |= is_fake_var(em->em_expr);
+ pathkeyExprs = lappend(pathkeyExprs, em->em_expr);
+
+ /*
+ * We need to calculate the number of comparisons for this column, which
+ * requires knowing the group size. So we estimate the number of groups
+ * by calling estimate_num_groups_incremental(), which estimates the
+ * group size for "new" pathkeys.
+ *
+ * Note: estimate_num_groups_incremntal does not handle fake Vars, so use
+ * a default estimate otherwise.
+ */
+ if (!has_fake_var)
+ nGroups = estimate_num_groups_incremental(root, pathkeyExprs,
+ tuplesPerPrevGroup, NULL, NULL,
+ &cache_varinfos,
+ list_length(pathkeyExprs) - 1);
+ else if (tuples > 4.0)
+ /*
+ * Use geometric mean as estimation if there are no stats.
+ *
+ * We don't use DEFAULT_NUM_DISTINCT here, because that’s used for
+ * a single column, but here we’re dealing with multiple columns.
+ */
+ nGroups = ceil(2.0 + sqrt(tuples) * (i + 1) / list_length(pathkeys));
+ else
+ nGroups = tuples;
+
+ /*
+ * Presorted keys are not considered in the cost above, but we still do
+ * have to compare them in the qsort comparator. So make sure to factor
+ * in the cost in that case.
+ */
+ if (i >= nPresortedKeys)
+ {
+ if (heapSort)
+ {
+ /* have to keep at least one group, and a multiple of group size */
+ correctedNGroups = ceil(output_tuples / tuplesPerPrevGroup);
+ }
+ else
+ /* all groups in the input */
+ correctedNGroups = nGroups;
+
+ correctedNGroups = Max(1.0, ceil(correctedNGroups));
+
+ per_tuple_cost += totalFuncCost * LOG2(correctedNGroups);
+ }
+
+ i++;
+
+ /*
+ * Uniform distributions with all groups being of the same size are the
+ * best case, with nice smooth behavior. Real-world distributions tend
+ * not to be uniform, though, and we don’t have any reliable easy-to-use
+ * information. As a basic defense against skewed distributions, we use
+ * a 1.5 factor to make the expected group a bit larger, but we need to
+ * be careful not to make the group larger than in the preceding step.
+ */
+ tuplesPerPrevGroup = Min(tuplesPerPrevGroup,
+ ceil(1.5 * tuplesPerPrevGroup / nGroups));
+
+ /*
+ * Once we get single-row group, it means tuples in the group are unique
+ * and we can skip all remaining columns.
+ */
+ if (tuplesPerPrevGroup <= 1.0)
+ break;
+ }
+
+ list_free(pathkeyExprs);
+
+ /* per_tuple_cost is in cpu_operator_cost units */
+ per_tuple_cost *= cpu_operator_cost;
+
+ /*
+ * Accordingly to "Introduction to algorithms", Thomas H. Cormen, Charles E.
+ * Leiserson, Ronald L. Rivest, ISBN 0-07-013143-0, quicksort estimation
+ * formula has additional term proportional to number of tuples (See Chapter
+ * 8.2 and Theorem 4.1). That affects cases with a low number of tuples,
+ * approximately less than 1e4. We could implement it as an additional
+ * multiplier under the logarithm, but we use a bit more complex formula
+ * which takes into account the number of unique tuples and it’s not clear
+ * how to combine the multiplier with the number of groups. Estimate it as
+ * 10 in cpu_operator_cost unit.
+ */
+ per_tuple_cost += 10 * cpu_operator_cost;
+
+ per_tuple_cost += comparison_cost;
+
+ return tuples * per_tuple_cost;
+}
+
+/*
+ * simple wrapper just to estimate best sort path
+ */
+Cost
+cost_sort_estimate(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
+ double tuples)
+{
+ return compute_cpu_sort_cost(root, pathkeys, nPresortedKeys,
+ 0, tuples, tuples, false);
+}
+
/*
* cost_tuplesort
* Determines and returns the cost of sorting a relation using tuplesort,
@@ -1771,7 +2087,7 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
* number of initial runs formed and M is the merge order used by tuplesort.c.
* Since the average initial run should be about sort_mem, we have
* disk traffic = 2 * relsize * ceil(logM(p / sort_mem))
- * cpu = comparison_cost * t * log2(t)
+ * and cpu cost (computed by compute_cpu_sort_cost()).
*
* If the sort is bounded (i.e., only the first k result tuples are needed)
* and k tuples can fit into sort_mem, we use a heap method that keeps only
@@ -1790,9 +2106,11 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
* 'comparison_cost' is the extra cost per comparison, if any
* 'sort_mem' is the number of kilobytes of work memory allowed for the sort
* 'limit_tuples' is the bound on the number of output tuples; -1 if no bound
+ * 'startup_cost' is expected to be 0 at input. If there is "input cost" it should
+ * be added by caller later
*/
static void
-cost_tuplesort(Cost *startup_cost, Cost *run_cost,
+cost_tuplesort(PlannerInfo *root, List *pathkeys, Cost *startup_cost, Cost *run_cost,
double tuples, int width,
Cost comparison_cost, int sort_mem,
double limit_tuples)
@@ -1809,9 +2127,6 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
if (tuples < 2.0)
tuples = 2.0;
- /* Include the default cost-per-comparison */
- comparison_cost += 2.0 * cpu_operator_cost;
-
/* Do we have a useful LIMIT? */
if (limit_tuples > 0 && limit_tuples < tuples)
{
@@ -1835,12 +2150,10 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
double log_runs;
double npageaccesses;
- /*
- * CPU costs
- *
- * Assume about N log2 N comparisons
- */
- *startup_cost = comparison_cost * tuples * LOG2(tuples);
+ /* CPU costs */
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ tuples, false);
/* Disk costs */
@@ -1856,18 +2169,17 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
}
else if (tuples > 2 * output_tuples || input_bytes > sort_mem_bytes)
{
- /*
- * We'll use a bounded heap-sort keeping just K tuples in memory, for
- * a total number of tuple comparisons of N log2 K; but the constant
- * factor is a bit higher than for quicksort. Tweak it so that the
- * cost curve is continuous at the crossover point.
- */
- *startup_cost = comparison_cost * tuples * LOG2(2.0 * output_tuples);
+ /* We'll use a bounded heap-sort keeping just K tuples in memory. */
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ output_tuples, true);
}
else
{
/* We'll use plain quicksort on all the input tuples */
- *startup_cost = comparison_cost * tuples * LOG2(tuples);
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ tuples, false);
}
/*
@@ -1900,8 +2212,8 @@ cost_incremental_sort(Path *path,
double input_tuples, int width, Cost comparison_cost, int sort_mem,
double limit_tuples)
{
- Cost startup_cost = 0,
- run_cost = 0,
+ Cost startup_cost,
+ run_cost,
input_run_cost = input_total_cost - input_startup_cost;
double group_tuples,
input_groups;
@@ -1986,7 +2298,7 @@ cost_incremental_sort(Path *path,
* pessimistic about incremental sort performance and increase its average
* group size by half.
*/
- cost_tuplesort(&group_startup_cost, &group_run_cost,
+ cost_tuplesort(root, pathkeys, &group_startup_cost, &group_run_cost,
1.5 * group_tuples, width, comparison_cost, sort_mem,
limit_tuples);
@@ -1994,7 +2306,7 @@ cost_incremental_sort(Path *path,
* Startup cost of incremental sort is the startup cost of its first group
* plus the cost of its input.
*/
- startup_cost += group_startup_cost
+ startup_cost = group_startup_cost
+ input_startup_cost + group_input_run_cost;
/*
@@ -2003,7 +2315,7 @@ cost_incremental_sort(Path *path,
* group, plus the total cost to process the remaining groups, plus the
* remaining cost of input.
*/
- run_cost += group_run_cost
+ run_cost = group_run_cost
+ (group_run_cost + group_startup_cost) * (input_groups - 1)
+ group_input_run_cost * (input_groups - 1);
@@ -2043,7 +2355,7 @@ cost_sort(Path *path, PlannerInfo *root,
Cost startup_cost;
Cost run_cost;
- cost_tuplesort(&startup_cost, &run_cost,
+ cost_tuplesort(root, pathkeys, &startup_cost, &run_cost,
tuples, width,
comparison_cost, sort_mem,
limit_tuples);
@@ -2141,7 +2453,7 @@ append_nonpartial_cost(List *subpaths, int numpaths, int parallel_workers)
* Determines and returns the cost of an Append node.
*/
void
-cost_append(AppendPath *apath)
+cost_append(AppendPath *apath, PlannerInfo *root)
{
ListCell *l;
@@ -2209,7 +2521,7 @@ cost_append(AppendPath *apath)
* any child.
*/
cost_sort(&sort_path,
- NULL, /* doesn't currently need root */
+ root,
pathkeys,
subpath->total_cost,
subpath->rows,
diff --git a/src/backend/optimizer/path/equivclass.c b/src/backend/optimizer/path/equivclass.c
index 8c6770de972..258302840f8 100644
--- a/src/backend/optimizer/path/equivclass.c
+++ b/src/backend/optimizer/path/equivclass.c
@@ -681,7 +681,18 @@ get_eclass_for_sort_expr(PlannerInfo *root,
if (opcintype == cur_em->em_datatype &&
equal(expr, cur_em->em_expr))
- return cur_ec; /* Match! */
+ {
+ /*
+ * Match!
+ *
+ * Copy the sortref if it wasn't set yet. That may happen if the
+ * ec was constructed from WHERE clause, i.e. it doesn't have a
+ * target reference at all.
+ */
+ if (cur_ec->ec_sortref == 0 && sortref > 0)
+ cur_ec->ec_sortref = sortref;
+ return cur_ec;
+ }
}
}
diff --git a/src/backend/optimizer/path/pathkeys.c b/src/backend/optimizer/path/pathkeys.c
index 86a35cdef17..75fe03fd04b 100644
--- a/src/backend/optimizer/path/pathkeys.c
+++ b/src/backend/optimizer/path/pathkeys.c
@@ -17,17 +17,22 @@
*/
#include "postgres.h"
+#include "miscadmin.h"
#include "access/stratnum.h"
#include "catalog/pg_opfamily.h"
#include "nodes/makefuncs.h"
#include "nodes/nodeFuncs.h"
#include "nodes/plannodes.h"
+#include "optimizer/cost.h"
#include "optimizer/optimizer.h"
#include "optimizer/pathnode.h"
#include "optimizer/paths.h"
#include "partitioning/partbounds.h"
#include "utils/lsyscache.h"
+#include "utils/selfuncs.h"
+/* Consider reordering of GROUP BY keys? */
+bool enable_group_by_reordering = true;
static bool pathkey_is_redundant(PathKey *new_pathkey, List *pathkeys);
static bool matches_boolean_partition_clause(RestrictInfo *rinfo,
@@ -334,6 +339,517 @@ pathkeys_contained_in(List *keys1, List *keys2)
return false;
}
+/*
+ * group_keys_reorder_by_pathkeys
+ * Reorder GROUP BY keys to match pathkeys of input path.
+ *
+ * Function returns new lists (pathkeys and clauses), original GROUP BY lists
+ * stay untouched.
+ *
+ * Returns the number of GROUP BY keys with a matching pathkey.
+ */
+int
+group_keys_reorder_by_pathkeys(List *pathkeys, List **group_pathkeys,
+ List **group_clauses)
+{
+ List *new_group_pathkeys= NIL,
+ *new_group_clauses = NIL;
+ ListCell *lc;
+ int n;
+
+ if (pathkeys == NIL || *group_pathkeys == NIL)
+ return 0;
+
+ /*
+ * Walk the pathkeys (determining ordering of the input path) and see if
+ * there's a matching GROUP BY key. If we find one, we append it to the
+ * list, and do the same for the clauses.
+ *
+ * Once we find the first pathkey without a matching GROUP BY key, the rest
+ * of the pathkeys are useless and can't be used to evaluate the grouping,
+ * so we abort the loop and ignore the remaining pathkeys.
+ *
+ * XXX Pathkeys are built in a way to allow simply comparing pointers.
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(lc);
+ SortGroupClause *sgc;
+
+ /* abort on first mismatch */
+ if (!list_member_ptr(*group_pathkeys, pathkey))
+ break;
+
+ new_group_pathkeys = lappend(new_group_pathkeys, pathkey);
+
+ sgc = get_sortgroupref_clause(pathkey->pk_eclass->ec_sortref,
+ *group_clauses);
+
+ new_group_clauses = lappend(new_group_clauses, sgc);
+ }
+
+ /* remember the number of pathkeys with a matching GROUP BY key */
+ n = list_length(new_group_pathkeys);
+
+ /* append the remaining group pathkeys (will be treated as not sorted) */
+ *group_pathkeys = list_concat_unique_ptr(new_group_pathkeys,
+ *group_pathkeys);
+ *group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+
+ return n;
+}
+
+/*
+ * Used to generate all permutations of a pathkey list.
+ */
+typedef struct PathkeyMutatorState {
+ List *elemsList;
+ ListCell **elemCells;
+ void **elems;
+ int *positions;
+ int mutatorNColumns;
+ int count;
+} PathkeyMutatorState;
+
+
+/*
+ * PathkeyMutatorInit
+ * Initialize state of the permutation generator.
+ *
+ * We want to generate permutations of elements in the "elems" list. We may want
+ * to skip some number of elements at the beginning (when treating as presorted)
+ * or at the end (we only permute a limited number of group keys).
+ *
+ * The list is decomposed into elements, and we also keep pointers to individual
+ * cells. This allows us to build the permuted list quickly and cheaply, without
+ * creating any copies.
+ */
+static void
+PathkeyMutatorInit(PathkeyMutatorState *state, List *elems, int start, int end)
+{
+ int i;
+ int n = end - start;
+ ListCell *lc;
+
+ memset(state, 0, sizeof(*state));
+
+ state->mutatorNColumns = n;
+
+ state->elemsList = list_copy(elems);
+
+ state->elems = palloc(sizeof(void*) * n);
+ state->elemCells = palloc(sizeof(ListCell*) * n);
+ state->positions = palloc(sizeof(int) * n);
+
+ i = 0;
+ for_each_cell(lc, state->elemsList, list_nth_cell(state->elemsList, start))
+ {
+ state->elemCells[i] = lc;
+ state->elems[i] = lfirst(lc);
+ state->positions[i] = i + 1;
+ i++;
+
+ if (i >= n)
+ break;
+ }
+}
+
+/* Swap two elements of an array. */
+static void
+PathkeyMutatorSwap(int *a, int i, int j)
+{
+ int s = a[i];
+
+ a[i] = a[j];
+ a[j] = s;
+}
+
+/*
+ * Generate the next permutation of elements.
+ */
+static bool
+PathkeyMutatorNextSet(int *a, int n)
+{
+ int j, k, l, r;
+
+ j = n - 2;
+
+ while (j >= 0 && a[j] >= a[j + 1])
+ j--;
+
+ if (j < 0)
+ return false;
+
+ k = n - 1;
+
+ while (k >= 0 && a[j] >= a[k])
+ k--;
+
+ PathkeyMutatorSwap(a, j, k);
+
+ l = j + 1;
+ r = n - 1;
+
+ while (l < r)
+ PathkeyMutatorSwap(a, l++, r--);
+
+ return true;
+}
+
+/*
+ * PathkeyMutatorNext
+ * Generate the next permutation of list of elements.
+ *
+ * Returns the next permutation (as a list of elements) or NIL if there are no
+ * more permutations.
+ */
+static List *
+PathkeyMutatorNext(PathkeyMutatorState *state)
+{
+ int i;
+
+ state->count++;
+
+ /* first permutation is original list */
+ if (state->count == 1)
+ return state->elemsList;
+
+ /* when there are no more permutations, return NIL */
+ if (!PathkeyMutatorNextSet(state->positions, state->mutatorNColumns))
+ {
+ pfree(state->elems);
+ pfree(state->elemCells);
+ pfree(state->positions);
+
+ list_free(state->elemsList);
+
+ return NIL;
+ }
+
+ /* update the list cells to point to the right elements */
+ for(i = 0; i < state->mutatorNColumns; i++)
+ lfirst(state->elemCells[i]) =
+ (void *) state->elems[ state->positions[i] - 1 ];
+
+ return state->elemsList;
+}
+
+/*
+ * Cost of comparing pathkeys.
+ */
+typedef struct PathkeySortCost
+{
+ Cost cost;
+ PathKey *pathkey;
+} PathkeySortCost;
+
+static int
+pathkey_sort_cost_comparator(const void *_a, const void *_b)
+{
+ const PathkeySortCost *a = (PathkeySortCost *) _a;
+ const PathkeySortCost *b = (PathkeySortCost *) _b;
+
+ if (a->cost < b->cost)
+ return -1;
+ else if (a->cost == b->cost)
+ return 0;
+ return 1;
+}
+
+/*
+ * get_cheapest_group_keys_order
+ * Reorders the group pathkeys/clauses to minimize the comparison cost.
+ *
+ * Given a list of pathkeys, we try to reorder them in a way that minimizes
+ * the CPU cost of sorting. This depends mainly on the cost of comparator
+ * function (the pathkeys may use different data types) and the number of
+ * distinct values in each column (which affects the number of comparator
+ * calls for the following pathkeys).
+ *
+ * In case the input is partially sorted, only the remaining pathkeys are
+ * considered.
+ *
+ * Returns true if the keys/clauses have been reordered (or might have been),
+ * and a new list is returned through an argument. The list is a new copy
+ * and may be freed using list_free.
+ *
+ * Returns false if no reordering was possible.
+ */
+static bool
+get_cheapest_group_keys_order(PlannerInfo *root, double nrows,
+ List **group_pathkeys, List **group_clauses,
+ int n_preordered)
+{
+ List *new_group_pathkeys = NIL,
+ *new_group_clauses = NIL,
+ *var_group_pathkeys;
+
+ ListCell *cell;
+ PathkeyMutatorState mstate;
+ double cheapest_sort_cost = -1.0;
+
+ int nFreeKeys;
+ int nToPermute;
+
+ /* If there are less than 2 unsorted pathkeys, we're done. */
+ if (list_length(*group_pathkeys) - n_preordered < 2)
+ return false;
+
+ /*
+ * We could exhaustively cost all possible orderings of the pathkeys, but for
+ * a large number of pathkeys it might be prohibitively expensive. So we try
+ * to apply simple cheap heuristics first - we sort the pathkeys by sort cost
+ * (as if the pathkey was sorted independently) and then check only the four
+ * cheapest pathkeys. The remaining pathkeys are kept ordered by cost.
+ *
+ * XXX This is a very simple heuristics, but likely to work fine for most
+ * cases (because the number of GROUP BY clauses tends to be lower than 4).
+ * But it ignores how the number of distinct values in each pathkey affects
+ * the following steps. It might be better to use "more expensive" pathkey
+ * first if it has many distinct values, because it then limits the number
+ * of comparisons for the remaining pathkeys. But evaluating that is likely
+ * quite the expensive.
+ */
+ nFreeKeys = list_length(*group_pathkeys) - n_preordered;
+ nToPermute = 4;
+ if (nFreeKeys > nToPermute)
+ {
+ int i;
+ PathkeySortCost *costs = palloc(sizeof(PathkeySortCost) * nFreeKeys);
+
+ /* skip the pre-ordered pathkeys */
+ cell = list_nth_cell(*group_pathkeys, n_preordered);
+
+ /* estimate cost for sorting individual pathkeys */
+ for (i = 0; cell != NULL; i++, (cell = lnext(*group_pathkeys, cell)))
+ {
+ List *to_cost = list_make1(lfirst(cell));
+
+ Assert(i < nFreeKeys);
+
+ costs[i].pathkey = lfirst(cell);
+ costs[i].cost = cost_sort_estimate(root, to_cost, 0, nrows);
+
+ pfree(to_cost);
+ }
+
+ /* sort the pathkeys by sort cost in ascending order */
+ qsort(costs, nFreeKeys, sizeof(*costs), pathkey_sort_cost_comparator);
+
+ /*
+ * Rebuild the list of pathkeys - first the preordered ones, then the
+ * rest ordered by cost.
+ */
+ new_group_pathkeys = list_truncate(list_copy(*group_pathkeys), n_preordered);
+
+ for (i = 0; i < nFreeKeys; i++)
+ new_group_pathkeys = lappend(new_group_pathkeys, costs[i].pathkey);
+
+ pfree(costs);
+ }
+ else
+ {
+ /* Copy the list, so that we can free the new list by list_free. */
+ new_group_pathkeys = list_copy(*group_pathkeys);
+ nToPermute = nFreeKeys;
+ }
+
+ Assert(list_length(new_group_pathkeys) == list_length(*group_pathkeys));
+
+ /*
+ * Generate pathkey lists with permutations of the first nToPermute pathkeys.
+ *
+ * XXX We simply calculate sort cost for each individual pathkey list, but
+ * there's room for two dynamic programming optimizations here. Firstly, we
+ * may pass the current "best" cost to cost_sort_estimate so that it can
+ * "abort" if the estimated pathkeys list exceeds it. Secondly, it could pass
+ * the return information about the position when it exceeded the cost, and
+ * we could skip all permutations with the same prefix.
+ *
+ * Imagine we've already found ordering with cost C1, and we're evaluating
+ * another ordering - cost_sort_estimate() calculates cost by adding the
+ * pathkeys one by one (more or less), and the cost only grows. If at any
+ * point it exceeds C1, it can't possibly be "better" so we can discard it.
+ * But we also know that we can discard all ordering with the same prefix,
+ * because if we're estimating (a,b,c,d) and we exceed C1 at (a,b) then the
+ * same thing will happen for any ordering with this prefix.
+ */
+ PathkeyMutatorInit(&mstate, new_group_pathkeys, n_preordered, n_preordered + nToPermute);
+
+ while((var_group_pathkeys = PathkeyMutatorNext(&mstate)) != NIL)
+ {
+ Cost cost;
+
+ cost = cost_sort_estimate(root, var_group_pathkeys, n_preordered, nrows);
+
+ if (cost < cheapest_sort_cost || cheapest_sort_cost < 0)
+ {
+ list_free(new_group_pathkeys);
+ new_group_pathkeys = list_copy(var_group_pathkeys);
+ cheapest_sort_cost = cost;
+ }
+ }
+
+ /* Reorder the group clauses according to the reordered pathkeys. */
+ foreach(cell, new_group_pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(cell);
+
+ new_group_clauses = lappend(new_group_clauses,
+ get_sortgroupref_clause(pathkey->pk_eclass->ec_sortref,
+ *group_clauses));
+ }
+
+ /* Just append the rest GROUP BY clauses */
+ new_group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+
+ *group_pathkeys = new_group_pathkeys;
+ *group_clauses = new_group_clauses;
+
+ return true;
+}
+
+/*
+ * get_useful_group_keys_orderings
+ * Determine which orderings of GROUP BY keys are potentially interesting.
+ *
+ * Returns list of PathKeyInfo items, each representing an interesting ordering
+ * of GROUP BY keys. Each item stores pathkeys and clauses in matching order.
+ *
+ * The function considers (and keeps) multiple group by orderings:
+ *
+ * - the original ordering, as specified by the GROUP BY clause
+ *
+ * - GROUP BY keys reordered to minimize the sort cost
+ *
+ * - GROUP BY keys reordered to match path ordering (as much as possible), with
+ * the tail reordered to minimize the sort cost
+ *
+ * - GROUP BY keys to match target ORDER BY clause (as much as possible), with
+ * the tail reordered to minimize the sort cost
+ *
+ * There are other potentially interesting orderings (e.g. it might be best to
+ * match the first ORDER BY key, order the remaining keys differently and then
+ * rely on the incremental sort to fix this), but we ignore those for now. To
+ * make this work we'd have to pretty much generate all possible permutations.
+ */
+List *
+get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys,
+ List *group_pathkeys, List *group_clauses)
+{
+ Query *parse = root->parse;
+ List *infos = NIL;
+ PathKeyInfo *info;
+ int n_preordered = 0;
+
+ List *pathkeys = group_pathkeys;
+ List *clauses = group_clauses;
+
+ /* always return at least the original pathkeys/clauses */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+
+ /*
+ * Should we try generating alternative orderings of the group keys? If not,
+ * we produce only the order specified in the query, i.e. the optimization
+ * is effectively disabled.
+ */
+ if (!enable_group_by_reordering)
+ return infos;
+
+ /* for grouping sets we can't do any reordering */
+ if (parse->groupingSets)
+ return infos;
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost, ignoring both the
+ * target ordering (ORDER BY) and ordering of the input path.
+ */
+ if (get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered))
+ {
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * If the path is sorted in some way, try reordering the group keys to match
+ * as much of the ordering as possible - we get this sort for free (mostly).
+ *
+ * We must not do this when there are no grouping sets, because those use
+ * more complex logic to decide the ordering.
+ *
+ * XXX Isn't this somewhat redundant with presorted_keys? Actually, it's
+ * more a complement, because it allows benefiting from incremental sort
+ * as much as possible.
+ *
+ * XXX This does nothing if (n_preordered == 0). We shouldn't create the
+ * info in this case.
+ */
+ if (path_pathkeys)
+ {
+ n_preordered = group_keys_reorder_by_pathkeys(path_pathkeys,
+ &pathkeys,
+ &clauses);
+
+ /* reorder the tail to minimize sort cost */
+ get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, in
+ * which case get_cheapest_group_keys_order returns false. But we
+ * still want to keep the keys reordered to path_pathkeys.
+ */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost (this time consider
+ * the ORDER BY clause, but only if set debug_group_by_match_order_by).
+ */
+ if (root->sort_pathkeys)
+ {
+ n_preordered = group_keys_reorder_by_pathkeys(root->sort_pathkeys,
+ &pathkeys,
+ &clauses);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, in
+ * which case get_cheapest_group_keys_order returns false. But we
+ * still want to keep the keys reordered to sort_pathkeys.
+ */
+ get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered);
+
+ /* keep the group keys reordered to match ordering of input path */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ return infos;
+}
+
/*
* pathkeys_count_contained_in
* Same as pathkeys_contained_in, but also sets length of longest
@@ -1862,6 +2378,54 @@ pathkeys_useful_for_ordering(PlannerInfo *root, List *pathkeys)
return n_common_pathkeys;
}
+/*
+ * pathkeys_useful_for_grouping
+ * Count the number of pathkeys that are useful for grouping (instead of
+ * explicit sort)
+ *
+ * Group pathkeys could be reordered to benefit from the odering. The ordering
+ * may not be "complete" and may require incremental sort, but that's fine. So
+ * we simply count prefix pathkeys with a matching group key, and stop once we
+ * find the first pathkey without a match.
+ *
+ * So e.g. with pathkeys (a,b,c) and group keys (a,b,e) this determines (a,b)
+ * pathkeys are useful for grouping, and we might do incremental sort to get
+ * path ordered by (a,b,e).
+ *
+ * This logic is necessary to retain paths with ordeding not matching grouping
+ * keys directly, without the reordering.
+ *
+ * Returns the length of pathkey prefix with matching group keys.
+ */
+static int
+pathkeys_useful_for_grouping(PlannerInfo *root, List *pathkeys)
+{
+ ListCell *key;
+ int n = 0;
+
+ /* no special ordering requested for grouping */
+ if (root->group_pathkeys == NIL)
+ return 0;
+
+ /* unordered path */
+ if (pathkeys == NIL)
+ return 0;
+
+ /* walk the pathkeys and search for matching group key */
+ foreach(key, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(key);
+
+ /* no matching group key, we're done */
+ if (!list_member_ptr(root->group_pathkeys, pathkey))
+ break;
+
+ n++;
+ }
+
+ return n;
+}
+
/*
* truncate_useless_pathkeys
* Shorten the given pathkey list to just the useful pathkeys.
@@ -1876,6 +2440,9 @@ truncate_useless_pathkeys(PlannerInfo *root,
nuseful = pathkeys_useful_for_merging(root, rel, pathkeys);
nuseful2 = pathkeys_useful_for_ordering(root, pathkeys);
+ if (nuseful2 > nuseful)
+ nuseful = nuseful2;
+ nuseful2 = pathkeys_useful_for_grouping(root, pathkeys);
if (nuseful2 > nuseful)
nuseful = nuseful2;
@@ -1911,6 +2478,8 @@ has_useful_pathkeys(PlannerInfo *root, RelOptInfo *rel)
{
if (rel->joininfo != NIL || rel->has_eclass_joins)
return true; /* might be able to use pathkeys for merging */
+ if (root->group_pathkeys != NIL)
+ return true; /* might be able to use pathkeys for grouping */
if (root->query_pathkeys != NIL)
return true; /* might be able to use them for ordering */
return false; /* definitely useless */
diff --git a/src/backend/optimizer/plan/planner.c b/src/backend/optimizer/plan/planner.c
index 547fda20a23..b2569c5d0c7 100644
--- a/src/backend/optimizer/plan/planner.c
+++ b/src/backend/optimizer/plan/planner.c
@@ -6218,24 +6218,121 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
Path *path_original = path;
- bool is_sorted;
- int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ List *pathkey_orderings = NIL;
+
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
+
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
- if (path == cheapest_path || is_sorted)
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- /* Sort the cheapest-total path if it isn't already sorted */
- if (!is_sorted)
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+ /* restore the path (we replace it in the loop) */
+ path = path_original;
+
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (path == cheapest_path || is_sorted)
+ {
+ /* Sort the cheapest-total path if it isn't already sorted */
+ if (!is_sorted)
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+
+ /* Now decide what to stick atop it */
+ if (parse->groupingSets)
+ {
+ consider_groupingsets_paths(root, grouped_rel,
+ path, true, can_hash,
+ gd, agg_costs, dNumGroups);
+ }
+ else if (parse->hasAggs)
+ {
+ /*
+ * We have aggregation, possibly with plain GROUP BY. Make
+ * an AggPath.
+ */
+ add_path(grouped_rel, (Path *)
+ create_agg_path(root,
+ grouped_rel,
+ path,
+ grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_SIMPLE,
+ info->clauses,
+ havingQual,
+ agg_costs,
+ dNumGroups));
+ }
+ else if (group_clauses)
+ {
+ /*
+ * We have GROUP BY without aggregation or grouping sets.
+ * Make a GroupPath.
+ */
+ add_path(grouped_rel, (Path *)
+ create_group_path(root,
+ grouped_rel,
+ path,
+ info->clauses,
+ havingQual,
+ dNumGroups));
+ }
+ else
+ {
+ /* Other cases should have been handled above */
+ Assert(false);
+ }
+ }
+
+ /*
+ * Now we may consider incremental sort on this path, but only
+ * when the path is not already sorted and when incremental sort
+ * is enabled.
+ */
+ if (is_sorted || !enable_incremental_sort)
+ continue;
+
+ /* Restore the input path (we might have added Sort on top). */
+ path = path_original;
+
+ /* no shared prefix, no point in building incremental sort */
+ if (presorted_keys == 0)
+ continue;
+
+ /*
+ * We should have already excluded pathkeys of length 1 because
+ * then presorted_keys > 0 would imply is_sorted was true.
+ */
+ Assert(list_length(root->group_pathkeys) != 1);
+
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
/* Now decide what to stick atop it */
if (parse->groupingSets)
@@ -6247,17 +6344,17 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
else if (parse->hasAggs)
{
/*
- * We have aggregation, possibly with plain GROUP BY. Make
- * an AggPath.
+ * We have aggregation, possibly with plain GROUP BY. Make an
+ * AggPath.
*/
add_path(grouped_rel, (Path *)
create_agg_path(root,
grouped_rel,
path,
grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_SIMPLE,
- parse->groupClause,
+ info->clauses,
havingQual,
agg_costs,
dNumGroups));
@@ -6265,14 +6362,14 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
else if (parse->groupClause)
{
/*
- * We have GROUP BY without aggregation or grouping sets.
- * Make a GroupPath.
+ * We have GROUP BY without aggregation or grouping sets. Make
+ * a GroupPath.
*/
add_path(grouped_rel, (Path *)
create_group_path(root,
grouped_rel,
path,
- parse->groupClause,
+ info->clauses,
havingQual,
dNumGroups));
}
@@ -6282,79 +6379,6 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
Assert(false);
}
}
-
- /*
- * Now we may consider incremental sort on this path, but only
- * when the path is not already sorted and when incremental sort
- * is enabled.
- */
- if (is_sorted || !enable_incremental_sort)
- continue;
-
- /* Restore the input path (we might have added Sort on top). */
- path = path_original;
-
- /* no shared prefix, no point in building incremental sort */
- if (presorted_keys == 0)
- continue;
-
- /*
- * We should have already excluded pathkeys of length 1 because
- * then presorted_keys > 0 would imply is_sorted was true.
- */
- Assert(list_length(root->group_pathkeys) != 1);
-
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
-
- /* Now decide what to stick atop it */
- if (parse->groupingSets)
- {
- consider_groupingsets_paths(root, grouped_rel,
- path, true, can_hash,
- gd, agg_costs, dNumGroups);
- }
- else if (parse->hasAggs)
- {
- /*
- * We have aggregation, possibly with plain GROUP BY. Make an
- * AggPath.
- */
- add_path(grouped_rel, (Path *)
- create_agg_path(root,
- grouped_rel,
- path,
- grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_SIMPLE,
- parse->groupClause,
- havingQual,
- agg_costs,
- dNumGroups));
- }
- else if (parse->groupClause)
- {
- /*
- * We have GROUP BY without aggregation or grouping sets. Make
- * a GroupPath.
- */
- add_path(grouped_rel, (Path *)
- create_group_path(root,
- grouped_rel,
- path,
- parse->groupClause,
- havingQual,
- dNumGroups));
- }
- else
- {
- /* Other cases should have been handled above */
- Assert(false);
- }
}
/*
@@ -6365,100 +6389,124 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
{
foreach(lc, partially_grouped_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
Path *path_original = path;
- bool is_sorted;
- int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ List *pathkey_orderings = NIL;
- /*
- * Insert a Sort node, if required. But there's no point in
- * sorting anything but the cheapest path.
- */
- if (!is_sorted)
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
+
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- if (path != partially_grouped_rel->cheapest_total_path)
- continue;
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- }
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- if (parse->hasAggs)
- add_path(grouped_rel, (Path *)
- create_agg_path(root,
- grouped_rel,
- path,
- grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_FINAL_DESERIAL,
- parse->groupClause,
- havingQual,
- agg_final_costs,
- dNumGroups));
- else
- add_path(grouped_rel, (Path *)
- create_group_path(root,
- grouped_rel,
- path,
- parse->groupClause,
- havingQual,
- dNumGroups));
+ /* restore the path (we replace it in the loop) */
+ path = path_original;
- /*
- * Now we may consider incremental sort on this path, but only
- * when the path is not already sorted and when incremental
- * sort is enabled.
- */
- if (is_sorted || !enable_incremental_sort)
- continue;
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
- /* Restore the input path (we might have added Sort on top). */
- path = path_original;
+ /*
+ * Insert a Sort node, if required. But there's no point in
+ * sorting anything but the cheapest path.
+ */
+ if (!is_sorted)
+ {
+ if (path != partially_grouped_rel->cheapest_total_path)
+ continue;
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ }
- /* no shared prefix, not point in building incremental sort */
- if (presorted_keys == 0)
- continue;
+ if (parse->hasAggs)
+ add_path(grouped_rel, (Path *)
+ create_agg_path(root,
+ grouped_rel,
+ path,
+ grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_FINAL_DESERIAL,
+ info->clauses,
+ havingQual,
+ agg_final_costs,
+ dNumGroups));
+ else
+ add_path(grouped_rel, (Path *)
+ create_group_path(root,
+ grouped_rel,
+ path,
+ info->clauses,
+ havingQual,
+ dNumGroups));
- /*
- * We should have already excluded pathkeys of length 1
- * because then presorted_keys > 0 would imply is_sorted was
- * true.
- */
- Assert(list_length(root->group_pathkeys) != 1);
+ /*
+ * Now we may consider incremental sort on this path, but only
+ * when the path is not already sorted and when incremental
+ * sort is enabled.
+ */
+ if (is_sorted || !enable_incremental_sort)
+ continue;
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
+ /* Restore the input path (we might have added Sort on top). */
+ path = path_original;
- if (parse->hasAggs)
- add_path(grouped_rel, (Path *)
- create_agg_path(root,
- grouped_rel,
- path,
- grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_FINAL_DESERIAL,
- parse->groupClause,
- havingQual,
- agg_final_costs,
- dNumGroups));
- else
- add_path(grouped_rel, (Path *)
- create_group_path(root,
- grouped_rel,
- path,
- parse->groupClause,
- havingQual,
- dNumGroups));
+ /* no shared prefix, not point in building incremental sort */
+ if (presorted_keys == 0)
+ continue;
+
+ /*
+ * We should have already excluded pathkeys of length 1
+ * because then presorted_keys > 0 would imply is_sorted was
+ * true.
+ */
+ Assert(list_length(root->group_pathkeys) != 1);
+
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+
+ if (parse->hasAggs)
+ add_path(grouped_rel, (Path *)
+ create_agg_path(root,
+ grouped_rel,
+ path,
+ grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_FINAL_DESERIAL,
+ info->clauses,
+ havingQual,
+ agg_final_costs,
+ dNumGroups));
+ else
+ add_path(grouped_rel, (Path *)
+ create_group_path(root,
+ grouped_rel,
+ path,
+ info->clauses,
+ havingQual,
+ dNumGroups));
+ }
}
}
}
@@ -6661,41 +6709,71 @@ create_partial_grouping_paths(PlannerInfo *root,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
+ Path *path_save = path;
- is_sorted = pathkeys_contained_in(root->group_pathkeys,
- path->pathkeys);
- if (path == cheapest_total_path || is_sorted)
+ List *pathkey_orderings = NIL;
+
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
+
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- /* Sort the cheapest partial path, if it isn't already */
- if (!is_sorted)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- if (parse->hasAggs)
- add_path(partially_grouped_rel, (Path *)
- create_agg_path(root,
- partially_grouped_rel,
- path,
- partially_grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_INITIAL_SERIAL,
- parse->groupClause,
- NIL,
- agg_partial_costs,
- dNumPartialGroups));
- else
- add_path(partially_grouped_rel, (Path *)
- create_group_path(root,
- partially_grouped_rel,
- path,
- parse->groupClause,
- NIL,
- dNumPartialGroups));
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
+
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (path == cheapest_total_path || is_sorted)
+ {
+ /* Sort the cheapest partial path, if it isn't already */
+ if (!is_sorted)
+ {
+ path = (Path *) create_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_path(partially_grouped_rel, (Path *)
+ create_agg_path(root,
+ partially_grouped_rel,
+ path,
+ partially_grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_INITIAL_SERIAL,
+ info->clauses,
+ NIL,
+ agg_partial_costs,
+ dNumPartialGroups));
+ else
+ add_path(partially_grouped_rel, (Path *)
+ create_group_path(root,
+ partially_grouped_rel,
+ path,
+ info->clauses,
+ NIL,
+ dNumPartialGroups));
+ }
}
}
@@ -6705,6 +6783,8 @@ create_partial_grouping_paths(PlannerInfo *root,
* We can also skip the entire loop when we only have a single-item
* group_pathkeys because then we can't possibly have a presorted
* prefix of the list without having the list be fully sorted.
+ *
+ * XXX Shouldn't this also consider the group-key-reordering?
*/
if (enable_incremental_sort && list_length(root->group_pathkeys) > 1)
{
@@ -6763,24 +6843,100 @@ create_partial_grouping_paths(PlannerInfo *root,
/* Similar to above logic, but for partial paths. */
foreach(lc, input_rel->partial_pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
Path *path_original = path;
- bool is_sorted;
- int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ List *pathkey_orderings = NIL;
+
+ List *group_pathkeys = root->group_pathkeys;
+ List *group_clauses = parse->groupClause;
- if (path == cheapest_partial_path || is_sorted)
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys,
+ group_pathkeys,
+ group_clauses);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach (lc2, pathkey_orderings)
{
- /* Sort the cheapest partial path, if it isn't already */
- if (!is_sorted)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
+ bool is_sorted;
+ int presorted_keys = 0;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+ /* restore the path (we replace it in the loop) */
+ path = path_original;
+
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (path == cheapest_partial_path || is_sorted)
+ {
+
+ /* Sort the cheapest partial path, if it isn't already */
+ if (!is_sorted)
+ {
+ path = (Path *) create_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_partial_path(partially_grouped_rel, (Path *)
+ create_agg_path(root,
+ partially_grouped_rel,
+ path,
+ partially_grouped_rel->reltarget,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
+ AGGSPLIT_INITIAL_SERIAL,
+ info->clauses,
+ NIL,
+ agg_partial_costs,
+ dNumPartialPartialGroups));
+ else
+ add_partial_path(partially_grouped_rel, (Path *)
+ create_group_path(root,
+ partially_grouped_rel,
+ path,
+ info->clauses,
+ NIL,
+ dNumPartialPartialGroups));
+ }
+
+ /*
+ * Now we may consider incremental sort on this path, but only
+ * when the path is not already sorted and when incremental sort
+ * is enabled.
+ */
+ if (is_sorted || !enable_incremental_sort)
+ continue;
+
+ /* Restore the input path (we might have added Sort on top). */
+ path = path_original;
+
+ /* no shared prefix, not point in building incremental sort */
+ if (presorted_keys == 0)
+ continue;
+
+ /*
+ * We should have already excluded pathkeys of length 1 because
+ * then presorted_keys > 0 would imply is_sorted was true.
+ */
+ Assert(list_length(root->group_pathkeys) != 1);
+
+ path = (Path *) create_incremental_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
if (parse->hasAggs)
add_partial_path(partially_grouped_rel, (Path *)
@@ -6788,9 +6944,9 @@ create_partial_grouping_paths(PlannerInfo *root,
partially_grouped_rel,
path,
partially_grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
+ info->clauses ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_INITIAL_SERIAL,
- parse->groupClause,
+ info->clauses,
NIL,
agg_partial_costs,
dNumPartialPartialGroups));
@@ -6799,59 +6955,10 @@ create_partial_grouping_paths(PlannerInfo *root,
create_group_path(root,
partially_grouped_rel,
path,
- parse->groupClause,
+ info->clauses,
NIL,
dNumPartialPartialGroups));
}
-
- /*
- * Now we may consider incremental sort on this path, but only
- * when the path is not already sorted and when incremental sort
- * is enabled.
- */
- if (is_sorted || !enable_incremental_sort)
- continue;
-
- /* Restore the input path (we might have added Sort on top). */
- path = path_original;
-
- /* no shared prefix, not point in building incremental sort */
- if (presorted_keys == 0)
- continue;
-
- /*
- * We should have already excluded pathkeys of length 1 because
- * then presorted_keys > 0 would imply is_sorted was true.
- */
- Assert(list_length(root->group_pathkeys) != 1);
-
- path = (Path *) create_incremental_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
-
- if (parse->hasAggs)
- add_partial_path(partially_grouped_rel, (Path *)
- create_agg_path(root,
- partially_grouped_rel,
- path,
- partially_grouped_rel->reltarget,
- parse->groupClause ? AGG_SORTED : AGG_PLAIN,
- AGGSPLIT_INITIAL_SERIAL,
- parse->groupClause,
- NIL,
- agg_partial_costs,
- dNumPartialPartialGroups));
- else
- add_partial_path(partially_grouped_rel, (Path *)
- create_group_path(root,
- partially_grouped_rel,
- path,
- parse->groupClause,
- NIL,
- dNumPartialPartialGroups));
}
}
diff --git a/src/backend/optimizer/util/pathnode.c b/src/backend/optimizer/util/pathnode.c
index 99df76b6b71..1670e546440 100644
--- a/src/backend/optimizer/util/pathnode.c
+++ b/src/backend/optimizer/util/pathnode.c
@@ -1342,7 +1342,7 @@ create_append_path(PlannerInfo *root,
pathnode->path.pathkeys = child->pathkeys;
}
else
- cost_append(pathnode);
+ cost_append(pathnode, root);
/* If the caller provided a row estimate, override the computed value. */
if (rows >= 0)
diff --git a/src/backend/utils/adt/selfuncs.c b/src/backend/utils/adt/selfuncs.c
index 1fbb0b28c3b..fb4fb987e7f 100644
--- a/src/backend/utils/adt/selfuncs.c
+++ b/src/backend/utils/adt/selfuncs.c
@@ -3368,11 +3368,28 @@ double
estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
List **pgset, EstimationInfo *estinfo)
{
- List *varinfos = NIL;
+ return estimate_num_groups_incremental(root, groupExprs,
+ input_rows, pgset, estinfo,
+ NULL, 0);
+}
+
+/*
+ * estimate_num_groups_incremental
+ * An estimate_num_groups variant, optimized for cases that are adding the
+ * expressions incrementally (e.g. one by one).
+ */
+double
+estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
+ double input_rows,
+ List **pgset, EstimationInfo *estinfo,
+ List **cache_varinfos, int prevNExprs)
+{
+ List *varinfos = (cache_varinfos) ? *cache_varinfos : NIL;
double srf_multiplier = 1.0;
double numdistinct;
ListCell *l;
- int i;
+ int i,
+ j;
/* Zero the estinfo output parameter, if non-NULL */
if (estinfo != NULL)
@@ -3403,7 +3420,7 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
*/
numdistinct = 1.0;
- i = 0;
+ i = j = 0;
foreach(l, groupExprs)
{
Node *groupexpr = (Node *) lfirst(l);
@@ -3412,6 +3429,14 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
List *varshere;
ListCell *l2;
+ /* was done on previous call */
+ if (cache_varinfos && j++ < prevNExprs)
+ {
+ if (pgset)
+ i++; /* to keep in sync with lines below */
+ continue;
+ }
+
/* is expression in this grouping set? */
if (pgset && !list_member_int(*pgset, i++))
continue;
@@ -3481,7 +3506,11 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
if (varshere == NIL)
{
if (contain_volatile_functions(groupexpr))
+ {
+ if (cache_varinfos)
+ *cache_varinfos = varinfos;
return input_rows;
+ }
continue;
}
@@ -3498,6 +3527,9 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
}
}
+ if (cache_varinfos)
+ *cache_varinfos = varinfos;
+
/*
* If now no Vars, we must have an all-constant or all-boolean GROUP BY
* list.
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index eb3a03b9762..9e8ab1420d9 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -1182,6 +1182,16 @@ static struct config_bool ConfigureNamesBool[] =
true,
NULL, NULL, NULL
},
+ {
+ {"enable_group_by_reordering", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("enable reordering of GROUP BY key"),
+ NULL,
+ GUC_EXPLAIN
+ },
+ &enable_group_by_reordering,
+ true,
+ NULL, NULL, NULL
+ },
{
{"geqo", PGC_USERSET, QUERY_TUNING_GEQO,
gettext_noop("Enables genetic query optimization."),
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index b933fade8c6..93d221a37b1 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -382,6 +382,7 @@
#enable_seqscan = on
#enable_sort = on
#enable_tidscan = on
+#enable_group_by_reordering = on
# - Planner Cost Constants -
diff --git a/src/include/nodes/nodes.h b/src/include/nodes/nodes.h
index e8f30367a48..e74a7c72ea6 100644
--- a/src/include/nodes/nodes.h
+++ b/src/include/nodes/nodes.h
@@ -271,6 +271,7 @@ typedef enum NodeTag
T_EquivalenceClass,
T_EquivalenceMember,
T_PathKey,
+ T_PathKeyInfo,
T_PathTarget,
T_RestrictInfo,
T_IndexClause,
diff --git a/src/include/nodes/pathnodes.h b/src/include/nodes/pathnodes.h
index 365000bdcdb..6cbcb67bdf1 100644
--- a/src/include/nodes/pathnodes.h
+++ b/src/include/nodes/pathnodes.h
@@ -1070,6 +1070,16 @@ typedef struct PathKey
bool pk_nulls_first; /* do NULLs come before normal values? */
} PathKey;
+/*
+ * Combines information about pathkeys and the associated clauses.
+ */
+typedef struct PathKeyInfo
+{
+ NodeTag type;
+ List *pathkeys;
+ List *clauses;
+} PathKeyInfo;
+
/*
* VolatileFunctionStatus -- allows nodes to cache their
* contain_volatile_functions properties. VOLATILITY_UNKNOWN means not yet
diff --git a/src/include/optimizer/cost.h b/src/include/optimizer/cost.h
index bc12071af6e..dc7fc174114 100644
--- a/src/include/optimizer/cost.h
+++ b/src/include/optimizer/cost.h
@@ -114,7 +114,9 @@ extern void cost_incremental_sort(Path *path,
Cost input_startup_cost, Cost input_total_cost,
double input_tuples, int width, Cost comparison_cost, int sort_mem,
double limit_tuples);
-extern void cost_append(AppendPath *path);
+extern Cost cost_sort_estimate(PlannerInfo *root, List *pathkeys,
+ int nPresortedKeys, double tuples);
+extern void cost_append(AppendPath *path, PlannerInfo *root);
extern void cost_merge_append(Path *path, PlannerInfo *root,
List *pathkeys, int n_streams,
Cost input_startup_cost, Cost input_total_cost,
diff --git a/src/include/optimizer/paths.h b/src/include/optimizer/paths.h
index 0c3a0b90c85..2ffa24aa899 100644
--- a/src/include/optimizer/paths.h
+++ b/src/include/optimizer/paths.h
@@ -24,6 +24,7 @@ extern PGDLLIMPORT bool enable_geqo;
extern PGDLLIMPORT int geqo_threshold;
extern PGDLLIMPORT int min_parallel_table_scan_size;
extern PGDLLIMPORT int min_parallel_index_scan_size;
+extern PGDLLIMPORT bool enable_group_by_reordering;
/* Hook for plugins to get control in set_rel_pathlist() */
typedef void (*set_rel_pathlist_hook_type) (PlannerInfo *root,
@@ -203,6 +204,12 @@ typedef enum
extern PathKeysComparison compare_pathkeys(List *keys1, List *keys2);
extern bool pathkeys_contained_in(List *keys1, List *keys2);
extern bool pathkeys_count_contained_in(List *keys1, List *keys2, int *n_common);
+extern int group_keys_reorder_by_pathkeys(List *pathkeys,
+ List **group_pathkeys,
+ List **group_clauses);
+extern List *get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys,
+ List *pathkeys, List *clauses);
extern Path *get_cheapest_path_for_pathkeys(List *paths, List *pathkeys,
Relids required_outer,
CostSelector cost_criterion,
diff --git a/src/include/utils/selfuncs.h b/src/include/utils/selfuncs.h
index 8f3d73edfb2..c313a08d541 100644
--- a/src/include/utils/selfuncs.h
+++ b/src/include/utils/selfuncs.h
@@ -214,6 +214,11 @@ extern double estimate_num_groups(PlannerInfo *root, List *groupExprs,
double input_rows, List **pgset,
EstimationInfo *estinfo);
+extern double estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
+ double input_rows, List **pgset,
+ EstimationInfo *estinfo,
+ List **cache_varinfos, int prevNExprs);
+
extern void estimate_hash_bucket_stats(PlannerInfo *root,
Node *hashkey, double nbuckets,
Selectivity *mcv_freq,
diff --git a/src/test/regress/expected/aggregates.out b/src/test/regress/expected/aggregates.out
index 0a23a39aa29..601047fa3dd 100644
--- a/src/test/regress/expected/aggregates.out
+++ b/src/test/regress/expected/aggregates.out
@@ -1210,7 +1210,8 @@ explain (costs off)
select distinct min(f1), max(f1) from minmaxtest;
QUERY PLAN
---------------------------------------------------------------------------------------------
- Unique
+ HashAggregate
+ Group Key: $0, $1
InitPlan 1 (returns $0)
-> Limit
-> Merge Append
@@ -1233,10 +1234,8 @@ explain (costs off)
-> Index Only Scan using minmaxtest2i on minmaxtest2 minmaxtest_8
Index Cond: (f1 IS NOT NULL)
-> Index Only Scan Backward using minmaxtest3i on minmaxtest3 minmaxtest_9
- -> Sort
- Sort Key: ($0), ($1)
- -> Result
-(26 rows)
+ -> Result
+(25 rows)
select distinct min(f1), max(f1) from minmaxtest;
min | max
@@ -2448,6 +2447,241 @@ SELECT balk(hundred) FROM tenk1;
(1 row)
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+SELECT
+ i AS id,
+ i/2 AS p,
+ format('%60s', i%2) AS v,
+ i/4 AS c,
+ i/8 AS d,
+ (random() * (10000/8))::int as e --the same as d but no correlation with p
+ INTO btg
+FROM
+ generate_series(1, 10000) i;
+VACUUM btg;
+ANALYZE btg;
+-- GROUP BY optimization by reorder columns by frequency
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Sort
+ Sort Key: p, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Sort
+ Sort Key: p, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, c, v
+ -> Sort
+ Sort Key: p, c, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: v, p, c
+ -> Sort
+ Sort Key: v, p, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: p, d, c, v
+ -> Sort
+ Sort Key: p, d, c, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: v, p, d, c
+ -> Sort
+ Sort Key: v, p, d, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: p, v, d, c
+ -> Sort
+ Sort Key: p, v, d, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, d, e
+ -> Sort
+ Sort Key: p, d, e
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, e, d
+ -> Sort
+ Sort Key: p, e, d
+ -> Seq Scan on btg
+(5 rows)
+
+CREATE STATISTICS btg_dep ON d, e, p FROM btg;
+ANALYZE btg;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, d, e
+ -> Sort
+ Sort Key: p, d, e
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, e, d
+ -> Sort
+ Sort Key: p, e, d
+ -> Seq Scan on btg
+(5 rows)
+
+-- GROUP BY optimization by reorder columns by index scan
+CREATE INDEX ON btg(p, v);
+SET enable_seqscan=off;
+SET enable_bitmapscan=off;
+VACUUM btg;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v ORDER BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p ORDER BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, c, v
+ -> Incremental Sort
+ Sort Key: p, c, v
+ Presorted Key: p
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY p, v;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, v, c
+ -> Incremental Sort
+ Sort Key: p, v, c
+ Presorted Key: p, v
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, c, d, v
+ -> Incremental Sort
+ Sort Key: p, c, d, v
+ Presorted Key: p
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d ORDER BY p, v;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, v, c, d
+ -> Incremental Sort
+ Sort Key: p, v, c, d
+ Presorted Key: p, v
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+DROP TABLE btg;
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+RESET enable_seqscan;
+RESET enable_bitmapscan;
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/expected/incremental_sort.out b/src/test/regress/expected/incremental_sort.out
index 545e301e482..21c429226f7 100644
--- a/src/test/regress/expected/incremental_sort.out
+++ b/src/test/regress/expected/incremental_sort.out
@@ -1439,7 +1439,7 @@ set parallel_setup_cost = 0;
set parallel_tuple_cost = 0;
set max_parallel_workers_per_gather = 2;
create table t (a int, b int, c int);
-insert into t select mod(i,10),mod(i,10),i from generate_series(1,10000) s(i);
+insert into t select mod(i,10),mod(i,10),i from generate_series(1,60000) s(i);
create index on t (a);
analyze t;
set enable_incremental_sort = off;
diff --git a/src/test/regress/expected/join.out b/src/test/regress/expected/join.out
index 19caebabd01..bf1a2db2cf0 100644
--- a/src/test/regress/expected/join.out
+++ b/src/test/regress/expected/join.out
@@ -1984,8 +1984,8 @@ USING (name);
------+----+----
bb | 12 | 13
cc | 22 | 23
- dd | | 33
ee | 42 |
+ dd | | 33
(4 rows)
-- Cases with non-nullable expressions in subquery results;
@@ -2019,8 +2019,8 @@ NATURAL FULL JOIN
------+------+------+------+------
bb | 12 | 2 | 13 | 3
cc | 22 | 2 | 23 | 3
- dd | | | 33 | 3
ee | 42 | 2 | |
+ dd | | | 33 | 3
(4 rows)
SELECT * FROM
@@ -4618,18 +4618,20 @@ select d.* from d left join (select * from b group by b.id, b.c_id) s
explain (costs off)
select d.* from d left join (select distinct * from b) s
on d.a = s.id;
- QUERY PLAN
---------------------------------------
- Merge Right Join
- Merge Cond: (b.id = d.a)
- -> Unique
- -> Sort
- Sort Key: b.id, b.c_id
- -> Seq Scan on b
+ QUERY PLAN
+---------------------------------------------
+ Merge Left Join
+ Merge Cond: (d.a = s.id)
-> Sort
Sort Key: d.a
-> Seq Scan on d
-(9 rows)
+ -> Sort
+ Sort Key: s.id
+ -> Subquery Scan on s
+ -> HashAggregate
+ Group Key: b.id, b.c_id
+ -> Seq Scan on b
+(11 rows)
-- check join removal works when uniqueness of the join condition is enforced
-- by a UNION
@@ -6336,44 +6338,39 @@ select * from j1 natural join j2;
explain (verbose, costs off)
select * from j1
inner join (select distinct id from j3) j3 on j1.id = j3.id;
- QUERY PLAN
------------------------------------------
+ QUERY PLAN
+-----------------------------------
Nested Loop
Output: j1.id, j3.id
Inner Unique: true
Join Filter: (j1.id = j3.id)
- -> Unique
+ -> HashAggregate
Output: j3.id
- -> Sort
+ Group Key: j3.id
+ -> Seq Scan on public.j3
Output: j3.id
- Sort Key: j3.id
- -> Seq Scan on public.j3
- Output: j3.id
-> Seq Scan on public.j1
Output: j1.id
-(13 rows)
+(11 rows)
-- ensure group by clause allows the inner to become unique
explain (verbose, costs off)
select * from j1
inner join (select id from j3 group by id) j3 on j1.id = j3.id;
- QUERY PLAN
------------------------------------------
+ QUERY PLAN
+-----------------------------------
Nested Loop
Output: j1.id, j3.id
Inner Unique: true
Join Filter: (j1.id = j3.id)
- -> Group
+ -> HashAggregate
Output: j3.id
Group Key: j3.id
- -> Sort
+ -> Seq Scan on public.j3
Output: j3.id
- Sort Key: j3.id
- -> Seq Scan on public.j3
- Output: j3.id
-> Seq Scan on public.j1
Output: j1.id
-(14 rows)
+(11 rows)
drop table j1;
drop table j2;
diff --git a/src/test/regress/expected/partition_aggregate.out b/src/test/regress/expected/partition_aggregate.out
index dfa4b036b52..a08a3825ff6 100644
--- a/src/test/regress/expected/partition_aggregate.out
+++ b/src/test/regress/expected/partition_aggregate.out
@@ -952,32 +952,30 @@ SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HA
--------------------------------------------------------------------------------------
Sort
Sort Key: pagg_tab_ml.a, (sum(pagg_tab_ml.b)), (array_agg(DISTINCT pagg_tab_ml.c))
- -> Gather
- Workers Planned: 2
- -> Parallel Append
- -> GroupAggregate
- Group Key: pagg_tab_ml.a
- Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml.a
- -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
- -> GroupAggregate
- Group Key: pagg_tab_ml_5.a
- Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_5.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
- -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
- -> GroupAggregate
- Group Key: pagg_tab_ml_2.a
- Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_2.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
- -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
-(27 rows)
+ -> Append
+ -> GroupAggregate
+ Group Key: pagg_tab_ml.a
+ Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml.a
+ -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_2.a
+ Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_2.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
+ -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_5.a
+ Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_5.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
+ -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
+(25 rows)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3 ORDER BY 1, 2, 3;
a | sum | array_agg | count
@@ -996,34 +994,32 @@ SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HA
-- Without ORDER BY clause, to test Gather at top-most path
EXPLAIN (COSTS OFF)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3;
- QUERY PLAN
----------------------------------------------------------------------------
- Gather
- Workers Planned: 2
- -> Parallel Append
- -> GroupAggregate
- Group Key: pagg_tab_ml.a
- Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml.a
- -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
- -> GroupAggregate
- Group Key: pagg_tab_ml_5.a
- Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_5.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
- -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
- -> GroupAggregate
- Group Key: pagg_tab_ml_2.a
- Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_2.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
- -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
-(25 rows)
+ QUERY PLAN
+---------------------------------------------------------------------
+ Append
+ -> GroupAggregate
+ Group Key: pagg_tab_ml.a
+ Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml.a
+ -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_2.a
+ Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_2.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
+ -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_5.a
+ Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_5.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
+ -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
+(23 rows)
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
@@ -1379,28 +1375,26 @@ SELECT x, sum(y), avg(y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) <
-- When GROUP BY clause does not match; partial aggregation is performed for each partition.
EXPLAIN (COSTS OFF)
SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) < 12 ORDER BY 1, 2, 3;
- QUERY PLAN
--------------------------------------------------------------------------------------------
+ QUERY PLAN
+-------------------------------------------------------------------------------------
Sort
Sort Key: pagg_tab_para.y, (sum(pagg_tab_para.x)), (avg(pagg_tab_para.x))
- -> Finalize GroupAggregate
+ -> Finalize HashAggregate
Group Key: pagg_tab_para.y
Filter: (avg(pagg_tab_para.x) < '12'::numeric)
- -> Gather Merge
+ -> Gather
Workers Planned: 2
- -> Sort
- Sort Key: pagg_tab_para.y
- -> Parallel Append
- -> Partial HashAggregate
- Group Key: pagg_tab_para.y
- -> Parallel Seq Scan on pagg_tab_para_p1 pagg_tab_para
- -> Partial HashAggregate
- Group Key: pagg_tab_para_1.y
- -> Parallel Seq Scan on pagg_tab_para_p2 pagg_tab_para_1
- -> Partial HashAggregate
- Group Key: pagg_tab_para_2.y
- -> Parallel Seq Scan on pagg_tab_para_p3 pagg_tab_para_2
-(19 rows)
+ -> Parallel Append
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para.y
+ -> Parallel Seq Scan on pagg_tab_para_p1 pagg_tab_para
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para_1.y
+ -> Parallel Seq Scan on pagg_tab_para_p2 pagg_tab_para_1
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para_2.y
+ -> Parallel Seq Scan on pagg_tab_para_p3 pagg_tab_para_2
+(17 rows)
SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) < 12 ORDER BY 1, 2, 3;
y | sum | avg | count
diff --git a/src/test/regress/expected/partition_join.out b/src/test/regress/expected/partition_join.out
index bb5b7c47a45..03926a84138 100644
--- a/src/test/regress/expected/partition_join.out
+++ b/src/test/regress/expected/partition_join.out
@@ -466,52 +466,41 @@ EXPLAIN (COSTS OFF)
SELECT a, b FROM prt1 FULL JOIN prt2 p2(b,a,c) USING(a,b)
WHERE a BETWEEN 490 AND 510
GROUP BY 1, 2 ORDER BY 1, 2;
- QUERY PLAN
------------------------------------------------------------------------------------------------------------------
+ QUERY PLAN
+-----------------------------------------------------------------------------------------------------------
Group
Group Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Merge Append
+ -> Sort
Sort Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Group
- Group Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Sort
- Sort Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Merge Full Join
- Merge Cond: ((prt1.a = p2.a) AND (prt1.b = p2.b))
- Filter: ((COALESCE(prt1.a, p2.a) >= 490) AND (COALESCE(prt1.a, p2.a) <= 510))
- -> Sort
- Sort Key: prt1.a, prt1.b
- -> Seq Scan on prt1_p1 prt1
- -> Sort
- Sort Key: p2.a, p2.b
- -> Seq Scan on prt2_p1 p2
- -> Group
- Group Key: (COALESCE(prt1_1.a, p2_1.a)), (COALESCE(prt1_1.b, p2_1.b))
- -> Sort
- Sort Key: (COALESCE(prt1_1.a, p2_1.a)), (COALESCE(prt1_1.b, p2_1.b))
- -> Merge Full Join
- Merge Cond: ((prt1_1.a = p2_1.a) AND (prt1_1.b = p2_1.b))
- Filter: ((COALESCE(prt1_1.a, p2_1.a) >= 490) AND (COALESCE(prt1_1.a, p2_1.a) <= 510))
- -> Sort
- Sort Key: prt1_1.a, prt1_1.b
- -> Seq Scan on prt1_p2 prt1_1
- -> Sort
- Sort Key: p2_1.a, p2_1.b
- -> Seq Scan on prt2_p2 p2_1
- -> Group
- Group Key: (COALESCE(prt1_2.a, p2_2.a)), (COALESCE(prt1_2.b, p2_2.b))
- -> Sort
- Sort Key: (COALESCE(prt1_2.a, p2_2.a)), (COALESCE(prt1_2.b, p2_2.b))
- -> Merge Full Join
- Merge Cond: ((prt1_2.a = p2_2.a) AND (prt1_2.b = p2_2.b))
- Filter: ((COALESCE(prt1_2.a, p2_2.a) >= 490) AND (COALESCE(prt1_2.a, p2_2.a) <= 510))
- -> Sort
- Sort Key: prt1_2.a, prt1_2.b
- -> Seq Scan on prt1_p3 prt1_2
- -> Sort
- Sort Key: p2_2.a, p2_2.b
- -> Seq Scan on prt2_p3 p2_2
-(43 rows)
+ -> Append
+ -> Merge Full Join
+ Merge Cond: ((prt1_1.a = p2_1.a) AND (prt1_1.b = p2_1.b))
+ Filter: ((COALESCE(prt1_1.a, p2_1.a) >= 490) AND (COALESCE(prt1_1.a, p2_1.a) <= 510))
+ -> Sort
+ Sort Key: prt1_1.a, prt1_1.b
+ -> Seq Scan on prt1_p1 prt1_1
+ -> Sort
+ Sort Key: p2_1.a, p2_1.b
+ -> Seq Scan on prt2_p1 p2_1
+ -> Merge Full Join
+ Merge Cond: ((prt1_2.a = p2_2.a) AND (prt1_2.b = p2_2.b))
+ Filter: ((COALESCE(prt1_2.a, p2_2.a) >= 490) AND (COALESCE(prt1_2.a, p2_2.a) <= 510))
+ -> Sort
+ Sort Key: prt1_2.a, prt1_2.b
+ -> Seq Scan on prt1_p2 prt1_2
+ -> Sort
+ Sort Key: p2_2.a, p2_2.b
+ -> Seq Scan on prt2_p2 p2_2
+ -> Merge Full Join
+ Merge Cond: ((prt1_3.b = p2_3.b) AND (prt1_3.a = p2_3.a))
+ Filter: ((COALESCE(prt1_3.a, p2_3.a) >= 490) AND (COALESCE(prt1_3.a, p2_3.a) <= 510))
+ -> Sort
+ Sort Key: prt1_3.b, prt1_3.a
+ -> Seq Scan on prt1_p3 prt1_3
+ -> Sort
+ Sort Key: p2_3.b, p2_3.a
+ -> Seq Scan on prt2_p3 p2_3
+(32 rows)
SELECT a, b FROM prt1 FULL JOIN prt2 p2(b,a,c) USING(a,b)
WHERE a BETWEEN 490 AND 510
diff --git a/src/test/regress/expected/sysviews.out b/src/test/regress/expected/sysviews.out
index 579b861d84f..4e775af1758 100644
--- a/src/test/regress/expected/sysviews.out
+++ b/src/test/regress/expected/sysviews.out
@@ -114,6 +114,7 @@ select name, setting from pg_settings where name like 'enable%';
enable_async_append | on
enable_bitmapscan | on
enable_gathermerge | on
+ enable_group_by_reordering | on
enable_hashagg | on
enable_hashjoin | on
enable_incremental_sort | on
@@ -131,7 +132,7 @@ select name, setting from pg_settings where name like 'enable%';
enable_seqscan | on
enable_sort | on
enable_tidscan | on
-(20 rows)
+(21 rows)
-- Test that the pg_timezone_names and pg_timezone_abbrevs views are
-- more-or-less working. We can't test their contents in any great detail
diff --git a/src/test/regress/expected/union.out b/src/test/regress/expected/union.out
index dece7310cfe..7ac4a9380e2 100644
--- a/src/test/regress/expected/union.out
+++ b/src/test/regress/expected/union.out
@@ -1303,24 +1303,22 @@ select distinct q1 from
union all
select distinct * from int8_tbl i82) ss
where q2 = q2;
- QUERY PLAN
-----------------------------------------------------------
- Unique
- -> Merge Append
- Sort Key: "*SELECT* 1".q1
+ QUERY PLAN
+----------------------------------------------------
+ HashAggregate
+ Group Key: "*SELECT* 1".q1
+ -> Append
-> Subquery Scan on "*SELECT* 1"
- -> Unique
- -> Sort
- Sort Key: i81.q1, i81.q2
- -> Seq Scan on int8_tbl i81
- Filter: (q2 IS NOT NULL)
+ -> HashAggregate
+ Group Key: i81.q1, i81.q2
+ -> Seq Scan on int8_tbl i81
+ Filter: (q2 IS NOT NULL)
-> Subquery Scan on "*SELECT* 2"
- -> Unique
- -> Sort
- Sort Key: i82.q1, i82.q2
- -> Seq Scan on int8_tbl i82
- Filter: (q2 IS NOT NULL)
-(15 rows)
+ -> HashAggregate
+ Group Key: i82.q1, i82.q2
+ -> Seq Scan on int8_tbl i82
+ Filter: (q2 IS NOT NULL)
+(13 rows)
select distinct q1 from
(select distinct * from int8_tbl i81
@@ -1339,24 +1337,22 @@ select distinct q1 from
union all
select distinct * from int8_tbl i82) ss
where -q1 = q2;
- QUERY PLAN
---------------------------------------------------------
- Unique
- -> Merge Append
- Sort Key: "*SELECT* 1".q1
+ QUERY PLAN
+--------------------------------------------------
+ HashAggregate
+ Group Key: "*SELECT* 1".q1
+ -> Append
-> Subquery Scan on "*SELECT* 1"
- -> Unique
- -> Sort
- Sort Key: i81.q1, i81.q2
- -> Seq Scan on int8_tbl i81
- Filter: ((- q1) = q2)
+ -> HashAggregate
+ Group Key: i81.q1, i81.q2
+ -> Seq Scan on int8_tbl i81
+ Filter: ((- q1) = q2)
-> Subquery Scan on "*SELECT* 2"
- -> Unique
- -> Sort
- Sort Key: i82.q1, i82.q2
- -> Seq Scan on int8_tbl i82
- Filter: ((- q1) = q2)
-(15 rows)
+ -> HashAggregate
+ Group Key: i82.q1, i82.q2
+ -> Seq Scan on int8_tbl i82
+ Filter: ((- q1) = q2)
+(13 rows)
select distinct q1 from
(select distinct * from int8_tbl i81
diff --git a/src/test/regress/sql/aggregates.sql b/src/test/regress/sql/aggregates.sql
index 2f5d0e00f3d..c6e0d7ba2b7 100644
--- a/src/test/regress/sql/aggregates.sql
+++ b/src/test/regress/sql/aggregates.sql
@@ -1025,6 +1025,105 @@ SELECT balk(hundred) FROM tenk1;
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+
+SELECT
+ i AS id,
+ i/2 AS p,
+ format('%60s', i%2) AS v,
+ i/4 AS c,
+ i/8 AS d,
+ (random() * (10000/8))::int as e --the same as d but no correlation with p
+ INTO btg
+FROM
+ generate_series(1, 10000) i;
+
+VACUUM btg;
+ANALYZE btg;
+
+-- GROUP BY optimization by reorder columns by frequency
+
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+
+CREATE STATISTICS btg_dep ON d, e, p FROM btg;
+ANALYZE btg;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+
+
+-- GROUP BY optimization by reorder columns by index scan
+
+CREATE INDEX ON btg(p, v);
+SET enable_seqscan=off;
+SET enable_bitmapscan=off;
+VACUUM btg;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d ORDER BY p, v;
+
+DROP TABLE btg;
+
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+RESET enable_seqscan;
+RESET enable_bitmapscan;
+
+
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/sql/incremental_sort.sql b/src/test/regress/sql/incremental_sort.sql
index d8768a6b54d..1de163e0395 100644
--- a/src/test/regress/sql/incremental_sort.sql
+++ b/src/test/regress/sql/incremental_sort.sql
@@ -213,7 +213,7 @@ set parallel_tuple_cost = 0;
set max_parallel_workers_per_gather = 2;
create table t (a int, b int, c int);
-insert into t select mod(i,10),mod(i,10),i from generate_series(1,10000) s(i);
+insert into t select mod(i,10),mod(i,10),i from generate_series(1,60000) s(i);
create index on t (a);
analyze t;
--
2.34.1
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-03-30 23:19 Tomas Vondra <[email protected]>
parent: Tomas Vondra <[email protected]>
0 siblings, 2 replies; 38+ messages in thread
From: Tomas Vondra @ 2022-03-30 23:19 UTC (permalink / raw)
To: Zhihong Yu <[email protected]>; +Cc: Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
Pushed, after going through the patch once more, running check-world
under valgrind, and updating the commit message.
Thanks to everyone who reviewed/tested this patch!
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-07-13 03:37 David Rowley <[email protected]>
parent: Tomas Vondra <[email protected]>
1 sibling, 1 reply; 38+ messages in thread
From: David Rowley @ 2022-07-13 03:37 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; +Cc: Zhihong Yu <[email protected]>; Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On Thu, 31 Mar 2022 at 12:19, Tomas Vondra
<[email protected]> wrote:
> Pushed, after going through the patch once more, running check-world
> under valgrind, and updating the commit message.
I'm just in this general area of the code again today and wondered
about the header comment for the preprocess_groupclause() function.
It says:
* In principle it might be interesting to consider other orderings of the
* GROUP BY elements, which could match the sort ordering of other
* possible plans (eg an indexscan) and thereby reduce cost. We don't
* bother with that, though. Hashed grouping will frequently win anyway.
I'd say this commit makes that paragraph mostly obsolete. It's only
true now in the sense that we don't try orders that suit some index
that would provide pre-sorted results for a GroupAggregate path. The
comment leads me to believe that we don't do anything at all to find a
better order, and that's not true now.
David
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-07-15 05:18 David Rowley <[email protected]>
parent: Tomas Vondra <[email protected]>
1 sibling, 1 reply; 38+ messages in thread
From: David Rowley @ 2022-07-15 05:18 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; +Cc: Zhihong Yu <[email protected]>; Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On Thu, 31 Mar 2022 at 12:19, Tomas Vondra
<[email protected]> wrote:
> Pushed, after going through the patch once more, running check-world
> under valgrind, and updating the commit message.
I'm still working in this area and I noticed that db0d67db2 updated
some regression tests in partition_aggregate.out without any care as
to what the test was testing.
The comment above the test reads:
-- Without ORDER BY clause, to test Gather at top-most path
and you've changed the expected plan from being a parallel plan with a
Gather to being a serial plan. So it looks like the test might have
become useless.
I see that the original plan appears to come back with some
adjustments to parallel_setup_cost and parallel_tuple_cost. It seems a
bit strange to me that the changes with this patch would cause a
change of plan for this. There is only 1 GROUP BY column in the query
in question. There's no rearrangement to do with a single column GROUP
BY.
David
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-07-15 19:46 Tomas Vondra <[email protected]>
parent: David Rowley <[email protected]>
0 siblings, 1 reply; 38+ messages in thread
From: Tomas Vondra @ 2022-07-15 19:46 UTC (permalink / raw)
To: David Rowley <[email protected]>; +Cc: Zhihong Yu <[email protected]>; Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On 7/15/22 07:18, David Rowley wrote:
> On Thu, 31 Mar 2022 at 12:19, Tomas Vondra
> <[email protected]> wrote:
>> Pushed, after going through the patch once more, running check-world
>> under valgrind, and updating the commit message.
>
> I'm still working in this area and I noticed that db0d67db2 updated
> some regression tests in partition_aggregate.out without any care as
> to what the test was testing.
>
> The comment above the test reads:
>
> -- Without ORDER BY clause, to test Gather at top-most path
>
> and you've changed the expected plan from being a parallel plan with a
> Gather to being a serial plan. So it looks like the test might have
> become useless.
>
I agree this is a mistake in db0d67db2 that makes the test useless.
> I see that the original plan appears to come back with some
> adjustments to parallel_setup_cost and parallel_tuple_cost. It seems a
> bit strange to me that the changes with this patch would cause a
> change of plan for this. There is only 1 GROUP BY column in the query
> in question. There's no rearrangement to do with a single column GROUP
> BY.
>
It might seem a bit strange, but the patch tweaked the costing a bit, so
it's not entirely unexpected. I'd bet the plan cost changed just a teeny
bit, but enough to change the cheapest plan. The costing changed for all
group counts, including a single group.
regards
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-08-02 10:21 Michael Paquier <[email protected]>
parent: Tomas Vondra <[email protected]>
0 siblings, 1 reply; 38+ messages in thread
From: Michael Paquier @ 2022-08-02 10:21 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; +Cc: David Rowley <[email protected]>; Zhihong Yu <[email protected]>; Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On Fri, Jul 15, 2022 at 09:46:51PM +0200, Tomas Vondra wrote:
> I agree this is a mistake in db0d67db2 that makes the test useless.
Tomas, please note that this is an open item assigned to you. Are you
planning to do something with these regression tests by beta3?
--
Michael
Attachments:
[application/pgp-signature] signature.asc (833B, ../../[email protected]/2-signature.asc)
download
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-08-02 11:14 David Rowley <[email protected]>
parent: Michael Paquier <[email protected]>
0 siblings, 1 reply; 38+ messages in thread
From: David Rowley @ 2022-08-02 11:14 UTC (permalink / raw)
To: Michael Paquier <[email protected]>; +Cc: Tomas Vondra <[email protected]>; Zhihong Yu <[email protected]>; Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On Tue, 2 Aug 2022 at 22:21, Michael Paquier <[email protected]> wrote:
>
> On Fri, Jul 15, 2022 at 09:46:51PM +0200, Tomas Vondra wrote:
> > I agree this is a mistake in db0d67db2 that makes the test useless.
>
> Tomas, please note that this is an open item assigned to you. Are you
> planning to do something with these regression tests by beta3?
There's still something to do there for PG15, but 1349d2790 (just gone
in to master) made those two tests run as a parallel query again.
David
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-08-17 14:46 Tomas Vondra <[email protected]>
parent: David Rowley <[email protected]>
0 siblings, 1 reply; 38+ messages in thread
From: Tomas Vondra @ 2022-08-17 14:46 UTC (permalink / raw)
To: David Rowley <[email protected]>; Michael Paquier <[email protected]>; +Cc: Zhihong Yu <[email protected]>; Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On 8/2/22 13:14, David Rowley wrote:
> On Tue, 2 Aug 2022 at 22:21, Michael Paquier <[email protected]> wrote:
>>
>> On Fri, Jul 15, 2022 at 09:46:51PM +0200, Tomas Vondra wrote:
>>> I agree this is a mistake in db0d67db2 that makes the test useless.
>>
>> Tomas, please note that this is an open item assigned to you. Are you
>> planning to do something with these regression tests by beta3?
>
> There's still something to do there for PG15, but 1349d2790 (just gone
> in to master) made those two tests run as a parallel query again.
>
Hi,
I've been looking at this, and as I speculated before it's due to the
sort costing changing a little bit. On PG14, the costing of the plans
looks like this:
Gather (cost=1869.39..2823.15 rows=8 width=52)
and with disabled parallelism, it's like this:
Append (cost=998.04..3000.64 rows=10 width=52)
so there's a (fairly small) diffrence in favor of the parallel plan. But
on PG15 it's the other way around - the selected non-parallel one is
costed like this:
Append (cost=779.41..2490.61 rows=10 width=52)
and by setting parallel_setup_cost=0 you get this:
Gather (cost=700.34..1531.76 rows=8 width=52)
So with the setup cost included it's ~2531, and it loses to the simple
plan. This happens because the patch changed sort costing - the same
sort on PG14 and PG15 looks like this:
PG14: -> Sort (cost=998.04..1028.04 rows=12000 width=13)
PG15: -> Sort (cost=779.41..809.41 rows=12000 width=13)
As mentioned, the commit tweaked sort costing - before it was pretty
much just
comparison_cost * tuples * LOG2(tuples)
but the patch needs to cost different pathkey orderings, and consider
that maybe we don't need to compare all the keys (which the original
costing kind assumes). That's the whole point of this optimization.
The costing (compute_cpu_sort_cost) considers a couple other things,
like cost of the comparator function for the data type, width of the
values, groupings determined by preceding keys, and so on.
It might seem strange that a query with a single pathkey changes, but
that single pathkey is costed the same way, of course. In principle we
might have a special costing for this case, but I'd bet that would
result in pretty surprising inconsistencies when adding a sort key
(going from 1 to 2 keys).
So I don't think the current costing is wrong, but it certainly is more
complex. But the test does not test what it intended - I have two ideas
how to make it work:
1) increase the number of rows in the table
2) increase cpu_operator_cost (for that one test?)
3) tweak the costing somehow, to increase the cost a bit
Both (1) and (2) work - I've tried doubling the number of rows or
setting the operator cost to 0.005, and that does the trick, but maybe a
smaller change would be enough.
I don't like (3) very much - changing the costing just to get the same
test behavior as on older release does not seem very principled. Yes,
maybe it should be tweaked, but not because of a regression test.
regards
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
Attachments:
[text/x-log] plans-14.log (3.9K, ../../[email protected]/2-plans-14.log)
download | inline:
-- Without ORDER BY clause, to test Gather at top-most path
EXPLAIN (COSTS ON)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3;
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------
Gather (cost=1869.39..2823.15 rows=8 width=52)
Workers Planned: 2
-> Parallel Append (cost=869.39..1822.35 rows=4 width=52)
-> GroupAggregate (cost=998.04..1178.25 rows=4 width=52)
Group Key: pagg_tab_ml.a
Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
-> Sort (cost=998.04..1028.04 rows=12000 width=13)
Sort Key: pagg_tab_ml.a
-> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml (cost=0.00..185.00 rows=12000 width=13)
-> GroupAggregate (cost=869.39..1019.56 rows=3 width=52)
Group Key: pagg_tab_ml_5.a
Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
-> Sort (cost=869.39..894.39 rows=10000 width=13)
Sort Key: pagg_tab_ml_5.a
-> Append (cost=0.00..205.00 rows=10000 width=13)
-> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5 (cost=0.00..108.00 rows=7000 width=13)
-> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6 (cost=0.00..47.00 rows=3000 width=13)
-> GroupAggregate (cost=682.63..802.77 rows=3 width=52)
Group Key: pagg_tab_ml_2.a
Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
-> Sort (cost=682.63..702.63 rows=8000 width=13)
Sort Key: pagg_tab_ml_2.a
-> Append (cost=0.00..164.00 rows=8000 width=13)
-> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2 (cost=0.00..93.00 rows=6000 width=13)
-> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3 (cost=0.00..31.00 rows=2000 width=13)
(25 rows)
SET max_parallel_workers_per_gather = 0;
EXPLAIN (COSTS ON)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3;
QUERY PLAN
-------------------------------------------------------------------------------------------------------------
Append (cost=998.04..3000.64 rows=10 width=52)
-> GroupAggregate (cost=998.04..1178.25 rows=4 width=52)
Group Key: pagg_tab_ml.a
Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
-> Sort (cost=998.04..1028.04 rows=12000 width=13)
Sort Key: pagg_tab_ml.a
-> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml (cost=0.00..185.00 rows=12000 width=13)
-> GroupAggregate (cost=682.63..802.77 rows=3 width=52)
Group Key: pagg_tab_ml_2.a
Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
-> Sort (cost=682.63..702.63 rows=8000 width=13)
Sort Key: pagg_tab_ml_2.a
-> Append (cost=0.00..164.00 rows=8000 width=13)
-> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2 (cost=0.00..93.00 rows=6000 width=13)
-> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3 (cost=0.00..31.00 rows=2000 width=13)
-> GroupAggregate (cost=869.39..1019.56 rows=3 width=52)
Group Key: pagg_tab_ml_5.a
Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
-> Sort (cost=869.39..894.39 rows=10000 width=13)
Sort Key: pagg_tab_ml_5.a
-> Append (cost=0.00..205.00 rows=10000 width=13)
-> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5 (cost=0.00..108.00 rows=7000 width=13)
-> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6 (cost=0.00..47.00 rows=3000 width=13)
(23 rows)
RESET max_parallel_workers_per_gather;
[text/x-log] plans-15.log (3.9K, ../../[email protected]/3-plans-15.log)
download | inline:
-- Without ORDER BY clause, to test Gather at top-most path
EXPLAIN (COSTS ON)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3;
QUERY PLAN
-------------------------------------------------------------------------------------------------------------
Append (cost=779.41..2490.61 rows=10 width=52)
-> GroupAggregate (cost=779.41..959.62 rows=4 width=52)
Group Key: pagg_tab_ml.a
Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
-> Sort (cost=779.41..809.41 rows=12000 width=13)
Sort Key: pagg_tab_ml.a
-> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml (cost=0.00..185.00 rows=12000 width=13)
-> GroupAggregate (cost=560.28..680.42 rows=3 width=52)
Group Key: pagg_tab_ml_2.a
Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
-> Sort (cost=560.28..580.28 rows=8000 width=13)
Sort Key: pagg_tab_ml_2.a
-> Append (cost=0.00..164.00 rows=8000 width=13)
-> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2 (cost=0.00..93.00 rows=6000 width=13)
-> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3 (cost=0.00..31.00 rows=2000 width=13)
-> GroupAggregate (cost=700.34..850.52 rows=3 width=52)
Group Key: pagg_tab_ml_5.a
Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
-> Sort (cost=700.34..725.34 rows=10000 width=13)
Sort Key: pagg_tab_ml_5.a
-> Append (cost=0.00..205.00 rows=10000 width=13)
-> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5 (cost=0.00..108.00 rows=7000 width=13)
-> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6 (cost=0.00..47.00 rows=3000 width=13)
(23 rows)
SET parallel_setup_cost = 0;
EXPLAIN (COSTS ON)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3;
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------
Gather (cost=700.34..1531.76 rows=8 width=52)
Workers Planned: 2
-> Parallel Append (cost=700.34..1530.96 rows=4 width=52)
-> GroupAggregate (cost=779.41..959.62 rows=4 width=52)
Group Key: pagg_tab_ml.a
Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
-> Sort (cost=779.41..809.41 rows=12000 width=13)
Sort Key: pagg_tab_ml.a
-> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml (cost=0.00..185.00 rows=12000 width=13)
-> GroupAggregate (cost=700.34..850.52 rows=3 width=52)
Group Key: pagg_tab_ml_5.a
Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
-> Sort (cost=700.34..725.34 rows=10000 width=13)
Sort Key: pagg_tab_ml_5.a
-> Append (cost=0.00..205.00 rows=10000 width=13)
-> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5 (cost=0.00..108.00 rows=7000 width=13)
-> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6 (cost=0.00..47.00 rows=3000 width=13)
-> GroupAggregate (cost=560.28..680.42 rows=3 width=52)
Group Key: pagg_tab_ml_2.a
Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
-> Sort (cost=560.28..580.28 rows=8000 width=13)
Sort Key: pagg_tab_ml_2.a
-> Append (cost=0.00..164.00 rows=8000 width=13)
-> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2 (cost=0.00..93.00 rows=6000 width=13)
-> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3 (cost=0.00..31.00 rows=2000 width=13)
(25 rows)
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-08-18 01:32 David Rowley <[email protected]>
parent: Tomas Vondra <[email protected]>
0 siblings, 1 reply; 38+ messages in thread
From: David Rowley @ 2022-08-18 01:32 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; +Cc: Michael Paquier <[email protected]>; Zhihong Yu <[email protected]>; Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On Thu, 18 Aug 2022 at 02:46, Tomas Vondra
<[email protected]> wrote:
> So I don't think the current costing is wrong, but it certainly is more
> complex. But the test does not test what it intended - I have two ideas
> how to make it work:
>
> 1) increase the number of rows in the table
>
> 2) increase cpu_operator_cost (for that one test?)
>
> 3) tweak the costing somehow, to increase the cost a bit
Why not, 4) SET parallel_setup_cost = 0; there are plenty of other
places we do just that so we get a parallel plan without having to
generate enough cost to drown out the parallel worker startup cost.
Here are a couple of patches to demo the idea.
David
diff --git a/src/test/regress/expected/partition_aggregate.out b/src/test/regress/expected/partition_aggregate.out
index 4b41ccf1aa..db36e3a150 100644
--- a/src/test/regress/expected/partition_aggregate.out
+++ b/src/test/regress/expected/partition_aggregate.out
@@ -942,6 +942,7 @@ INSERT INTO pagg_tab_ml SELECT i % 30, i % 10, to_char(i % 4, 'FM0000') FROM gen
ANALYZE pagg_tab_ml;
-- For Parallel Append
SET max_parallel_workers_per_gather TO 2;
+SET parallel_setup_cost = 0;
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
-- PARTITION KEY, but still we do not see a partial aggregation as array_agg()
@@ -1025,6 +1026,7 @@ SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HA
-> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
(25 rows)
+RESET parallel_setup_cost;
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
-- PARTITION KEY, thus we will have a partial aggregation for them.
diff --git a/src/test/regress/sql/partition_aggregate.sql b/src/test/regress/sql/partition_aggregate.sql
index c17294b15b..ab070fee24 100644
--- a/src/test/regress/sql/partition_aggregate.sql
+++ b/src/test/regress/sql/partition_aggregate.sql
@@ -222,6 +222,7 @@ ANALYZE pagg_tab_ml;
-- For Parallel Append
SET max_parallel_workers_per_gather TO 2;
+SET parallel_setup_cost = 0;
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
@@ -235,6 +236,8 @@ SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HA
EXPLAIN (COSTS OFF)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3;
+RESET parallel_setup_cost;
+
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
-- PARTITION KEY, thus we will have a partial aggregation for them.
diff --git a/src/test/regress/expected/partition_aggregate.out b/src/test/regress/expected/partition_aggregate.out
index a08a3825ff..0dc6d63347 100644
--- a/src/test/regress/expected/partition_aggregate.out
+++ b/src/test/regress/expected/partition_aggregate.out
@@ -942,40 +942,43 @@ INSERT INTO pagg_tab_ml SELECT i % 30, i % 10, to_char(i % 4, 'FM0000') FROM gen
ANALYZE pagg_tab_ml;
-- For Parallel Append
SET max_parallel_workers_per_gather TO 2;
+SET parallel_setup_cost = 0;
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
-- PARTITION KEY, but still we do not see a partial aggregation as array_agg()
-- is not partial agg safe.
EXPLAIN (COSTS OFF)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3 ORDER BY 1, 2, 3;
- QUERY PLAN
---------------------------------------------------------------------------------------
- Sort
- Sort Key: pagg_tab_ml.a, (sum(pagg_tab_ml.b)), (array_agg(DISTINCT pagg_tab_ml.c))
- -> Append
- -> GroupAggregate
- Group Key: pagg_tab_ml.a
- Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml.a
- -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
- -> GroupAggregate
- Group Key: pagg_tab_ml_2.a
- Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_2.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
- -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
- -> GroupAggregate
- Group Key: pagg_tab_ml_5.a
- Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_5.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
- -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
-(25 rows)
+ QUERY PLAN
+--------------------------------------------------------------------------------------------
+ Gather Merge
+ Workers Planned: 2
+ -> Sort
+ Sort Key: pagg_tab_ml.a, (sum(pagg_tab_ml.b)), (array_agg(DISTINCT pagg_tab_ml.c))
+ -> Parallel Append
+ -> GroupAggregate
+ Group Key: pagg_tab_ml.a
+ Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml.a
+ -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_5.a
+ Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_5.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
+ -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_2.a
+ Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_2.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
+ -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
+(27 rows)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3 ORDER BY 1, 2, 3;
a | sum | array_agg | count
@@ -994,33 +997,36 @@ SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HA
-- Without ORDER BY clause, to test Gather at top-most path
EXPLAIN (COSTS OFF)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3;
- QUERY PLAN
----------------------------------------------------------------------
- Append
- -> GroupAggregate
- Group Key: pagg_tab_ml.a
- Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml.a
- -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
- -> GroupAggregate
- Group Key: pagg_tab_ml_2.a
- Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_2.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
- -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
- -> GroupAggregate
- Group Key: pagg_tab_ml_5.a
- Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_5.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
- -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
-(23 rows)
+ QUERY PLAN
+---------------------------------------------------------------------------
+ Gather
+ Workers Planned: 2
+ -> Parallel Append
+ -> GroupAggregate
+ Group Key: pagg_tab_ml.a
+ Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml.a
+ -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_5.a
+ Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_5.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
+ -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_2.a
+ Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_2.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
+ -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
+(25 rows)
+RESET parallel_setup_cost;
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
-- PARTITION KEY, thus we will have a partial aggregation for them.
diff --git a/src/test/regress/sql/partition_aggregate.sql b/src/test/regress/sql/partition_aggregate.sql
index c17294b15b..ab070fee24 100644
--- a/src/test/regress/sql/partition_aggregate.sql
+++ b/src/test/regress/sql/partition_aggregate.sql
@@ -222,6 +222,7 @@ ANALYZE pagg_tab_ml;
-- For Parallel Append
SET max_parallel_workers_per_gather TO 2;
+SET parallel_setup_cost = 0;
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
@@ -235,6 +236,8 @@ SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HA
EXPLAIN (COSTS OFF)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3;
+RESET parallel_setup_cost;
+
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
-- PARTITION KEY, thus we will have a partial aggregation for them.
Attachments:
[text/plain] fix_partitionwise_aggregate_test_master.patch (2.1K, ../../CAApHDvr+95wJcABsqSXWcBa+yyUmJ2bq_mnZs9dUnaOi3XRVMg@mail.gmail.com/2-fix_partitionwise_aggregate_test_master.patch)
download | inline diff:
diff --git a/src/test/regress/expected/partition_aggregate.out b/src/test/regress/expected/partition_aggregate.out
index 4b41ccf1aa..db36e3a150 100644
--- a/src/test/regress/expected/partition_aggregate.out
+++ b/src/test/regress/expected/partition_aggregate.out
@@ -942,6 +942,7 @@ INSERT INTO pagg_tab_ml SELECT i % 30, i % 10, to_char(i % 4, 'FM0000') FROM gen
ANALYZE pagg_tab_ml;
-- For Parallel Append
SET max_parallel_workers_per_gather TO 2;
+SET parallel_setup_cost = 0;
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
-- PARTITION KEY, but still we do not see a partial aggregation as array_agg()
@@ -1025,6 +1026,7 @@ SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HA
-> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
(25 rows)
+RESET parallel_setup_cost;
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
-- PARTITION KEY, thus we will have a partial aggregation for them.
diff --git a/src/test/regress/sql/partition_aggregate.sql b/src/test/regress/sql/partition_aggregate.sql
index c17294b15b..ab070fee24 100644
--- a/src/test/regress/sql/partition_aggregate.sql
+++ b/src/test/regress/sql/partition_aggregate.sql
@@ -222,6 +222,7 @@ ANALYZE pagg_tab_ml;
-- For Parallel Append
SET max_parallel_workers_per_gather TO 2;
+SET parallel_setup_cost = 0;
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
@@ -235,6 +236,8 @@ SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HA
EXPLAIN (COSTS OFF)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3;
+RESET parallel_setup_cost;
+
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
-- PARTITION KEY, thus we will have a partial aggregation for them.
[text/plain] fix_partitionwise_aggregate_test_pg15.patch (7.7K, ../../CAApHDvr+95wJcABsqSXWcBa+yyUmJ2bq_mnZs9dUnaOi3XRVMg@mail.gmail.com/3-fix_partitionwise_aggregate_test_pg15.patch)
download | inline diff:
diff --git a/src/test/regress/expected/partition_aggregate.out b/src/test/regress/expected/partition_aggregate.out
index a08a3825ff..0dc6d63347 100644
--- a/src/test/regress/expected/partition_aggregate.out
+++ b/src/test/regress/expected/partition_aggregate.out
@@ -942,40 +942,43 @@ INSERT INTO pagg_tab_ml SELECT i % 30, i % 10, to_char(i % 4, 'FM0000') FROM gen
ANALYZE pagg_tab_ml;
-- For Parallel Append
SET max_parallel_workers_per_gather TO 2;
+SET parallel_setup_cost = 0;
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
-- PARTITION KEY, but still we do not see a partial aggregation as array_agg()
-- is not partial agg safe.
EXPLAIN (COSTS OFF)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3 ORDER BY 1, 2, 3;
- QUERY PLAN
---------------------------------------------------------------------------------------
- Sort
- Sort Key: pagg_tab_ml.a, (sum(pagg_tab_ml.b)), (array_agg(DISTINCT pagg_tab_ml.c))
- -> Append
- -> GroupAggregate
- Group Key: pagg_tab_ml.a
- Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml.a
- -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
- -> GroupAggregate
- Group Key: pagg_tab_ml_2.a
- Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_2.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
- -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
- -> GroupAggregate
- Group Key: pagg_tab_ml_5.a
- Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_5.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
- -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
-(25 rows)
+ QUERY PLAN
+--------------------------------------------------------------------------------------------
+ Gather Merge
+ Workers Planned: 2
+ -> Sort
+ Sort Key: pagg_tab_ml.a, (sum(pagg_tab_ml.b)), (array_agg(DISTINCT pagg_tab_ml.c))
+ -> Parallel Append
+ -> GroupAggregate
+ Group Key: pagg_tab_ml.a
+ Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml.a
+ -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_5.a
+ Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_5.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
+ -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_2.a
+ Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_2.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
+ -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
+(27 rows)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3 ORDER BY 1, 2, 3;
a | sum | array_agg | count
@@ -994,33 +997,36 @@ SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HA
-- Without ORDER BY clause, to test Gather at top-most path
EXPLAIN (COSTS OFF)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3;
- QUERY PLAN
----------------------------------------------------------------------
- Append
- -> GroupAggregate
- Group Key: pagg_tab_ml.a
- Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml.a
- -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
- -> GroupAggregate
- Group Key: pagg_tab_ml_2.a
- Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_2.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
- -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
- -> GroupAggregate
- Group Key: pagg_tab_ml_5.a
- Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
- -> Sort
- Sort Key: pagg_tab_ml_5.a
- -> Append
- -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
- -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
-(23 rows)
+ QUERY PLAN
+---------------------------------------------------------------------------
+ Gather
+ Workers Planned: 2
+ -> Parallel Append
+ -> GroupAggregate
+ Group Key: pagg_tab_ml.a
+ Filter: (avg(pagg_tab_ml.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml.a
+ -> Seq Scan on pagg_tab_ml_p1 pagg_tab_ml
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_5.a
+ Filter: (avg(pagg_tab_ml_5.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_5.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p3_s1 pagg_tab_ml_5
+ -> Seq Scan on pagg_tab_ml_p3_s2 pagg_tab_ml_6
+ -> GroupAggregate
+ Group Key: pagg_tab_ml_2.a
+ Filter: (avg(pagg_tab_ml_2.b) < '3'::numeric)
+ -> Sort
+ Sort Key: pagg_tab_ml_2.a
+ -> Append
+ -> Seq Scan on pagg_tab_ml_p2_s1 pagg_tab_ml_2
+ -> Seq Scan on pagg_tab_ml_p2_s2 pagg_tab_ml_3
+(25 rows)
+RESET parallel_setup_cost;
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
-- PARTITION KEY, thus we will have a partial aggregation for them.
diff --git a/src/test/regress/sql/partition_aggregate.sql b/src/test/regress/sql/partition_aggregate.sql
index c17294b15b..ab070fee24 100644
--- a/src/test/regress/sql/partition_aggregate.sql
+++ b/src/test/regress/sql/partition_aggregate.sql
@@ -222,6 +222,7 @@ ANALYZE pagg_tab_ml;
-- For Parallel Append
SET max_parallel_workers_per_gather TO 2;
+SET parallel_setup_cost = 0;
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
@@ -235,6 +236,8 @@ SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HA
EXPLAIN (COSTS OFF)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3;
+RESET parallel_setup_cost;
+
-- Full aggregation at level 1 as GROUP BY clause matches with PARTITION KEY
-- for level 1 only. For subpartitions, GROUP BY clause does not match with
-- PARTITION KEY, thus we will have a partial aggregation for them.
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-08-18 07:29 Tomas Vondra <[email protected]>
parent: David Rowley <[email protected]>
0 siblings, 1 reply; 38+ messages in thread
From: Tomas Vondra @ 2022-08-18 07:29 UTC (permalink / raw)
To: David Rowley <[email protected]>; +Cc: Michael Paquier <[email protected]>; Zhihong Yu <[email protected]>; Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On 8/18/22 03:32, David Rowley wrote:
> On Thu, 18 Aug 2022 at 02:46, Tomas Vondra
> <[email protected]> wrote:
>> So I don't think the current costing is wrong, but it certainly is more
>> complex. But the test does not test what it intended - I have two ideas
>> how to make it work:
>>
>> 1) increase the number of rows in the table
>>
>> 2) increase cpu_operator_cost (for that one test?)
>>
>> 3) tweak the costing somehow, to increase the cost a bit
>
> Why not, 4) SET parallel_setup_cost = 0; there are plenty of other
> places we do just that so we get a parallel plan without having to
> generate enough cost to drown out the parallel worker startup cost.
>
> Here are a couple of patches to demo the idea.
>
Yeah, that's an option too. I should have mentioned it along with the
cpu_operator_cost.
BTW would you mind taking a look at the costing? I think it's fine, but
it would be good if someone not involved in the patch takes a look.
regards
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-08-30 13:15 Jonathan S. Katz <[email protected]>
parent: Tomas Vondra <[email protected]>
0 siblings, 1 reply; 38+ messages in thread
From: Jonathan S. Katz @ 2022-08-30 13:15 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; David Rowley <[email protected]>; +Cc: Michael Paquier <[email protected]>; Zhihong Yu <[email protected]>; Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On 8/18/22 3:29 AM, Tomas Vondra wrote:
>
>
> On 8/18/22 03:32, David Rowley wrote:
>> Here are a couple of patches to demo the idea.
>>
>
> Yeah, that's an option too. I should have mentioned it along with the
> cpu_operator_cost.
>
> BTW would you mind taking a look at the costing? I think it's fine, but
> it would be good if someone not involved in the patch takes a look.
With RMT hat on, just a friendly reminder that this is still on the open
items list[1] -- we have the Beta 4 release next week and it would be
great if we can get this resolved prior to it.
Thanks!
Jonathan
[1] https://wiki.postgresql.org/wiki/PostgreSQL_15_Open_Items
[2]
https://www.postgresql.org/message-id/[email protected]
Attachments:
[application/pgp-signature] OpenPGP_signature (840B, ../../[email protected]/2-OpenPGP_signature)
download
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-09-01 13:06 Tomas Vondra <[email protected]>
parent: Jonathan S. Katz <[email protected]>
0 siblings, 1 reply; 38+ messages in thread
From: Tomas Vondra @ 2022-09-01 13:06 UTC (permalink / raw)
To: Jonathan S. Katz <[email protected]>; David Rowley <[email protected]>; +Cc: Michael Paquier <[email protected]>; Zhihong Yu <[email protected]>; Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On 8/30/22 15:15, Jonathan S. Katz wrote:
> On 8/18/22 3:29 AM, Tomas Vondra wrote:
>>
>>
>> On 8/18/22 03:32, David Rowley wrote:
>
>>> Here are a couple of patches to demo the idea.
>>>
>>
>> Yeah, that's an option too. I should have mentioned it along with the
>> cpu_operator_cost.
>>
>> BTW would you mind taking a look at the costing? I think it's fine, but
>> it would be good if someone not involved in the patch takes a look.
>
> With RMT hat on, just a friendly reminder that this is still on the open
> items list[1] -- we have the Beta 4 release next week and it would be
> great if we can get this resolved prior to it.
>
I know. I'll fix this by tweaking one of the cost gucs, mentioned by David.
regards
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-09-01 14:05 Jonathan S. Katz <[email protected]>
parent: Tomas Vondra <[email protected]>
0 siblings, 1 reply; 38+ messages in thread
From: Jonathan S. Katz @ 2022-09-01 14:05 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; David Rowley <[email protected]>; +Cc: Michael Paquier <[email protected]>; Zhihong Yu <[email protected]>; Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On 9/1/22 9:06 AM, Tomas Vondra wrote:
>
>
> On 8/30/22 15:15, Jonathan S. Katz wrote:
>> On 8/18/22 3:29 AM, Tomas Vondra wrote:
>>>
>>>
>>> On 8/18/22 03:32, David Rowley wrote:
>>
>>>> Here are a couple of patches to demo the idea.
>>>>
>>>
>>> Yeah, that's an option too. I should have mentioned it along with the
>>> cpu_operator_cost.
>>>
>>> BTW would you mind taking a look at the costing? I think it's fine, but
>>> it would be good if someone not involved in the patch takes a look.
>>
>> With RMT hat on, just a friendly reminder that this is still on the open
>> items list[1] -- we have the Beta 4 release next week and it would be
>> great if we can get this resolved prior to it.
>>
>
> I know. I'll fix this by tweaking one of the cost gucs, mentioned by David.
Great -- thanks for the update!
Jonathan
Attachments:
[application/pgp-signature] OpenPGP_signature (840B, ../../[email protected]/2-OpenPGP_signature)
download
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-09-04 22:14 Tomas Vondra <[email protected]>
parent: Jonathan S. Katz <[email protected]>
0 siblings, 2 replies; 38+ messages in thread
From: Tomas Vondra @ 2022-09-04 22:14 UTC (permalink / raw)
To: Jonathan S. Katz <[email protected]>; David Rowley <[email protected]>; +Cc: Michael Paquier <[email protected]>; Zhihong Yu <[email protected]>; Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On 9/1/22 16:05, Jonathan S. Katz wrote:
> On 9/1/22 9:06 AM, Tomas Vondra wrote:
>>
>>
>> On 8/30/22 15:15, Jonathan S. Katz wrote:
>>> On 8/18/22 3:29 AM, Tomas Vondra wrote:
>>>>
>>>>
>>>> On 8/18/22 03:32, David Rowley wrote:
>>>
>>>>> Here are a couple of patches to demo the idea.
>>>>>
>>>>
>>>> Yeah, that's an option too. I should have mentioned it along with the
>>>> cpu_operator_cost.
>>>>
>>>> BTW would you mind taking a look at the costing? I think it's fine, but
>>>> it would be good if someone not involved in the patch takes a look.
>>>
>>> With RMT hat on, just a friendly reminder that this is still on the open
>>> items list[1] -- we have the Beta 4 release next week and it would be
>>> great if we can get this resolved prior to it.
>>>
>>
>> I know. I'll fix this by tweaking one of the cost gucs, mentioned by
>> David.
>
> Great -- thanks for the update!
>
I've pushed the fix to 15+master. In the end I just used David's patches
that set parallel_setup_cost to 0.
regards
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-09-05 00:50 Jonathan S. Katz <[email protected]>
parent: Tomas Vondra <[email protected]>
1 sibling, 0 replies; 38+ messages in thread
From: Jonathan S. Katz @ 2022-09-05 00:50 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; David Rowley <[email protected]>; +Cc: Michael Paquier <[email protected]>; Zhihong Yu <[email protected]>; Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On 9/4/22 6:14 PM, Tomas Vondra wrote:
> I've pushed the fix to 15+master. In the end I just used David's patches
> that set parallel_setup_cost to 0.
Thanks! I have closed the open item.
Jonathan
Attachments:
[application/pgp-signature] OpenPGP_signature (840B, ../../[email protected]/2-OpenPGP_signature)
download
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-09-05 06:38 Michael Paquier <[email protected]>
parent: Tomas Vondra <[email protected]>
1 sibling, 0 replies; 38+ messages in thread
From: Michael Paquier @ 2022-09-05 06:38 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; +Cc: Jonathan S. Katz <[email protected]>; David Rowley <[email protected]>; Zhihong Yu <[email protected]>; Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On Mon, Sep 05, 2022 at 12:14:48AM +0200, Tomas Vondra wrote:
> I've pushed the fix to 15+master. In the end I just used David's patches
> that set parallel_setup_cost to 0.
Thanks, Tomas!
--
Michael
Attachments:
[application/pgp-signature] signature.asc (833B, ../../YxWZTGMr%[email protected]/2-signature.asc)
download
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2022-09-19 21:16 David Rowley <[email protected]>
parent: David Rowley <[email protected]>
0 siblings, 1 reply; 38+ messages in thread
From: David Rowley @ 2022-09-19 21:16 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; +Cc: Zhihong Yu <[email protected]>; Andrey V. Lepikhov <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>
On Wed, 13 Jul 2022 at 15:37, David Rowley <[email protected]> wrote:
> I'm just in this general area of the code again today and wondered
> about the header comment for the preprocess_groupclause() function.
>
> It says:
>
> * In principle it might be interesting to consider other orderings of the
> * GROUP BY elements, which could match the sort ordering of other
> * possible plans (eg an indexscan) and thereby reduce cost. We don't
> * bother with that, though. Hashed grouping will frequently win anyway.
>
> I'd say this commit makes that paragraph mostly obsolete. It's only
> true now in the sense that we don't try orders that suit some index
> that would provide pre-sorted results for a GroupAggregate path. The
> comment leads me to believe that we don't do anything at all to find a
> better order, and that's not true now.
I've just pushed a fix for this.
David
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2023-07-20 06:37 Andrey Lepikhov <[email protected]>
parent: David Rowley <[email protected]>
0 siblings, 1 reply; 38+ messages in thread
From: Andrey Lepikhov @ 2023-07-20 06:37 UTC (permalink / raw)
To: PostgreSQL Developers <[email protected]>; Tom Lane <[email protected]>; +Cc: Zhihong Yu <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; David Rowley <[email protected]>; Tomas Vondra <[email protected]>; a.rybakina <[email protected]>; Белялов Дамир Наилевич <[email protected]>
On 3/10/2022 21:56, Tom Lane wrote:
> Revert "Optimize order of GROUP BY keys".
>
> This reverts commit db0d67db2401eb6238ccc04c6407a4fd4f985832 and
> several follow-on fixes.
> ...
> Since we're hard up against the release deadline for v15, let's
> revert these changes for now. We can always try again later.
It may be time to restart the project. As a first step, I rebased the
patch on the current master. It wasn't trivial because of some latest
optimizations (a29eab, 1349d27 and 8d83a5d).
Now, Let's repeat the review and rewrite the current path according to
the reasons uttered in the revert commit.
--
regards,
Andrey Lepikhov
Postgres Professional
From 913d55ee887dccfeba360f5f44ed347dd1ba9044 Mon Sep 17 00:00:00 2001
From: "Andrey V. Lepikhov" <[email protected]>
Date: Fri, 14 Jul 2023 10:29:36 +0700
Subject: [PATCH] When evaluating a query with a multi-column GROUP BY clause
using sort, the cost may be heavily dependent on the order in which the keys
are compared when building the groups. Grouping does not imply any ordering,
so we're allowed to compare the keys in arbitrary order, and a Hash Agg
leverages this. But for Group Agg, we simply compared keys in the order as
specified in the query. This commit explores alternative ordering of the
keys, trying to find a cheaper one.
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
In principle, we might generate grouping paths for all permutations of
the keys, and leave the rest to the optimizer. But that might get very
expensive, so we try to pick only a couple interesting orderings based
on both local and global information.
When planning the grouping path, we explore statistics (number of
distinct values, cost of the comparison function) for the keys and
reorder them to minimize comparison costs. Intuitively, it may be better
to perform more expensive comparisons (for complex data types etc.)
last, because maybe the cheaper comparisons will be enough. Similarly,
the higher the cardinality of a key, the lower the probability we’ll
need to compare more keys. The patch generates and costs various
orderings, picking the cheapest ones.
The ordering of group keys may interact with other parts of the query,
some of which may not be known while planning the grouping. E.g. there
may be an explicit ORDER BY clause, or some other ordering-dependent
operation, higher up in the query, and using the same ordering may allow
using either incremental sort or even eliminate the sort entirely.
The patch generates orderings and picks those minimizing the comparison
cost (for various pathkeys), and then adds orderings that might be
useful for operations higher up in the plan (ORDER BY, etc.). Finally,
it always keeps the ordering specified in the query, on the assumption
the user might have additional insights.
This introduces a new GUC enable_group_by_reordering, so that the
optimization may be disabled if needed.
---
.../postgres_fdw/expected/postgres_fdw.out | 15 +-
src/backend/optimizer/path/costsize.c | 363 ++++++++++-
src/backend/optimizer/path/equivclass.c | 13 +-
src/backend/optimizer/path/pathkeys.c | 602 ++++++++++++++++++
src/backend/optimizer/plan/planner.c | 467 ++++++++------
src/backend/optimizer/util/pathnode.c | 4 +-
src/backend/utils/adt/selfuncs.c | 38 +-
src/backend/utils/misc/guc_tables.c | 10 +
src/backend/utils/misc/postgresql.conf.sample | 1 +
src/include/nodes/pathnodes.h | 10 +
src/include/optimizer/cost.h | 4 +-
src/include/optimizer/paths.h | 7 +
src/include/utils/selfuncs.h | 5 +
src/test/regress/expected/aggregates.out | 244 ++++++-
.../regress/expected/incremental_sort.out | 2 +-
src/test/regress/expected/join.out | 51 +-
src/test/regress/expected/merge.out | 15 +-
.../regress/expected/partition_aggregate.out | 44 +-
src/test/regress/expected/partition_join.out | 75 +--
src/test/regress/expected/sysviews.out | 3 +-
src/test/regress/expected/union.out | 60 +-
src/test/regress/sql/aggregates.sql | 99 +++
src/test/regress/sql/incremental_sort.sql | 2 +-
23 files changed, 1760 insertions(+), 374 deletions(-)
diff --git a/contrib/postgres_fdw/expected/postgres_fdw.out b/contrib/postgres_fdw/expected/postgres_fdw.out
index 852b5b4707..3f3c181ac8 100644
--- a/contrib/postgres_fdw/expected/postgres_fdw.out
+++ b/contrib/postgres_fdw/expected/postgres_fdw.out
@@ -9593,13 +9593,16 @@ SELECT t1.a,t2.b,t3.c FROM fprt1 t1 INNER JOIN fprt2 t2 ON (t1.a = t2.b) INNER J
-- left outer join + nullable clause
EXPLAIN (VERBOSE, COSTS OFF)
SELECT t1.a,t2.b,t2.c FROM fprt1 t1 LEFT JOIN (SELECT * FROM fprt2 WHERE a < 10) t2 ON (t1.a = t2.b and t1.b = t2.a) WHERE t1.a < 10 ORDER BY 1,2,3;
- QUERY PLAN
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
- Foreign Scan
+ QUERY PLAN
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+ Sort
Output: t1.a, fprt2.b, fprt2.c
- Relations: (public.ftprt1_p1 t1) LEFT JOIN (public.ftprt2_p1 fprt2)
- Remote SQL: SELECT r5.a, r6.b, r6.c FROM (public.fprt1_p1 r5 LEFT JOIN public.fprt2_p1 r6 ON (((r5.a = r6.b)) AND ((r5.b = r6.a)) AND ((r6.a < 10)))) WHERE ((r5.a < 10)) ORDER BY r5.a ASC NULLS LAST, r6.b ASC NULLS LAST, r6.c ASC NULLS LAST
-(4 rows)
+ Sort Key: t1.a, fprt2.b, fprt2.c
+ -> Foreign Scan
+ Output: t1.a, fprt2.b, fprt2.c
+ Relations: (public.ftprt1_p1 t1) LEFT JOIN (public.ftprt2_p1 fprt2)
+ Remote SQL: SELECT r5.a, r6.b, r6.c FROM (public.fprt1_p1 r5 LEFT JOIN public.fprt2_p1 r6 ON (((r5.a = r6.b)) AND ((r5.b = r6.a)) AND ((r6.a < 10)))) WHERE ((r5.a < 10))
+(7 rows)
SELECT t1.a,t2.b,t2.c FROM fprt1 t1 LEFT JOIN (SELECT * FROM fprt2 WHERE a < 10) t2 ON (t1.a = t2.b and t1.b = t2.a) WHERE t1.a < 10 ORDER BY 1,2,3;
a | b | c
diff --git a/src/backend/optimizer/path/costsize.c b/src/backend/optimizer/path/costsize.c
index ef475d95a1..1e979d57f2 100644
--- a/src/backend/optimizer/path/costsize.c
+++ b/src/backend/optimizer/path/costsize.c
@@ -1816,6 +1816,327 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
rterm->pathtarget->width);
}
+/*
+ * is_fake_var
+ * Workaround for generate_append_tlist() which generates fake Vars with
+ * varno == 0, that will cause a fail of estimate_num_group() call
+ *
+ * XXX Ummm, why would estimate_num_group fail with this?
+ */
+static bool
+is_fake_var(Expr *expr)
+{
+ if (IsA(expr, RelabelType))
+ expr = (Expr *) ((RelabelType *) expr)->arg;
+
+ return (IsA(expr, Var) && ((Var *) expr)->varno == 0);
+}
+
+/*
+ * get_width_cost_multiplier
+ * Returns relative complexity of comparing two values based on its width.
+ * The idea behind is that the comparison becomes more expensive the longer the
+ * value is. Return value is in cpu_operator_cost units.
+ */
+static double
+get_width_cost_multiplier(PlannerInfo *root, Expr *expr)
+{
+ double width = -1.0; /* fake value */
+
+ if (IsA(expr, RelabelType))
+ expr = (Expr *) ((RelabelType *) expr)->arg;
+
+ /* Try to find actual stat in corresponding relation */
+ if (IsA(expr, Var))
+ {
+ Var *var = (Var *) expr;
+
+ if (var->varno > 0 && var->varno < root->simple_rel_array_size)
+ {
+ RelOptInfo *rel = root->simple_rel_array[var->varno];
+
+ if (rel != NULL &&
+ var->varattno >= rel->min_attr &&
+ var->varattno <= rel->max_attr)
+ {
+ int ndx = var->varattno - rel->min_attr;
+
+ if (rel->attr_widths[ndx] > 0)
+ width = rel->attr_widths[ndx];
+ }
+ }
+ }
+
+ /* Didn't find any actual stats, try using type width instead. */
+ if (width < 0.0)
+ {
+ Node *node = (Node *) expr;
+
+ width = get_typavgwidth(exprType(node), exprTypmod(node));
+ }
+
+ /*
+ * Values are passed as Datum type, so comparisons can't be cheaper than
+ * comparing a Datum value.
+ *
+ * FIXME I find this reasoning questionable. We may pass int2, and
+ * comparing it is probably a bit cheaper than comparing a bigint.
+ */
+ if (width <= sizeof(Datum))
+ return 1.0;
+
+ /*
+ * We consider the cost of a comparison not to be directly proportional to
+ * width of the argument, because widths of the arguments could be
+ * slightly different (we only know the average width for the whole
+ * column). So we use log16(width) as an estimate.
+ */
+ return 1.0 + 0.125 * LOG2(width / sizeof(Datum));
+}
+
+/*
+ * compute_cpu_sort_cost
+ * compute CPU cost of sort (i.e. in-memory)
+ *
+ * The main thing we need to calculate to estimate sort CPU costs is the number
+ * of calls to the comparator functions. The difficulty is that for multi-column
+ * sorts there may be different data types involved (for some of which the calls
+ * may be much more expensive). Furthermore, columns may have a very different
+ * number of distinct values - the higher the number, the fewer comparisons will
+ * be needed for the following columns.
+ *
+ * The algorithm is incremental - we add pathkeys one by one, and at each step we
+ * estimate the number of necessary comparisons (based on the number of distinct
+ * groups in the current pathkey prefix and the new pathkey), and the comparison
+ * costs (which is data type specific).
+ *
+ * Estimation of the number of comparisons is based on ideas from:
+ *
+ * "Quicksort Is Optimal", Robert Sedgewick, Jon Bentley, 2002
+ * [https://www.cs.princeton.edu/~rs/talks/QuicksortIsOptimal.pdf]
+ *
+ * In term of that paper, let N - number of tuples, Xi - number of identical
+ * tuples with value Ki, then the estimate of number of comparisons is:
+ *
+ * log(N! / (X1! * X2! * ..)) ~ sum(Xi * log(N/Xi))
+ *
+ * We assume all Xi the same because now we don't have any estimation of
+ * group sizes, we have only know the estimate of number of groups (distinct
+ * values). In that case, formula becomes:
+ *
+ * N * log(NumberOfGroups)
+ *
+ * For multi-column sorts we need to estimate the number of comparisons for
+ * each individual column - for example with columns (c1, c2, ..., ck) we
+ * can estimate that number of comparisons on ck is roughly
+ *
+ * ncomparisons(c1, c2, ..., ck) / ncomparisons(c1, c2, ..., c(k-1))
+ *
+ * Let k be a column number, Gk - number of groups defined by k columns, and Fk
+ * the cost of the comparison is
+ *
+ * N * sum( Fk * log(Gk) )
+ *
+ * Note: We also consider column width, not just the comparator cost.
+ *
+ * NOTE: some callers currently pass NIL for pathkeys because they
+ * can't conveniently supply the sort keys. In this case, it will fallback to
+ * simple comparison cost estimate.
+ */
+static Cost
+compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
+ Cost comparison_cost, double tuples, double output_tuples,
+ bool heapSort)
+{
+ Cost per_tuple_cost = 0.0;
+ ListCell *lc;
+ List *pathkeyExprs = NIL;
+ double tuplesPerPrevGroup = tuples;
+ double totalFuncCost = 1.0;
+ bool has_fake_var = false;
+ int i = 0;
+ Oid prev_datatype = InvalidOid;
+ List *cache_varinfos = NIL;
+
+ /* fallback if pathkeys is unknown */
+ if (list_length(pathkeys) == 0)
+ {
+ /*
+ * If we'll use a bounded heap-sort keeping just K tuples in memory,
+ * for a total number of tuple comparisons of N log2 K; but the
+ * constant factor is a bit higher than for quicksort. Tweak it so
+ * that the cost curve is continuous at the crossover point.
+ */
+ output_tuples = (heapSort) ? 2.0 * output_tuples : tuples;
+ per_tuple_cost += 2.0 * cpu_operator_cost * LOG2(output_tuples);
+
+ /* add cost provided by caller */
+ per_tuple_cost += comparison_cost;
+
+ return per_tuple_cost * tuples;
+ }
+
+ /*
+ * Computing total cost of sorting takes into account the per-column
+ * comparison function cost. We try to compute the needed number of
+ * comparisons per column.
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(lc);
+ EquivalenceMember *em;
+ double nGroups,
+ correctedNGroups;
+ Cost funcCost = 1.0;
+
+ /*
+ * We believe that equivalence members aren't very different, so, to
+ * estimate cost we consider just the first member.
+ */
+ em = (EquivalenceMember *) linitial(pathkey->pk_eclass->ec_members);
+
+ if (em->em_datatype != InvalidOid)
+ {
+ /* do not lookup funcCost if the data type is the same */
+ if (prev_datatype != em->em_datatype)
+ {
+ Oid sortop;
+ QualCost cost;
+
+ sortop = get_opfamily_member(pathkey->pk_opfamily,
+ em->em_datatype, em->em_datatype,
+ pathkey->pk_strategy);
+
+ cost.startup = 0;
+ cost.per_tuple = 0;
+ add_function_cost(root, get_opcode(sortop), NULL, &cost);
+
+ /*
+ * add_function_cost returns the product of cpu_operator_cost
+ * and procost, but we need just procost, co undo that.
+ */
+ funcCost = cost.per_tuple / cpu_operator_cost;
+
+ prev_datatype = em->em_datatype;
+ }
+ }
+
+ /* factor in the width of the values in this column */
+ funcCost *= get_width_cost_multiplier(root, em->em_expr);
+
+ /* now we have per-key cost, so add to the running total */
+ totalFuncCost += funcCost;
+
+ /* remember if we have found a fake Var in pathkeys */
+ has_fake_var |= is_fake_var(em->em_expr);
+ pathkeyExprs = lappend(pathkeyExprs, em->em_expr);
+
+ /*
+ * We need to calculate the number of comparisons for this column,
+ * which requires knowing the group size. So we estimate the number of
+ * groups by calling estimate_num_groups_incremental(), which
+ * estimates the group size for "new" pathkeys.
+ *
+ * Note: estimate_num_groups_incremental does not handle fake Vars, so
+ * use a default estimate otherwise.
+ */
+ if (!has_fake_var)
+ nGroups = estimate_num_groups_incremental(root, pathkeyExprs,
+ tuplesPerPrevGroup, NULL, NULL,
+ &cache_varinfos,
+ list_length(pathkeyExprs) - 1);
+ else if (tuples > 4.0)
+
+ /*
+ * Use geometric mean as estimation if there are no stats.
+ *
+ * We don't use DEFAULT_NUM_DISTINCT here, because that's used for
+ * a single column, but here we're dealing with multiple columns.
+ */
+ nGroups = ceil(2.0 + sqrt(tuples) * (i + 1) / list_length(pathkeys));
+ else
+ nGroups = tuples;
+
+ /*
+ * Presorted keys are not considered in the cost above, but we still
+ * do have to compare them in the qsort comparator. So make sure to
+ * factor in the cost in that case.
+ */
+ if (i >= nPresortedKeys)
+ {
+ if (heapSort)
+ {
+ /*
+ * have to keep at least one group, and a multiple of group
+ * size
+ */
+ correctedNGroups = ceil(output_tuples / tuplesPerPrevGroup);
+ }
+ else
+ /* all groups in the input */
+ correctedNGroups = nGroups;
+
+ correctedNGroups = Max(1.0, ceil(correctedNGroups));
+
+ per_tuple_cost += totalFuncCost * LOG2(correctedNGroups);
+ }
+
+ i++;
+
+ /*
+ * Once we get single-row group, it means tuples in the group are
+ * unique and we can skip all remaining columns.
+ */
+ if (tuplesPerPrevGroup <= nGroups)
+ break;
+
+ /*
+ * Uniform distributions with all groups being of the same size are
+ * the best case, with nice smooth behavior. Real-world distributions
+ * tend not to be uniform, though, and we don't have any reliable
+ * easy-to-use information. As a basic defense against skewed
+ * distributions, we use a 1.5 factor to make the expected group a bit
+ * larger, but we need to be careful not to make the group larger than
+ * in the preceding step.
+ */
+ tuplesPerPrevGroup = Min(tuplesPerPrevGroup,
+ ceil(1.5 * tuplesPerPrevGroup / nGroups));
+ }
+
+ list_free(pathkeyExprs);
+
+ /* per_tuple_cost is in cpu_operator_cost units */
+ per_tuple_cost *= cpu_operator_cost;
+
+ /*
+ * Accordingly to "Introduction to algorithms", Thomas H. Cormen, Charles
+ * E. Leiserson, Ronald L. Rivest, ISBN 0-07-013143-0, quicksort
+ * estimation formula has additional term proportional to number of tuples
+ * (see Chapter 8.2 and Theorem 4.1). That affects cases with a low number
+ * of tuples, approximately less than 1e4. We could implement it as an
+ * additional multiplier under the logarithm, but we use a bit more
+ * complex formula which takes into account the number of unique tuples
+ * and it's not clear how to combine the multiplier with the number of
+ * groups. Estimate it as 10 cpu_operator_cost units.
+ */
+ per_tuple_cost += 10 * cpu_operator_cost;
+
+ per_tuple_cost += comparison_cost;
+
+ return tuples * per_tuple_cost;
+}
+
+/*
+ * simple wrapper just to estimate best sort path
+ */
+Cost
+cost_sort_estimate(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
+ double tuples)
+{
+ return compute_cpu_sort_cost(root, pathkeys, nPresortedKeys,
+ 0, tuples, tuples, false);
+}
+
/*
* cost_tuplesort
* Determines and returns the cost of sorting a relation using tuplesort,
@@ -1832,7 +2153,7 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
* number of initial runs formed and M is the merge order used by tuplesort.c.
* Since the average initial run should be about sort_mem, we have
* disk traffic = 2 * relsize * ceil(logM(p / sort_mem))
- * cpu = comparison_cost * t * log2(t)
+ * and cpu cost (computed by compute_cpu_sort_cost()).
*
* If the sort is bounded (i.e., only the first k result tuples are needed)
* and k tuples can fit into sort_mem, we use a heap method that keeps only
@@ -1851,9 +2172,11 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
* 'comparison_cost' is the extra cost per comparison, if any
* 'sort_mem' is the number of kilobytes of work memory allowed for the sort
* 'limit_tuples' is the bound on the number of output tuples; -1 if no bound
+ * 'startup_cost' is expected to be 0 at input. If there is "input cost" it should
+ * be added by caller later
*/
static void
-cost_tuplesort(Cost *startup_cost, Cost *run_cost,
+cost_tuplesort(PlannerInfo *root, List *pathkeys, Cost *startup_cost, Cost *run_cost,
double tuples, int width,
Cost comparison_cost, int sort_mem,
double limit_tuples)
@@ -1870,9 +2193,6 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
if (tuples < 2.0)
tuples = 2.0;
- /* Include the default cost-per-comparison */
- comparison_cost += 2.0 * cpu_operator_cost;
-
/* Do we have a useful LIMIT? */
if (limit_tuples > 0 && limit_tuples < tuples)
{
@@ -1896,12 +2216,10 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
double log_runs;
double npageaccesses;
- /*
- * CPU costs
- *
- * Assume about N log2 N comparisons
- */
- *startup_cost = comparison_cost * tuples * LOG2(tuples);
+ /* CPU costs */
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ tuples, false);
/* Disk costs */
@@ -1917,18 +2235,17 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
}
else if (tuples > 2 * output_tuples || input_bytes > sort_mem_bytes)
{
- /*
- * We'll use a bounded heap-sort keeping just K tuples in memory, for
- * a total number of tuple comparisons of N log2 K; but the constant
- * factor is a bit higher than for quicksort. Tweak it so that the
- * cost curve is continuous at the crossover point.
- */
- *startup_cost = comparison_cost * tuples * LOG2(2.0 * output_tuples);
+ /* We'll use a bounded heap-sort keeping just K tuples in memory. */
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ output_tuples, true);
}
else
{
/* We'll use plain quicksort on all the input tuples */
- *startup_cost = comparison_cost * tuples * LOG2(tuples);
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ tuples, false);
}
/*
@@ -2042,7 +2359,7 @@ cost_incremental_sort(Path *path,
* Estimate the average cost of sorting of one group where presorted keys
* are equal.
*/
- cost_tuplesort(&group_startup_cost, &group_run_cost,
+ cost_tuplesort(root, pathkeys, &group_startup_cost, &group_run_cost,
group_tuples, width, comparison_cost, sort_mem,
limit_tuples);
@@ -2102,7 +2419,7 @@ cost_sort(Path *path, PlannerInfo *root,
Cost startup_cost;
Cost run_cost;
- cost_tuplesort(&startup_cost, &run_cost,
+ cost_tuplesort(root, pathkeys, &startup_cost, &run_cost,
tuples, width,
comparison_cost, sort_mem,
limit_tuples);
@@ -2200,7 +2517,7 @@ append_nonpartial_cost(List *subpaths, int numpaths, int parallel_workers)
* Determines and returns the cost of an Append node.
*/
void
-cost_append(AppendPath *apath)
+cost_append(AppendPath *apath, PlannerInfo *root)
{
ListCell *l;
@@ -2268,7 +2585,7 @@ cost_append(AppendPath *apath)
* any child.
*/
cost_sort(&sort_path,
- NULL, /* doesn't currently need root */
+ root,
pathkeys,
subpath->total_cost,
subpath->rows,
diff --git a/src/backend/optimizer/path/equivclass.c b/src/backend/optimizer/path/equivclass.c
index 7fa502d6e2..07edd4f38e 100644
--- a/src/backend/optimizer/path/equivclass.c
+++ b/src/backend/optimizer/path/equivclass.c
@@ -652,7 +652,18 @@ get_eclass_for_sort_expr(PlannerInfo *root,
if (opcintype == cur_em->em_datatype &&
equal(expr, cur_em->em_expr))
- return cur_ec; /* Match! */
+ {
+ /*
+ * Match!
+ *
+ * Copy the sortref if it wasn't set yet. That may happen if
+ * the ec was constructed from WHERE clause, i.e. it doesn't
+ * have a target reference at all.
+ */
+ if (cur_ec->ec_sortref == 0 && sortref > 0)
+ cur_ec->ec_sortref = sortref;
+ return cur_ec;
+ }
}
}
diff --git a/src/backend/optimizer/path/pathkeys.c b/src/backend/optimizer/path/pathkeys.c
index e53ea84224..2143b39873 100644
--- a/src/backend/optimizer/path/pathkeys.c
+++ b/src/backend/optimizer/path/pathkeys.c
@@ -17,17 +17,24 @@
*/
#include "postgres.h"
+#include <float.h>
+
+#include "miscadmin.h"
#include "access/stratnum.h"
#include "catalog/pg_opfamily.h"
#include "nodes/makefuncs.h"
#include "nodes/nodeFuncs.h"
#include "nodes/plannodes.h"
+#include "optimizer/cost.h"
#include "optimizer/optimizer.h"
#include "optimizer/pathnode.h"
#include "optimizer/paths.h"
#include "partitioning/partbounds.h"
#include "utils/lsyscache.h"
+#include "utils/selfuncs.h"
+/* Consider reordering of GROUP BY keys? */
+bool enable_group_by_reordering = true;
static bool pathkey_is_redundant(PathKey *new_pathkey, List *pathkeys);
static bool matches_boolean_partition_clause(RestrictInfo *rinfo,
@@ -350,6 +357,548 @@ pathkeys_contained_in(List *keys1, List *keys2)
return false;
}
+/*
+ * group_keys_reorder_by_pathkeys
+ * Reorder GROUP BY keys to match pathkeys of input path.
+ *
+ * Function returns new lists (pathkeys and clauses), original GROUP BY lists
+ * stay untouched.
+ *
+ * Returns the number of GROUP BY keys with a matching pathkey.
+ */
+int
+group_keys_reorder_by_pathkeys(List *pathkeys, List **group_pathkeys,
+ List **group_clauses,
+ int num_groupby_pathkeys)
+{
+ List *new_group_pathkeys = NIL,
+ *new_group_clauses = NIL;
+ ListCell *lc;
+ int n;
+
+ if (pathkeys == NIL || *group_pathkeys == NIL)
+ return 0;
+
+ /*
+ * Walk the pathkeys (determining ordering of the input path) and see if
+ * there's a matching GROUP BY key. If we find one, we append it to the
+ * list, and do the same for the clauses.
+ *
+ * Once we find the first pathkey without a matching GROUP BY key, the
+ * rest of the pathkeys are useless and can't be used to evaluate the
+ * grouping, so we abort the loop and ignore the remaining pathkeys.
+ *
+ * XXX Pathkeys are built in a way to allow simply comparing pointers.
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(lc);
+ SortGroupClause *sgc;
+
+ if (foreach_current_index(lc) >= num_groupby_pathkeys ||
+ !list_member_ptr(*group_pathkeys, pathkey))
+ /* */
+ break;
+
+ /* abort on first mismatch */
+ sgc = get_sortgroupref_clause_noerr(pathkey->pk_eclass->ec_sortref,
+ *group_clauses);
+ if (!sgc)
+ /* The grouping clause is not cover this paathkey */
+ break;
+
+ new_group_pathkeys = lappend(new_group_pathkeys, pathkey);
+ new_group_clauses = lappend(new_group_clauses, sgc);
+ }
+
+ /* remember the number of pathkeys with a matching GROUP BY key */
+ n = list_length(new_group_pathkeys);
+
+ /* append the remaining group pathkeys (will be treated as not sorted) */
+ *group_pathkeys = list_concat_unique_ptr(new_group_pathkeys,
+ *group_pathkeys);
+ *group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+
+ return n;
+}
+
+/*
+ * Used to generate all permutations of a pathkey list.
+ */
+typedef struct PathkeyMutatorState
+{
+ List *elemsList;
+ ListCell **elemCells;
+ void **elems;
+ int *positions;
+ int mutatorNColumns;
+ int count;
+} PathkeyMutatorState;
+
+
+/*
+ * PathkeyMutatorInit
+ * Initialize state of the permutation generator.
+ *
+ * We want to generate permutations of elements in the "elems" list. We may want
+ * to skip some number of elements at the beginning (when treating as presorted)
+ * or at the end (we only permute a limited number of group keys).
+ *
+ * The list is decomposed into elements, and we also keep pointers to individual
+ * cells. This allows us to build the permuted list quickly and cheaply, without
+ * creating any copies.
+ */
+static void
+PathkeyMutatorInit(PathkeyMutatorState *state, List *elems, int start, int end)
+{
+ int i;
+ int n = end - start;
+ ListCell *lc;
+
+ memset(state, 0, sizeof(*state));
+
+ state->mutatorNColumns = n;
+
+ state->elemsList = list_copy(elems);
+
+ state->elems = palloc(sizeof(void *) * n);
+ state->elemCells = palloc(sizeof(ListCell *) * n);
+ state->positions = palloc(sizeof(int) * n);
+
+ i = 0;
+ for_each_cell(lc, state->elemsList, list_nth_cell(state->elemsList, start))
+ {
+ state->elemCells[i] = lc;
+ state->elems[i] = lfirst(lc);
+ state->positions[i] = i + 1;
+ i++;
+
+ if (i >= n)
+ break;
+ }
+}
+
+/* Swap two elements of an array. */
+static void
+PathkeyMutatorSwap(int *a, int i, int j)
+{
+ int s = a[i];
+
+ a[i] = a[j];
+ a[j] = s;
+}
+
+/*
+ * Generate the next permutation of elements.
+ */
+static bool
+PathkeyMutatorNextSet(int *a, int n)
+{
+ int j,
+ k,
+ l,
+ r;
+
+ j = n - 2;
+
+ while (j >= 0 && a[j] >= a[j + 1])
+ j--;
+
+ if (j < 0)
+ return false;
+
+ k = n - 1;
+
+ while (k >= 0 && a[j] >= a[k])
+ k--;
+
+ PathkeyMutatorSwap(a, j, k);
+
+ l = j + 1;
+ r = n - 1;
+
+ while (l < r)
+ PathkeyMutatorSwap(a, l++, r--);
+
+ return true;
+}
+
+/*
+ * PathkeyMutatorNext
+ * Generate the next permutation of list of elements.
+ *
+ * Returns the next permutation (as a list of elements) or NIL if there are no
+ * more permutations.
+ */
+static List *
+PathkeyMutatorNext(PathkeyMutatorState *state)
+{
+ int i;
+
+ state->count++;
+
+ /* first permutation is original list */
+ if (state->count == 1)
+ return state->elemsList;
+
+ /* when there are no more permutations, return NIL */
+ if (!PathkeyMutatorNextSet(state->positions, state->mutatorNColumns))
+ {
+ pfree(state->elems);
+ pfree(state->elemCells);
+ pfree(state->positions);
+
+ list_free(state->elemsList);
+
+ return NIL;
+ }
+
+ /* update the list cells to point to the right elements */
+ for (i = 0; i < state->mutatorNColumns; i++)
+ lfirst(state->elemCells[i]) =
+ (void *) state->elems[state->positions[i] - 1];
+
+ return state->elemsList;
+}
+
+/*
+ * Cost of comparing pathkeys.
+ */
+typedef struct PathkeySortCost
+{
+ Cost cost;
+ PathKey *pathkey;
+} PathkeySortCost;
+
+static int
+pathkey_sort_cost_comparator(const void *_a, const void *_b)
+{
+ const PathkeySortCost *a = (PathkeySortCost *) _a;
+ const PathkeySortCost *b = (PathkeySortCost *) _b;
+
+ if (a->cost < b->cost)
+ return -1;
+ else if (a->cost == b->cost)
+ return 0;
+ return 1;
+}
+
+/*
+ * get_cheapest_group_keys_order
+ * Reorders the group pathkeys / clauses to minimize the comparison cost.
+ *
+ * Given the list of pathkeys in '*group_pathkeys', we try to arrange these
+ * in an order that minimizes the sort costs that will be incurred by the
+ * GROUP BY. The costs mainly depend on the cost of the sort comparator
+ * function(s) and the number of distinct values in each column of the GROUP
+ * BY clause (*group_clauses). Sorting on subsequent columns is only required
+ * for tiebreak situations where two values sort equally.
+ *
+ * In case the input is partially sorted, only the remaining pathkeys are
+ * considered. 'n_preordered' denotes how many of the leading *group_pathkeys
+ * the input is presorted by.
+ *
+ * Returns true and sets *group_pathkeys and *group_clauses to the newly
+ * ordered versions of the lists that were passed in via these parameters.
+ * If no reordering was deemed necessary then we return false, in which case
+ * the *group_pathkeys and *group_clauses lists are left untouched. The
+ * original *group_pathkeys and *group_clauses parameter values are never
+ * destructively modified in place.
+ */
+static bool
+get_cheapest_group_keys_order(PlannerInfo *root, double nrows,
+ List **group_pathkeys, List **group_clauses,
+ int n_preordered)
+{
+ List *new_group_pathkeys = NIL;
+ List *new_group_clauses = NIL;
+ List *var_group_pathkeys;
+ ListCell *cell;
+ PathkeyMutatorState mstate;
+ double cheapest_sort_cost = DBL_MAX;
+ int nFreeKeys;
+ int nToPermute;
+
+ /* If there are less than 2 unsorted pathkeys, we're done. */
+ if (root->num_groupby_pathkeys - n_preordered < 2)
+ return false;
+
+ /*
+ * We could exhaustively cost all possible orderings of the pathkeys, but
+ * for a large number of pathkeys it might be prohibitively expensive. So
+ * we try to apply simple cheap heuristics first - we sort the pathkeys by
+ * sort cost (as if the pathkey was sorted independently) and then check
+ * only the four cheapest pathkeys. The remaining pathkeys are kept
+ * ordered by cost.
+ *
+ * XXX This is a very simple heuristics, but likely to work fine for most
+ * cases (because the number of GROUP BY clauses tends to be lower than
+ * 4). But it ignores how the number of distinct values in each pathkey
+ * affects the following steps. It might be better to use "more expensive"
+ * pathkey first if it has many distinct values, because it then limits
+ * the number of comparisons for the remaining pathkeys. But evaluating
+ * that is likely quite the expensive.
+ */
+ nFreeKeys = root->num_groupby_pathkeys - n_preordered;
+ nToPermute = 4;
+ if (nFreeKeys > nToPermute)
+ {
+ PathkeySortCost *costs = palloc(sizeof(PathkeySortCost) * nFreeKeys);
+ PathkeySortCost *cost = costs;
+ int i = 0;
+
+ /*
+ * Estimate cost for sorting individual pathkeys skipping the
+ * pre-ordered pathkeys.
+ */
+ for_each_from(cell, *group_pathkeys, n_preordered)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(cell);
+ List *to_cost;
+
+ if (foreach_current_index(cell) >= root->num_groupby_pathkeys)
+ break;
+
+ i++;
+ to_cost = list_make1(pathkey);
+ cost->pathkey = pathkey;
+ cost->cost = cost_sort_estimate(root, to_cost, 0, nrows);
+ cost++;
+
+ list_free(to_cost);
+ }
+
+ Assert(i == nFreeKeys);
+ /* sort the pathkeys by sort cost in ascending order */
+ qsort(costs, nFreeKeys, sizeof(*costs), pathkey_sort_cost_comparator);
+
+ /*
+ * Rebuild the list of pathkeys - first the preordered ones, then the
+ * rest ordered by cost.
+ */
+ new_group_pathkeys = list_copy_head(*group_pathkeys, n_preordered);
+
+ for (int i = 0; i < nFreeKeys; i++)
+ new_group_pathkeys = lappend(new_group_pathkeys, costs[i].pathkey);
+
+ new_group_pathkeys = list_copy_tail(*group_pathkeys,
+ root->num_groupby_pathkeys);
+ pfree(costs);
+ }
+ else
+ {
+ /* Copy the list, so that we can free the new list by list_free. */
+ new_group_pathkeys = list_copy(*group_pathkeys);
+ nToPermute = nFreeKeys;
+ }
+
+ Assert(list_length(new_group_pathkeys) == list_length(*group_pathkeys));
+
+ /*
+ * Generate pathkey lists with permutations of the first nToPermute
+ * pathkeys.
+ *
+ * XXX We simply calculate sort cost for each individual pathkey list, but
+ * there's room for two dynamic programming optimizations here. Firstly,
+ * we may pass the current "best" cost to cost_sort_estimate so that it
+ * can "abort" if the estimated pathkeys list exceeds it. Secondly, it
+ * could pass the return information about the position when it exceeded
+ * the cost, and we could skip all permutations with the same prefix.
+ *
+ * Imagine we've already found ordering with cost C1, and we're evaluating
+ * another ordering - cost_sort_estimate() calculates cost by adding the
+ * pathkeys one by one (more or less), and the cost only grows. If at any
+ * point it exceeds C1, it can't possibly be "better" so we can discard
+ * it. But we also know that we can discard all ordering with the same
+ * prefix, because if we're estimating (a,b,c,d) and we exceed C1 at (a,b)
+ * then the same thing will happen for any ordering with this prefix.
+ */
+ PathkeyMutatorInit(&mstate, new_group_pathkeys, n_preordered,
+ n_preordered + nToPermute);
+
+ while ((var_group_pathkeys = PathkeyMutatorNext(&mstate)) != NIL)
+ {
+ Cost cost;
+
+ cost = cost_sort_estimate(root, var_group_pathkeys, n_preordered, nrows);
+
+ if (cost < cheapest_sort_cost)
+ {
+ list_free(new_group_pathkeys);
+ new_group_pathkeys = list_copy(var_group_pathkeys);
+ cheapest_sort_cost = cost;
+ }
+ }
+
+ Assert(list_length(new_group_pathkeys) == list_length(root->group_pathkeys));
+
+ /* Reorder the group clauses according to the reordered pathkeys. */
+ foreach(cell, new_group_pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(cell);
+
+ if (foreach_current_index(cell) >= root->num_groupby_pathkeys)
+ break;
+
+ new_group_clauses =
+ lappend(new_group_clauses,
+ get_sortgroupref_clause(pathkey->pk_eclass->ec_sortref,
+ *group_clauses));
+ }
+
+ /* Just append the rest GROUP BY clauses */
+ new_group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+ *group_pathkeys = new_group_pathkeys;
+ *group_clauses = new_group_clauses;
+ Assert(list_length(new_group_clauses) == list_length(root->processed_groupClause));
+
+ return true;
+}
+
+/*
+ * get_useful_group_keys_orderings
+ * Determine which orderings of GROUP BY keys are potentially interesting.
+ *
+ * Returns list of PathKeyInfo items, each representing an interesting ordering
+ * of GROUP BY keys. Each item stores pathkeys and clauses in matching order.
+ *
+ * The function considers (and keeps) multiple group by orderings:
+ *
+ * - the original ordering, as specified by the GROUP BY clause
+ *
+ * - GROUP BY keys reordered to minimize the sort cost
+ *
+ * - GROUP BY keys reordered to match path ordering (as much as possible), with
+ * the tail reordered to minimize the sort cost
+ *
+ * - GROUP BY keys to match target ORDER BY clause (as much as possible), with
+ * the tail reordered to minimize the sort cost
+ *
+ * There are other potentially interesting orderings (e.g. it might be best to
+ * match the first ORDER BY key, order the remaining keys differently and then
+ * rely on the incremental sort to fix this), but we ignore those for now. To
+ * make this work we'd have to pretty much generate all possible permutations.
+ */
+List *
+get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys)
+{
+ Query *parse = root->parse;
+ List *infos = NIL;
+ PathKeyInfo *info;
+ int n_preordered = 0;
+
+ List *pathkeys = root->group_pathkeys;
+ List *clauses = root->processed_groupClause;
+
+ /* always return at least the original pathkeys/clauses */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+ infos = lappend(infos, info);
+
+ /*
+ * Should we try generating alternative orderings of the group keys? If
+ * not, we produce only the order specified in the query, i.e. the
+ * optimization is effectively disabled.
+ */
+ if (!enable_group_by_reordering)
+ return infos;
+
+ /* for grouping sets we can't do any reordering */
+ if (parse->groupingSets)
+ return infos;
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost, ignoring both the
+ * target ordering (ORDER BY) and ordering of the input path.
+ */
+ if (get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered))
+ {
+ Assert(list_length(pathkeys) == list_length(clauses));
+
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * If the path is sorted in some way, try reordering the group keys to
+ * match as much of the ordering as possible - we get this sort for free
+ * (mostly).
+ *
+ * We must not do this when there are no grouping sets, because those use
+ * more complex logic to decide the ordering.
+ *
+ * XXX Isn't this somewhat redundant with presorted_keys? Actually, it's
+ * more a complement, because it allows benefiting from incremental sort
+ * as much as possible.
+ *
+ * XXX This does nothing if (n_preordered == 0). We shouldn't create the
+ * info in this case.
+ */
+ if (path_pathkeys)
+ {
+ n_preordered = group_keys_reorder_by_pathkeys(path_pathkeys,
+ &pathkeys,
+ &clauses,
+ root->num_groupby_pathkeys);
+
+ /* reorder the tail to minimize sort cost */
+ get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, in
+ * which case get_cheapest_group_keys_order returns false. But we
+ * still want to keep the keys reordered to path_pathkeys.
+ */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost (this time consider
+ * the ORDER BY clause, but only if set debug_group_by_match_order_by).
+ */
+ if (root->sort_pathkeys)
+ {
+ n_preordered = group_keys_reorder_by_pathkeys(root->sort_pathkeys,
+ &pathkeys,
+ &clauses,
+ root->num_groupby_pathkeys);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, in
+ * which case get_cheapest_group_keys_order returns false. But we
+ * still want to keep the keys reordered to sort_pathkeys.
+ */
+ get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered);
+
+ /* keep the group keys reordered to match ordering of input path */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ return infos;
+}
+
/*
* pathkeys_count_contained_in
* Same as pathkeys_contained_in, but also sets length of longest
@@ -1937,6 +2486,54 @@ pathkeys_useful_for_ordering(PlannerInfo *root, List *pathkeys)
return n_common_pathkeys;
}
+/*
+ * pathkeys_useful_for_grouping
+ * Count the number of pathkeys that are useful for grouping (instead of
+ * explicit sort)
+ *
+ * Group pathkeys could be reordered to benefit from the ordering. The
+ * ordering may not be "complete" and may require incremental sort, but that's
+ * fine. So we simply count prefix pathkeys with a matching group key, and
+ * stop once we find the first pathkey without a match.
+ *
+ * So e.g. with pathkeys (a,b,c) and group keys (a,b,e) this determines (a,b)
+ * pathkeys are useful for grouping, and we might do incremental sort to get
+ * path ordered by (a,b,e).
+ *
+ * This logic is necessary to retain paths with ordering not matching grouping
+ * keys directly, without the reordering.
+ *
+ * Returns the length of pathkey prefix with matching group keys.
+ */
+static int
+pathkeys_useful_for_grouping(PlannerInfo *root, List *pathkeys)
+{
+ ListCell *key;
+ int n = 0;
+
+ /* no special ordering requested for grouping */
+ if (root->group_pathkeys == NIL)
+ return 0;
+
+ /* unordered path */
+ if (pathkeys == NIL)
+ return 0;
+
+ /* walk the pathkeys and search for matching group key */
+ foreach(key, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(key);
+
+ /* no matching group key, we're done */
+ if (!list_member_ptr(root->group_pathkeys, pathkey))
+ break;
+
+ n++;
+ }
+
+ return n;
+}
+
/*
* truncate_useless_pathkeys
* Shorten the given pathkey list to just the useful pathkeys.
@@ -1951,6 +2548,9 @@ truncate_useless_pathkeys(PlannerInfo *root,
nuseful = pathkeys_useful_for_merging(root, rel, pathkeys);
nuseful2 = pathkeys_useful_for_ordering(root, pathkeys);
+ if (nuseful2 > nuseful)
+ nuseful = nuseful2;
+ nuseful2 = pathkeys_useful_for_grouping(root, pathkeys);
if (nuseful2 > nuseful)
nuseful = nuseful2;
@@ -1986,6 +2586,8 @@ has_useful_pathkeys(PlannerInfo *root, RelOptInfo *rel)
{
if (rel->joininfo != NIL || rel->has_eclass_joins)
return true; /* might be able to use pathkeys for merging */
+ if (root->group_pathkeys != NIL)
+ return true; /* might be able to use pathkeys for grouping */
if (root->query_pathkeys != NIL)
return true; /* might be able to use them for ordering */
return false; /* definitely useless */
diff --git a/src/backend/optimizer/plan/planner.c b/src/backend/optimizer/plan/planner.c
index 44efb1f4eb..e4d3273f80 100644
--- a/src/backend/optimizer/plan/planner.c
+++ b/src/backend/optimizer/plan/planner.c
@@ -2810,8 +2810,9 @@ remove_useless_groupby_columns(PlannerInfo *root)
*
* In principle it might be interesting to consider other orderings of the
* GROUP BY elements, which could match the sort ordering of other
- * possible plans (eg an indexscan) and thereby reduce cost. We don't
- * bother with that, though. Hashed grouping will frequently win anyway.
+ * possible plans (eg an indexscan) and thereby reduce cost. However, we
+ * don't yet have sufficient information to do that here, so that's left until
+ * later in planning. See get_useful_group_keys_orderings().
*
* Note: we need no comparable processing of the distinctClause because
* the parser already enforced that that matches ORDER BY.
@@ -3468,6 +3469,7 @@ standard_qp_callback(PlannerInfo *root, void *extra)
tlist,
true,
&sortable);
+
if (!sortable)
{
/* Can't sort; no point in considering aggregate ordering either */
@@ -6842,90 +6844,108 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
+
+ Assert(list_length(pathkey_orderings) > 0);
- if (!is_sorted)
+ foreach(lc2, pathkey_orderings)
{
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially sorted
- * already (no need to deal with paths which have presorted
- * keys when incremental sort is disabled unless it's the
- * cheapest input path).
- */
- if (path != cheapest_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ bool is_sorted;
+ int presorted_keys;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- /*
- * We've no need to consider both a sort and incremental sort.
- * We'll just do a sort if there are no presorted keys and an
- * incremental sort when there are presorted keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
- /* Now decide what to stick atop it */
- if (parse->groupingSets)
- {
- consider_groupingsets_paths(root, grouped_rel,
- path, true, can_hash,
- gd, agg_costs, dNumGroups);
- }
- else if (parse->hasAggs)
- {
- /*
- * We have aggregation, possibly with plain GROUP BY. Make an
- * AggPath.
- */
- add_path(grouped_rel, (Path *)
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (!is_sorted)
+ {
+ /*
+ * Try at least sorting the cheapest path and also try
+ * incrementally sorting any path which is partially sorted
+ * already (no need to deal with paths which have presorted
+ * keys when incremental sort is disabled unless it's the
+ * cheapest input path).
+ */
+ if (path != cheapest_path &&
+ (presorted_keys == 0 || !enable_incremental_sort))
+ continue;
+
+ /*
+ * We've no need to consider both a sort and incremental sort.
+ * We'll just do a sort if there are no presorted keys and an
+ * incremental sort when there are presorted keys.
+ */
+ if (presorted_keys == 0 || !enable_incremental_sort)
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ else
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+ }
+
+ /* Now decide what to stick atop it */
+ if (parse->groupingSets)
+ {
+ consider_groupingsets_paths(root, grouped_rel,
+ path, true, can_hash,
+ gd, agg_costs, dNumGroups);
+ }
+ else if (parse->hasAggs)
+ {
+ /*
+ * We have aggregation, possibly with plain GROUP BY. Make an
+ * AggPath.
+ */
+ add_path(grouped_rel, (Path *)
create_agg_path(root,
grouped_rel,
path,
grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_SIMPLE,
- root->processed_groupClause,
+ info->clauses,
havingQual,
agg_costs,
dNumGroups));
- }
- else if (parse->groupClause)
- {
- /*
- * We have GROUP BY without aggregation or grouping sets. Make
- * a GroupPath.
- */
- add_path(grouped_rel, (Path *)
+ }
+ else if (parse->groupClause)
+ {
+ /*
+ * We have GROUP BY without aggregation or grouping sets. Make
+ * a GroupPath.
+ */
+ add_path(grouped_rel, (Path *)
create_group_path(root,
grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
havingQual,
dNumGroups));
- }
- else
- {
- /* Other cases should have been handled above */
- Assert(false);
+ }
+ else
+ {
+ /* Other cases should have been handled above */
+ Assert(false);
+ }
}
}
@@ -6937,69 +6957,84 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
{
foreach(lc, partially_grouped_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
+ List *pathkey_orderings = NIL;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
- if (!is_sorted)
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach(lc2, pathkey_orderings)
{
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially
- * sorted already (no need to deal with paths which have
- * presorted keys when incremental sort is disabled unless
- * it's the cheapest input path).
- */
- if (path != partially_grouped_rel->cheapest_total_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ bool is_sorted;
+ int presorted_keys;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- /*
- * We've no need to consider both a sort and incremental
- * sort. We'll just do a sort if there are no pre-sorted
- * keys and an incremental sort when there are presorted
- * keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
+ path->pathkeys,
+ &presorted_keys);
- if (parse->hasAggs)
- add_path(grouped_rel, (Path *)
+ if (!is_sorted)
+ {
+ /*
+ * Try at least sorting the cheapest path and also try
+ * incrementally sorting any path which is partially
+ * sorted already (no need to deal with paths which have
+ * presorted keys when incremental sort is disabled unless
+ * it's the cheapest input path).
+ */
+ if (path != partially_grouped_rel->cheapest_total_path &&
+ (presorted_keys == 0 || !enable_incremental_sort))
+ continue;
+
+ /*
+ * We've no need to consider both a sort and incremental
+ * sort. We'll just do a sort if there are no pre-sorted
+ * keys and an incremental sort when there are presorted
+ * keys.
+ */
+ if (presorted_keys == 0 || !enable_incremental_sort)
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ else
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_path(grouped_rel, (Path *)
create_agg_path(root,
grouped_rel,
path,
grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_FINAL_DESERIAL,
- root->processed_groupClause,
+ info->clauses,
havingQual,
agg_final_costs,
dNumGroups));
- else
- add_path(grouped_rel, (Path *)
+ else
+ add_path(grouped_rel, (Path *)
create_group_path(root,
grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
havingQual,
dNumGroups));
+ }
}
}
}
@@ -7202,66 +7237,85 @@ create_partial_grouping_paths(PlannerInfo *root,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
- if (!is_sorted)
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach(lc2, pathkey_orderings)
{
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially sorted
- * already (no need to deal with paths which have presorted
- * keys when incremental sort is disabled unless it's the
- * cheapest input path).
- */
- if (path != cheapest_total_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ bool is_sorted;
+ int presorted_keys;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- /*
- * We've no need to consider both a sort and incremental sort.
- * We'll just do a sort if there are no presorted keys and an
- * incremental sort when there are presorted keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
- if (parse->hasAggs)
- add_path(partially_grouped_rel, (Path *)
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+ if (!is_sorted)
+ {
+ /*
+ * Try at least sorting the cheapest path and also try
+ * incrementally sorting any path which is partially sorted
+ * already (no need to deal with paths which have presorted
+ * keys when incremental sort is disabled unless it's the
+ * cheapest input path).
+ */
+ if (path != cheapest_total_path &&
+ (presorted_keys == 0 || !enable_incremental_sort))
+ continue;
+
+ /*
+ * We've no need to consider both a sort and incremental sort.
+ * We'll just do a sort if there are no presorted keys and an
+ * incremental sort when there are presorted keys.
+ */
+ if (presorted_keys == 0 || !enable_incremental_sort)
+ path = (Path *) create_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ else
+ path = (Path *) create_incremental_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_path(partially_grouped_rel, (Path *)
create_agg_path(root,
partially_grouped_rel,
path,
partially_grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_INITIAL_SERIAL,
- root->processed_groupClause,
+ info->clauses,
NIL,
agg_partial_costs,
dNumPartialGroups));
- else
- add_path(partially_grouped_rel, (Path *)
- create_group_path(root,
- partially_grouped_rel,
- path,
- root->processed_groupClause,
- NIL,
- dNumPartialGroups));
+ else
+ add_path(partially_grouped_rel, (Path *)
+ create_group_path(root,
+ partially_grouped_rel,
+ path,
+ info->clauses,
+ NIL,
+ dNumPartialGroups));
+ }
}
}
@@ -7270,67 +7324,86 @@ create_partial_grouping_paths(PlannerInfo *root,
/* Similar to above logic, but for partial paths. */
foreach(lc, input_rel->partial_pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
- if (!is_sorted)
- {
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially sorted
- * already (no need to deal with paths which have presorted
- * keys when incremental sort is disabled unless it's the
- * cheapest input path).
- */
- if (path != cheapest_partial_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ Assert(list_length(pathkey_orderings) > 0);
- /*
- * We've no need to consider both a sort and incremental sort.
- * We'll just do a sort if there are no presorted keys and an
- * incremental sort when there are presorted keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ /* process all potentially interesting grouping reorderings */
+ foreach(lc2, pathkey_orderings)
+ {
+ bool is_sorted;
+ int presorted_keys;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
- if (parse->hasAggs)
- add_partial_path(partially_grouped_rel, (Path *)
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (!is_sorted)
+ {
+ /*
+ * Try at least sorting the cheapest path and also try
+ * incrementally sorting any path which is partially sorted
+ * already (no need to deal with paths which have presorted
+ * keys when incremental sort is disabled unless it's the
+ * cheapest input path).
+ */
+ if (path != cheapest_partial_path &&
+ (presorted_keys == 0 || !enable_incremental_sort))
+ continue;
+
+ /*
+ * We've no need to consider both a sort and incremental sort.
+ * We'll just do a sort if there are no presorted keys and an
+ * incremental sort when there are presorted keys.
+ */
+ if (presorted_keys == 0 || !enable_incremental_sort)
+ path = (Path *) create_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ else
+ path = (Path *) create_incremental_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_partial_path(partially_grouped_rel, (Path *)
create_agg_path(root,
partially_grouped_rel,
path,
partially_grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_INITIAL_SERIAL,
- root->processed_groupClause,
+ info->clauses,
NIL,
agg_partial_costs,
dNumPartialPartialGroups));
- else
- add_partial_path(partially_grouped_rel, (Path *)
+ else
+ add_partial_path(partially_grouped_rel, (Path *)
create_group_path(root,
partially_grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
NIL,
dNumPartialPartialGroups));
+ }
}
}
@@ -7444,6 +7517,8 @@ gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
* We can also skip the entire loop when we only have a single-item
* group_pathkeys because then we can't possibly have a presorted prefix
* of the list without having the list be fully sorted.
+ *
+ * XXX Shouldn't this also consider the group-key-reordering?
*/
if (!enable_incremental_sort || list_length(root->group_pathkeys) == 1)
return;
diff --git a/src/backend/optimizer/util/pathnode.c b/src/backend/optimizer/util/pathnode.c
index f123fcb41e..c72ea08447 100644
--- a/src/backend/optimizer/util/pathnode.c
+++ b/src/backend/optimizer/util/pathnode.c
@@ -1346,12 +1346,12 @@ create_append_path(PlannerInfo *root,
pathnode->path.total_cost = child->total_cost;
}
else
- cost_append(pathnode);
+ cost_append(pathnode, root);
/* Must do this last, else cost_append complains */
pathnode->path.pathkeys = child->pathkeys;
}
else
- cost_append(pathnode);
+ cost_append(pathnode, root);
/* If the caller provided a row estimate, override the computed value. */
if (rows >= 0)
diff --git a/src/backend/utils/adt/selfuncs.c b/src/backend/utils/adt/selfuncs.c
index c4fcd0076e..cb7e67099e 100644
--- a/src/backend/utils/adt/selfuncs.c
+++ b/src/backend/utils/adt/selfuncs.c
@@ -3386,11 +3386,28 @@ double
estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
List **pgset, EstimationInfo *estinfo)
{
- List *varinfos = NIL;
+ return estimate_num_groups_incremental(root, groupExprs,
+ input_rows, pgset, estinfo,
+ NULL, 0);
+}
+
+/*
+ * estimate_num_groups_incremental
+ * An estimate_num_groups variant, optimized for cases that are adding the
+ * expressions incrementally (e.g. one by one).
+ */
+double
+estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
+ double input_rows,
+ List **pgset, EstimationInfo *estinfo,
+ List **cache_varinfos, int prevNExprs)
+{
+ List *varinfos = (cache_varinfos) ? *cache_varinfos : NIL;
double srf_multiplier = 1.0;
double numdistinct;
ListCell *l;
- int i;
+ int i,
+ j;
/* Zero the estinfo output parameter, if non-NULL */
if (estinfo != NULL)
@@ -3421,7 +3438,7 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
*/
numdistinct = 1.0;
- i = 0;
+ i = j = 0;
foreach(l, groupExprs)
{
Node *groupexpr = (Node *) lfirst(l);
@@ -3430,6 +3447,14 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
List *varshere;
ListCell *l2;
+ /* was done on previous call */
+ if (cache_varinfos && j++ < prevNExprs)
+ {
+ if (pgset)
+ i++; /* to keep in sync with lines below */
+ continue;
+ }
+
/* is expression in this grouping set? */
if (pgset && !list_member_int(*pgset, i++))
continue;
@@ -3499,7 +3524,11 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
if (varshere == NIL)
{
if (contain_volatile_functions(groupexpr))
+ {
+ if (cache_varinfos)
+ *cache_varinfos = varinfos;
return input_rows;
+ }
continue;
}
@@ -3516,6 +3545,9 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
}
}
+ if (cache_varinfos)
+ *cache_varinfos = varinfos;
+
/*
* If now no Vars, we must have an all-constant or all-boolean GROUP BY
* list.
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index f9dba43b8c..1c10ab5a7d 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -1025,6 +1025,16 @@ struct config_bool ConfigureNamesBool[] =
true,
NULL, NULL, NULL
},
+ {
+ {"enable_group_by_reordering", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("Enables reordering of GROUP BY keys."),
+ NULL,
+ GUC_EXPLAIN
+ },
+ &enable_group_by_reordering,
+ true,
+ NULL, NULL, NULL
+ },
{
{"geqo", PGC_USERSET, QUERY_TUNING_GEQO,
gettext_noop("Enables genetic query optimization."),
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index c768af9a73..671b88fc60 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -393,6 +393,7 @@
#enable_seqscan = on
#enable_sort = on
#enable_tidscan = on
+#enable_group_by_reordering = on
# - Planner Cost Constants -
diff --git a/src/include/nodes/pathnodes.h b/src/include/nodes/pathnodes.h
index c17b53f7ad..eeeafd8e76 100644
--- a/src/include/nodes/pathnodes.h
+++ b/src/include/nodes/pathnodes.h
@@ -1456,6 +1456,16 @@ typedef struct PathKey
bool pk_nulls_first; /* do NULLs come before normal values? */
} PathKey;
+/*
+ * Combines information about pathkeys and the associated clauses.
+ */
+typedef struct PathKeyInfo
+{
+ NodeTag type;
+ List *pathkeys;
+ List *clauses;
+} PathKeyInfo;
+
/*
* VolatileFunctionStatus -- allows nodes to cache their
* contain_volatile_functions properties. VOLATILITY_UNKNOWN means not yet
diff --git a/src/include/optimizer/cost.h b/src/include/optimizer/cost.h
index 6cf49705d3..28175d3472 100644
--- a/src/include/optimizer/cost.h
+++ b/src/include/optimizer/cost.h
@@ -116,7 +116,9 @@ extern void cost_incremental_sort(Path *path,
Cost input_startup_cost, Cost input_total_cost,
double input_tuples, int width, Cost comparison_cost, int sort_mem,
double limit_tuples);
-extern void cost_append(AppendPath *apath);
+extern Cost cost_sort_estimate(PlannerInfo *root, List *pathkeys,
+ int nPresortedKeys, double tuples);
+extern void cost_append(AppendPath *path, PlannerInfo *root);
extern void cost_merge_append(Path *path, PlannerInfo *root,
List *pathkeys, int n_streams,
Cost input_startup_cost, Cost input_total_cost,
diff --git a/src/include/optimizer/paths.h b/src/include/optimizer/paths.h
index 50bc3b503a..e4ff16d0c4 100644
--- a/src/include/optimizer/paths.h
+++ b/src/include/optimizer/paths.h
@@ -24,6 +24,7 @@ extern PGDLLIMPORT bool enable_geqo;
extern PGDLLIMPORT int geqo_threshold;
extern PGDLLIMPORT int min_parallel_table_scan_size;
extern PGDLLIMPORT int min_parallel_index_scan_size;
+extern PGDLLIMPORT bool enable_group_by_reordering;
/* Hook for plugins to get control in set_rel_pathlist() */
typedef void (*set_rel_pathlist_hook_type) (PlannerInfo *root,
@@ -205,6 +206,12 @@ typedef enum
extern PathKeysComparison compare_pathkeys(List *keys1, List *keys2);
extern bool pathkeys_contained_in(List *keys1, List *keys2);
extern bool pathkeys_count_contained_in(List *keys1, List *keys2, int *n_common);
+extern int group_keys_reorder_by_pathkeys(List *pathkeys,
+ List **group_pathkeys,
+ List **group_clauses,
+ int num_groupby_pathkeys);
+extern List *get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys);
extern Path *get_cheapest_path_for_pathkeys(List *paths, List *pathkeys,
Relids required_outer,
CostSelector cost_criterion,
diff --git a/src/include/utils/selfuncs.h b/src/include/utils/selfuncs.h
index 2f76c473db..2b21c06cc9 100644
--- a/src/include/utils/selfuncs.h
+++ b/src/include/utils/selfuncs.h
@@ -214,6 +214,11 @@ extern double estimate_num_groups(PlannerInfo *root, List *groupExprs,
double input_rows, List **pgset,
EstimationInfo *estinfo);
+extern double estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
+ double input_rows, List **pgset,
+ EstimationInfo *estinfo,
+ List **cache_varinfos, int prevNExprs);
+
extern void estimate_hash_bucket_stats(PlannerInfo *root,
Node *hashkey, double nbuckets,
Selectivity *mcv_freq,
diff --git a/src/test/regress/expected/aggregates.out b/src/test/regress/expected/aggregates.out
index d8271da4d1..205be28b3d 100644
--- a/src/test/regress/expected/aggregates.out
+++ b/src/test/regress/expected/aggregates.out
@@ -1234,7 +1234,8 @@ explain (costs off)
select distinct min(f1), max(f1) from minmaxtest;
QUERY PLAN
---------------------------------------------------------------------------------------------
- Unique
+ HashAggregate
+ Group Key: $0, $1
InitPlan 1 (returns $0)
-> Limit
-> Merge Append
@@ -1257,10 +1258,8 @@ explain (costs off)
-> Index Only Scan using minmaxtest2i on minmaxtest2 minmaxtest_8
Index Cond: (f1 IS NOT NULL)
-> Index Only Scan Backward using minmaxtest3i on minmaxtest3 minmaxtest_9
- -> Sort
- Sort Key: ($0), ($1)
- -> Result
-(26 rows)
+ -> Result
+(25 rows)
select distinct min(f1), max(f1) from minmaxtest;
min | max
@@ -2715,6 +2714,241 @@ SELECT balk(hundred) FROM tenk1;
(1 row)
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+SELECT
+ i AS id,
+ i/2 AS p,
+ format('%60s', i%2) AS v,
+ i/4 AS c,
+ i/8 AS d,
+ (random() * (10000/8))::int as e --the same as d but no correlation with p
+ INTO btg
+FROM
+ generate_series(1, 10000) i;
+VACUUM btg;
+ANALYZE btg;
+-- GROUP BY optimization by reorder columns by frequency
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Sort
+ Sort Key: p, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Sort
+ Sort Key: p, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, c, v
+ -> Sort
+ Sort Key: p, c, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: v, p, c
+ -> Sort
+ Sort Key: v, p, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: p, d, v, c
+ -> Sort
+ Sort Key: p, d, v, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: v, p, d, c
+ -> Sort
+ Sort Key: v, p, d, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: p, v, d, c
+ -> Sort
+ Sort Key: p, v, d, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, d, e
+ -> Sort
+ Sort Key: p, d, e
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, e, d
+ -> Sort
+ Sort Key: p, e, d
+ -> Seq Scan on btg
+(5 rows)
+
+CREATE STATISTICS btg_dep ON d, e, p FROM btg;
+ANALYZE btg;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, d, e
+ -> Sort
+ Sort Key: p, d, e
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, e, d
+ -> Sort
+ Sort Key: p, e, d
+ -> Seq Scan on btg
+(5 rows)
+
+-- GROUP BY optimization by reorder columns by index scan
+CREATE INDEX ON btg(p, v);
+SET enable_seqscan=off;
+SET enable_bitmapscan=off;
+VACUUM btg;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v ORDER BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p ORDER BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, c, v
+ -> Incremental Sort
+ Sort Key: p, c, v
+ Presorted Key: p
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY p, v;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, v, c
+ -> Incremental Sort
+ Sort Key: p, v, c
+ Presorted Key: p, v
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, c, v, d
+ -> Incremental Sort
+ Sort Key: p, c, v, d
+ Presorted Key: p
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d ORDER BY p, v;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, v, c, d
+ -> Incremental Sort
+ Sort Key: p, v, c, d
+ Presorted Key: p, v
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+DROP TABLE btg;
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+RESET enable_seqscan;
+RESET enable_bitmapscan;
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/expected/incremental_sort.out b/src/test/regress/expected/incremental_sort.out
index 7fdb685313..b8a0968b06 100644
--- a/src/test/regress/expected/incremental_sort.out
+++ b/src/test/regress/expected/incremental_sort.out
@@ -1426,7 +1426,7 @@ set parallel_setup_cost = 0;
set parallel_tuple_cost = 0;
set max_parallel_workers_per_gather = 2;
create table t (a int, b int, c int);
-insert into t select mod(i,10),mod(i,10),i from generate_series(1,10000) s(i);
+insert into t select mod(i,10),mod(i,10),i from generate_series(1,60000) s(i);
create index on t (a);
analyze t;
set enable_incremental_sort = off;
diff --git a/src/test/regress/expected/join.out b/src/test/regress/expected/join.out
index 9b8638f286..27684bfce2 100644
--- a/src/test/regress/expected/join.out
+++ b/src/test/regress/expected/join.out
@@ -1984,8 +1984,8 @@ USING (name);
------+----+----
bb | 12 | 13
cc | 22 | 23
- dd | | 33
ee | 42 |
+ dd | | 33
(4 rows)
-- Cases with non-nullable expressions in subquery results;
@@ -2019,8 +2019,8 @@ NATURAL FULL JOIN
------+------+------+------+------
bb | 12 | 2 | 13 | 3
cc | 22 | 2 | 23 | 3
- dd | | | 33 | 3
ee | 42 | 2 | |
+ dd | | | 33 | 3
(4 rows)
SELECT * FROM
@@ -5645,18 +5645,20 @@ select d.* from d left join (select * from b group by b.id, b.c_id) s
explain (costs off)
select d.* from d left join (select distinct * from b) s
on d.a = s.id;
- QUERY PLAN
---------------------------------------
- Merge Right Join
- Merge Cond: (b.id = d.a)
- -> Unique
- -> Sort
- Sort Key: b.id, b.c_id
- -> Seq Scan on b
+ QUERY PLAN
+---------------------------------------------
+ Merge Left Join
+ Merge Cond: (d.a = s.id)
-> Sort
Sort Key: d.a
-> Seq Scan on d
-(9 rows)
+ -> Sort
+ Sort Key: s.id
+ -> Subquery Scan on s
+ -> HashAggregate
+ Group Key: b.id, b.c_id
+ -> Seq Scan on b
+(11 rows)
-- join removal is not possible here
explain (costs off)
@@ -7597,44 +7599,39 @@ select * from j1 natural join j2;
explain (verbose, costs off)
select * from j1
inner join (select distinct id from j3) j3 on j1.id = j3.id;
- QUERY PLAN
------------------------------------------
+ QUERY PLAN
+-----------------------------------
Nested Loop
Output: j1.id, j3.id
Inner Unique: true
Join Filter: (j1.id = j3.id)
- -> Unique
+ -> HashAggregate
Output: j3.id
- -> Sort
+ Group Key: j3.id
+ -> Seq Scan on public.j3
Output: j3.id
- Sort Key: j3.id
- -> Seq Scan on public.j3
- Output: j3.id
-> Seq Scan on public.j1
Output: j1.id
-(13 rows)
+(11 rows)
-- ensure group by clause allows the inner to become unique
explain (verbose, costs off)
select * from j1
inner join (select id from j3 group by id) j3 on j1.id = j3.id;
- QUERY PLAN
------------------------------------------
+ QUERY PLAN
+-----------------------------------
Nested Loop
Output: j1.id, j3.id
Inner Unique: true
Join Filter: (j1.id = j3.id)
- -> Group
+ -> HashAggregate
Output: j3.id
Group Key: j3.id
- -> Sort
+ -> Seq Scan on public.j3
Output: j3.id
- Sort Key: j3.id
- -> Seq Scan on public.j3
- Output: j3.id
-> Seq Scan on public.j1
Output: j1.id
-(14 rows)
+(11 rows)
drop table j1;
drop table j2;
diff --git a/src/test/regress/expected/merge.out b/src/test/regress/expected/merge.out
index 133d42117c..f7c4cc4833 100644
--- a/src/test/regress/expected/merge.out
+++ b/src/test/regress/expected/merge.out
@@ -1459,18 +1459,15 @@ WHEN MATCHED AND t.a < 10 THEN
explain_merge
--------------------------------------------------------------------
Merge on ex_mtarget t (actual rows=0 loops=1)
- -> Merge Join (actual rows=0 loops=1)
- Merge Cond: (t.a = s.a)
- -> Sort (actual rows=0 loops=1)
- Sort Key: t.a
- Sort Method: quicksort Memory: xxx
+ -> Hash Join (actual rows=0 loops=1)
+ Hash Cond: (s.a = t.a)
+ -> Seq Scan on ex_msource s (actual rows=1 loops=1)
+ -> Hash (actual rows=0 loops=1)
+ Buckets: xxx Batches: xxx Memory Usage: xxx
-> Seq Scan on ex_mtarget t (actual rows=0 loops=1)
Filter: (a < '-1000'::integer)
Rows Removed by Filter: 54
- -> Sort (never executed)
- Sort Key: s.a
- -> Seq Scan on ex_msource s (never executed)
-(12 rows)
+(9 rows)
DROP TABLE ex_msource, ex_mtarget;
DROP FUNCTION explain_merge(text);
diff --git a/src/test/regress/expected/partition_aggregate.out b/src/test/regress/expected/partition_aggregate.out
index 1b900fddf8..800114dfde 100644
--- a/src/test/regress/expected/partition_aggregate.out
+++ b/src/test/regress/expected/partition_aggregate.out
@@ -948,12 +948,12 @@ SET parallel_setup_cost = 0;
-- is not partial agg safe.
EXPLAIN (COSTS OFF)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3 ORDER BY 1, 2, 3;
- QUERY PLAN
---------------------------------------------------------------------------------------
- Sort
- Sort Key: pagg_tab_ml.a, (sum(pagg_tab_ml.b)), (array_agg(DISTINCT pagg_tab_ml.c))
- -> Gather
- Workers Planned: 2
+ QUERY PLAN
+--------------------------------------------------------------------------------------------
+ Gather Merge
+ Workers Planned: 2
+ -> Sort
+ Sort Key: pagg_tab_ml.a, (sum(pagg_tab_ml.b)), (array_agg(DISTINCT pagg_tab_ml.c))
-> Parallel Append
-> GroupAggregate
Group Key: pagg_tab_ml.a
@@ -1380,28 +1380,26 @@ SELECT x, sum(y), avg(y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) <
-- When GROUP BY clause does not match; partial aggregation is performed for each partition.
EXPLAIN (COSTS OFF)
SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) < 12 ORDER BY 1, 2, 3;
- QUERY PLAN
--------------------------------------------------------------------------------------------
+ QUERY PLAN
+-------------------------------------------------------------------------------------
Sort
Sort Key: pagg_tab_para.y, (sum(pagg_tab_para.x)), (avg(pagg_tab_para.x))
- -> Finalize GroupAggregate
+ -> Finalize HashAggregate
Group Key: pagg_tab_para.y
Filter: (avg(pagg_tab_para.x) < '12'::numeric)
- -> Gather Merge
+ -> Gather
Workers Planned: 2
- -> Sort
- Sort Key: pagg_tab_para.y
- -> Parallel Append
- -> Partial HashAggregate
- Group Key: pagg_tab_para.y
- -> Parallel Seq Scan on pagg_tab_para_p1 pagg_tab_para
- -> Partial HashAggregate
- Group Key: pagg_tab_para_1.y
- -> Parallel Seq Scan on pagg_tab_para_p2 pagg_tab_para_1
- -> Partial HashAggregate
- Group Key: pagg_tab_para_2.y
- -> Parallel Seq Scan on pagg_tab_para_p3 pagg_tab_para_2
-(19 rows)
+ -> Parallel Append
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para.y
+ -> Parallel Seq Scan on pagg_tab_para_p1 pagg_tab_para
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para_1.y
+ -> Parallel Seq Scan on pagg_tab_para_p2 pagg_tab_para_1
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para_2.y
+ -> Parallel Seq Scan on pagg_tab_para_p3 pagg_tab_para_2
+(17 rows)
SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) < 12 ORDER BY 1, 2, 3;
y | sum | avg | count
diff --git a/src/test/regress/expected/partition_join.out b/src/test/regress/expected/partition_join.out
index 762cc8e53b..a1add421f5 100644
--- a/src/test/regress/expected/partition_join.out
+++ b/src/test/regress/expected/partition_join.out
@@ -466,52 +466,41 @@ EXPLAIN (COSTS OFF)
SELECT a, b FROM prt1 FULL JOIN prt2 p2(b,a,c) USING(a,b)
WHERE a BETWEEN 490 AND 510
GROUP BY 1, 2 ORDER BY 1, 2;
- QUERY PLAN
------------------------------------------------------------------------------------------------------------------
+ QUERY PLAN
+-----------------------------------------------------------------------------------------------------------
Group
Group Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Merge Append
+ -> Sort
Sort Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Group
- Group Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Sort
- Sort Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Merge Full Join
- Merge Cond: ((prt1.a = p2.a) AND (prt1.b = p2.b))
- Filter: ((COALESCE(prt1.a, p2.a) >= 490) AND (COALESCE(prt1.a, p2.a) <= 510))
- -> Sort
- Sort Key: prt1.a, prt1.b
- -> Seq Scan on prt1_p1 prt1
- -> Sort
- Sort Key: p2.a, p2.b
- -> Seq Scan on prt2_p1 p2
- -> Group
- Group Key: (COALESCE(prt1_1.a, p2_1.a)), (COALESCE(prt1_1.b, p2_1.b))
- -> Sort
- Sort Key: (COALESCE(prt1_1.a, p2_1.a)), (COALESCE(prt1_1.b, p2_1.b))
- -> Merge Full Join
- Merge Cond: ((prt1_1.a = p2_1.a) AND (prt1_1.b = p2_1.b))
- Filter: ((COALESCE(prt1_1.a, p2_1.a) >= 490) AND (COALESCE(prt1_1.a, p2_1.a) <= 510))
- -> Sort
- Sort Key: prt1_1.a, prt1_1.b
- -> Seq Scan on prt1_p2 prt1_1
- -> Sort
- Sort Key: p2_1.a, p2_1.b
- -> Seq Scan on prt2_p2 p2_1
- -> Group
- Group Key: (COALESCE(prt1_2.a, p2_2.a)), (COALESCE(prt1_2.b, p2_2.b))
- -> Sort
- Sort Key: (COALESCE(prt1_2.a, p2_2.a)), (COALESCE(prt1_2.b, p2_2.b))
- -> Merge Full Join
- Merge Cond: ((prt1_2.a = p2_2.a) AND (prt1_2.b = p2_2.b))
- Filter: ((COALESCE(prt1_2.a, p2_2.a) >= 490) AND (COALESCE(prt1_2.a, p2_2.a) <= 510))
- -> Sort
- Sort Key: prt1_2.a, prt1_2.b
- -> Seq Scan on prt1_p3 prt1_2
- -> Sort
- Sort Key: p2_2.a, p2_2.b
- -> Seq Scan on prt2_p3 p2_2
-(43 rows)
+ -> Append
+ -> Merge Full Join
+ Merge Cond: ((prt1_1.a = p2_1.a) AND (prt1_1.b = p2_1.b))
+ Filter: ((COALESCE(prt1_1.a, p2_1.a) >= 490) AND (COALESCE(prt1_1.a, p2_1.a) <= 510))
+ -> Sort
+ Sort Key: prt1_1.a, prt1_1.b
+ -> Seq Scan on prt1_p1 prt1_1
+ -> Sort
+ Sort Key: p2_1.a, p2_1.b
+ -> Seq Scan on prt2_p1 p2_1
+ -> Merge Full Join
+ Merge Cond: ((prt1_2.a = p2_2.a) AND (prt1_2.b = p2_2.b))
+ Filter: ((COALESCE(prt1_2.a, p2_2.a) >= 490) AND (COALESCE(prt1_2.a, p2_2.a) <= 510))
+ -> Sort
+ Sort Key: prt1_2.a, prt1_2.b
+ -> Seq Scan on prt1_p2 prt1_2
+ -> Sort
+ Sort Key: p2_2.a, p2_2.b
+ -> Seq Scan on prt2_p2 p2_2
+ -> Merge Full Join
+ Merge Cond: ((prt1_3.a = p2_3.a) AND (prt1_3.b = p2_3.b))
+ Filter: ((COALESCE(prt1_3.a, p2_3.a) >= 490) AND (COALESCE(prt1_3.a, p2_3.a) <= 510))
+ -> Sort
+ Sort Key: prt1_3.a, prt1_3.b
+ -> Seq Scan on prt1_p3 prt1_3
+ -> Sort
+ Sort Key: p2_3.a, p2_3.b
+ -> Seq Scan on prt2_p3 p2_3
+(32 rows)
SELECT a, b FROM prt1 FULL JOIN prt2 p2(b,a,c) USING(a,b)
WHERE a BETWEEN 490 AND 510
diff --git a/src/test/regress/expected/sysviews.out b/src/test/regress/expected/sysviews.out
index 001c6e7eb9..d5e9a35096 100644
--- a/src/test/regress/expected/sysviews.out
+++ b/src/test/regress/expected/sysviews.out
@@ -114,6 +114,7 @@ select name, setting from pg_settings where name like 'enable%';
enable_async_append | on
enable_bitmapscan | on
enable_gathermerge | on
+ enable_group_by_reordering | on
enable_hashagg | on
enable_hashjoin | on
enable_incremental_sort | on
@@ -132,7 +133,7 @@ select name, setting from pg_settings where name like 'enable%';
enable_seqscan | on
enable_sort | on
enable_tidscan | on
-(21 rows)
+(22 rows)
-- Test that the pg_timezone_names and pg_timezone_abbrevs views are
-- more-or-less working. We can't test their contents in any great detail
diff --git a/src/test/regress/expected/union.out b/src/test/regress/expected/union.out
index e2613d6777..220f0c1629 100644
--- a/src/test/regress/expected/union.out
+++ b/src/test/regress/expected/union.out
@@ -1303,24 +1303,22 @@ select distinct q1 from
union all
select distinct * from int8_tbl i82) ss
where q2 = q2;
- QUERY PLAN
-----------------------------------------------------------
- Unique
- -> Merge Append
- Sort Key: "*SELECT* 1".q1
+ QUERY PLAN
+----------------------------------------------------
+ HashAggregate
+ Group Key: "*SELECT* 1".q1
+ -> Append
-> Subquery Scan on "*SELECT* 1"
- -> Unique
- -> Sort
- Sort Key: i81.q1, i81.q2
- -> Seq Scan on int8_tbl i81
- Filter: (q2 IS NOT NULL)
+ -> HashAggregate
+ Group Key: i81.q1, i81.q2
+ -> Seq Scan on int8_tbl i81
+ Filter: (q2 IS NOT NULL)
-> Subquery Scan on "*SELECT* 2"
- -> Unique
- -> Sort
- Sort Key: i82.q1, i82.q2
- -> Seq Scan on int8_tbl i82
- Filter: (q2 IS NOT NULL)
-(15 rows)
+ -> HashAggregate
+ Group Key: i82.q1, i82.q2
+ -> Seq Scan on int8_tbl i82
+ Filter: (q2 IS NOT NULL)
+(13 rows)
select distinct q1 from
(select distinct * from int8_tbl i81
@@ -1339,24 +1337,22 @@ select distinct q1 from
union all
select distinct * from int8_tbl i82) ss
where -q1 = q2;
- QUERY PLAN
---------------------------------------------------------
- Unique
- -> Merge Append
- Sort Key: "*SELECT* 1".q1
+ QUERY PLAN
+--------------------------------------------------
+ HashAggregate
+ Group Key: "*SELECT* 1".q1
+ -> Append
-> Subquery Scan on "*SELECT* 1"
- -> Unique
- -> Sort
- Sort Key: i81.q1, i81.q2
- -> Seq Scan on int8_tbl i81
- Filter: ((- q1) = q2)
+ -> HashAggregate
+ Group Key: i81.q1, i81.q2
+ -> Seq Scan on int8_tbl i81
+ Filter: ((- q1) = q2)
-> Subquery Scan on "*SELECT* 2"
- -> Unique
- -> Sort
- Sort Key: i82.q1, i82.q2
- -> Seq Scan on int8_tbl i82
- Filter: ((- q1) = q2)
-(15 rows)
+ -> HashAggregate
+ Group Key: i82.q1, i82.q2
+ -> Seq Scan on int8_tbl i82
+ Filter: ((- q1) = q2)
+(13 rows)
select distinct q1 from
(select distinct * from int8_tbl i81
diff --git a/src/test/regress/sql/aggregates.sql b/src/test/regress/sql/aggregates.sql
index 75c78be640..5c4e768825 100644
--- a/src/test/regress/sql/aggregates.sql
+++ b/src/test/regress/sql/aggregates.sql
@@ -1172,6 +1172,105 @@ SELECT balk(hundred) FROM tenk1;
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+
+SELECT
+ i AS id,
+ i/2 AS p,
+ format('%60s', i%2) AS v,
+ i/4 AS c,
+ i/8 AS d,
+ (random() * (10000/8))::int as e --the same as d but no correlation with p
+ INTO btg
+FROM
+ generate_series(1, 10000) i;
+
+VACUUM btg;
+ANALYZE btg;
+
+-- GROUP BY optimization by reorder columns by frequency
+
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+
+CREATE STATISTICS btg_dep ON d, e, p FROM btg;
+ANALYZE btg;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+
+
+-- GROUP BY optimization by reorder columns by index scan
+
+CREATE INDEX ON btg(p, v);
+SET enable_seqscan=off;
+SET enable_bitmapscan=off;
+VACUUM btg;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d ORDER BY p, v;
+
+DROP TABLE btg;
+
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+RESET enable_seqscan;
+RESET enable_bitmapscan;
+
+
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/sql/incremental_sort.sql b/src/test/regress/sql/incremental_sort.sql
index ab471bdfff..70f311a6f2 100644
--- a/src/test/regress/sql/incremental_sort.sql
+++ b/src/test/regress/sql/incremental_sort.sql
@@ -208,7 +208,7 @@ set parallel_tuple_cost = 0;
set max_parallel_workers_per_gather = 2;
create table t (a int, b int, c int);
-insert into t select mod(i,10),mod(i,10),i from generate_series(1,10000) s(i);
+insert into t select mod(i,10),mod(i,10),i from generate_series(1,60000) s(i);
create index on t (a);
analyze t;
--
2.41.0
Attachments:
[text/plain] 0001-When-evaluating-a-query-with-a-multi-column-GROUP-BY-20230720.patch (95.5K, ../../[email protected]/2-0001-When-evaluating-a-query-with-a-multi-column-GROUP-BY-20230720.patch)
download | inline diff:
From 913d55ee887dccfeba360f5f44ed347dd1ba9044 Mon Sep 17 00:00:00 2001
From: "Andrey V. Lepikhov" <[email protected]>
Date: Fri, 14 Jul 2023 10:29:36 +0700
Subject: [PATCH] When evaluating a query with a multi-column GROUP BY clause
using sort, the cost may be heavily dependent on the order in which the keys
are compared when building the groups. Grouping does not imply any ordering,
so we're allowed to compare the keys in arbitrary order, and a Hash Agg
leverages this. But for Group Agg, we simply compared keys in the order as
specified in the query. This commit explores alternative ordering of the
keys, trying to find a cheaper one.
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
In principle, we might generate grouping paths for all permutations of
the keys, and leave the rest to the optimizer. But that might get very
expensive, so we try to pick only a couple interesting orderings based
on both local and global information.
When planning the grouping path, we explore statistics (number of
distinct values, cost of the comparison function) for the keys and
reorder them to minimize comparison costs. Intuitively, it may be better
to perform more expensive comparisons (for complex data types etc.)
last, because maybe the cheaper comparisons will be enough. Similarly,
the higher the cardinality of a key, the lower the probability we’ll
need to compare more keys. The patch generates and costs various
orderings, picking the cheapest ones.
The ordering of group keys may interact with other parts of the query,
some of which may not be known while planning the grouping. E.g. there
may be an explicit ORDER BY clause, or some other ordering-dependent
operation, higher up in the query, and using the same ordering may allow
using either incremental sort or even eliminate the sort entirely.
The patch generates orderings and picks those minimizing the comparison
cost (for various pathkeys), and then adds orderings that might be
useful for operations higher up in the plan (ORDER BY, etc.). Finally,
it always keeps the ordering specified in the query, on the assumption
the user might have additional insights.
This introduces a new GUC enable_group_by_reordering, so that the
optimization may be disabled if needed.
---
.../postgres_fdw/expected/postgres_fdw.out | 15 +-
src/backend/optimizer/path/costsize.c | 363 ++++++++++-
src/backend/optimizer/path/equivclass.c | 13 +-
src/backend/optimizer/path/pathkeys.c | 602 ++++++++++++++++++
src/backend/optimizer/plan/planner.c | 467 ++++++++------
src/backend/optimizer/util/pathnode.c | 4 +-
src/backend/utils/adt/selfuncs.c | 38 +-
src/backend/utils/misc/guc_tables.c | 10 +
src/backend/utils/misc/postgresql.conf.sample | 1 +
src/include/nodes/pathnodes.h | 10 +
src/include/optimizer/cost.h | 4 +-
src/include/optimizer/paths.h | 7 +
src/include/utils/selfuncs.h | 5 +
src/test/regress/expected/aggregates.out | 244 ++++++-
.../regress/expected/incremental_sort.out | 2 +-
src/test/regress/expected/join.out | 51 +-
src/test/regress/expected/merge.out | 15 +-
.../regress/expected/partition_aggregate.out | 44 +-
src/test/regress/expected/partition_join.out | 75 +--
src/test/regress/expected/sysviews.out | 3 +-
src/test/regress/expected/union.out | 60 +-
src/test/regress/sql/aggregates.sql | 99 +++
src/test/regress/sql/incremental_sort.sql | 2 +-
23 files changed, 1760 insertions(+), 374 deletions(-)
diff --git a/contrib/postgres_fdw/expected/postgres_fdw.out b/contrib/postgres_fdw/expected/postgres_fdw.out
index 852b5b4707..3f3c181ac8 100644
--- a/contrib/postgres_fdw/expected/postgres_fdw.out
+++ b/contrib/postgres_fdw/expected/postgres_fdw.out
@@ -9593,13 +9593,16 @@ SELECT t1.a,t2.b,t3.c FROM fprt1 t1 INNER JOIN fprt2 t2 ON (t1.a = t2.b) INNER J
-- left outer join + nullable clause
EXPLAIN (VERBOSE, COSTS OFF)
SELECT t1.a,t2.b,t2.c FROM fprt1 t1 LEFT JOIN (SELECT * FROM fprt2 WHERE a < 10) t2 ON (t1.a = t2.b and t1.b = t2.a) WHERE t1.a < 10 ORDER BY 1,2,3;
- QUERY PLAN
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
- Foreign Scan
+ QUERY PLAN
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+ Sort
Output: t1.a, fprt2.b, fprt2.c
- Relations: (public.ftprt1_p1 t1) LEFT JOIN (public.ftprt2_p1 fprt2)
- Remote SQL: SELECT r5.a, r6.b, r6.c FROM (public.fprt1_p1 r5 LEFT JOIN public.fprt2_p1 r6 ON (((r5.a = r6.b)) AND ((r5.b = r6.a)) AND ((r6.a < 10)))) WHERE ((r5.a < 10)) ORDER BY r5.a ASC NULLS LAST, r6.b ASC NULLS LAST, r6.c ASC NULLS LAST
-(4 rows)
+ Sort Key: t1.a, fprt2.b, fprt2.c
+ -> Foreign Scan
+ Output: t1.a, fprt2.b, fprt2.c
+ Relations: (public.ftprt1_p1 t1) LEFT JOIN (public.ftprt2_p1 fprt2)
+ Remote SQL: SELECT r5.a, r6.b, r6.c FROM (public.fprt1_p1 r5 LEFT JOIN public.fprt2_p1 r6 ON (((r5.a = r6.b)) AND ((r5.b = r6.a)) AND ((r6.a < 10)))) WHERE ((r5.a < 10))
+(7 rows)
SELECT t1.a,t2.b,t2.c FROM fprt1 t1 LEFT JOIN (SELECT * FROM fprt2 WHERE a < 10) t2 ON (t1.a = t2.b and t1.b = t2.a) WHERE t1.a < 10 ORDER BY 1,2,3;
a | b | c
diff --git a/src/backend/optimizer/path/costsize.c b/src/backend/optimizer/path/costsize.c
index ef475d95a1..1e979d57f2 100644
--- a/src/backend/optimizer/path/costsize.c
+++ b/src/backend/optimizer/path/costsize.c
@@ -1816,6 +1816,327 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
rterm->pathtarget->width);
}
+/*
+ * is_fake_var
+ * Workaround for generate_append_tlist() which generates fake Vars with
+ * varno == 0, that will cause a fail of estimate_num_group() call
+ *
+ * XXX Ummm, why would estimate_num_group fail with this?
+ */
+static bool
+is_fake_var(Expr *expr)
+{
+ if (IsA(expr, RelabelType))
+ expr = (Expr *) ((RelabelType *) expr)->arg;
+
+ return (IsA(expr, Var) && ((Var *) expr)->varno == 0);
+}
+
+/*
+ * get_width_cost_multiplier
+ * Returns relative complexity of comparing two values based on its width.
+ * The idea behind is that the comparison becomes more expensive the longer the
+ * value is. Return value is in cpu_operator_cost units.
+ */
+static double
+get_width_cost_multiplier(PlannerInfo *root, Expr *expr)
+{
+ double width = -1.0; /* fake value */
+
+ if (IsA(expr, RelabelType))
+ expr = (Expr *) ((RelabelType *) expr)->arg;
+
+ /* Try to find actual stat in corresponding relation */
+ if (IsA(expr, Var))
+ {
+ Var *var = (Var *) expr;
+
+ if (var->varno > 0 && var->varno < root->simple_rel_array_size)
+ {
+ RelOptInfo *rel = root->simple_rel_array[var->varno];
+
+ if (rel != NULL &&
+ var->varattno >= rel->min_attr &&
+ var->varattno <= rel->max_attr)
+ {
+ int ndx = var->varattno - rel->min_attr;
+
+ if (rel->attr_widths[ndx] > 0)
+ width = rel->attr_widths[ndx];
+ }
+ }
+ }
+
+ /* Didn't find any actual stats, try using type width instead. */
+ if (width < 0.0)
+ {
+ Node *node = (Node *) expr;
+
+ width = get_typavgwidth(exprType(node), exprTypmod(node));
+ }
+
+ /*
+ * Values are passed as Datum type, so comparisons can't be cheaper than
+ * comparing a Datum value.
+ *
+ * FIXME I find this reasoning questionable. We may pass int2, and
+ * comparing it is probably a bit cheaper than comparing a bigint.
+ */
+ if (width <= sizeof(Datum))
+ return 1.0;
+
+ /*
+ * We consider the cost of a comparison not to be directly proportional to
+ * width of the argument, because widths of the arguments could be
+ * slightly different (we only know the average width for the whole
+ * column). So we use log16(width) as an estimate.
+ */
+ return 1.0 + 0.125 * LOG2(width / sizeof(Datum));
+}
+
+/*
+ * compute_cpu_sort_cost
+ * compute CPU cost of sort (i.e. in-memory)
+ *
+ * The main thing we need to calculate to estimate sort CPU costs is the number
+ * of calls to the comparator functions. The difficulty is that for multi-column
+ * sorts there may be different data types involved (for some of which the calls
+ * may be much more expensive). Furthermore, columns may have a very different
+ * number of distinct values - the higher the number, the fewer comparisons will
+ * be needed for the following columns.
+ *
+ * The algorithm is incremental - we add pathkeys one by one, and at each step we
+ * estimate the number of necessary comparisons (based on the number of distinct
+ * groups in the current pathkey prefix and the new pathkey), and the comparison
+ * costs (which is data type specific).
+ *
+ * Estimation of the number of comparisons is based on ideas from:
+ *
+ * "Quicksort Is Optimal", Robert Sedgewick, Jon Bentley, 2002
+ * [https://www.cs.princeton.edu/~rs/talks/QuicksortIsOptimal.pdf]
+ *
+ * In term of that paper, let N - number of tuples, Xi - number of identical
+ * tuples with value Ki, then the estimate of number of comparisons is:
+ *
+ * log(N! / (X1! * X2! * ..)) ~ sum(Xi * log(N/Xi))
+ *
+ * We assume all Xi the same because now we don't have any estimation of
+ * group sizes, we have only know the estimate of number of groups (distinct
+ * values). In that case, formula becomes:
+ *
+ * N * log(NumberOfGroups)
+ *
+ * For multi-column sorts we need to estimate the number of comparisons for
+ * each individual column - for example with columns (c1, c2, ..., ck) we
+ * can estimate that number of comparisons on ck is roughly
+ *
+ * ncomparisons(c1, c2, ..., ck) / ncomparisons(c1, c2, ..., c(k-1))
+ *
+ * Let k be a column number, Gk - number of groups defined by k columns, and Fk
+ * the cost of the comparison is
+ *
+ * N * sum( Fk * log(Gk) )
+ *
+ * Note: We also consider column width, not just the comparator cost.
+ *
+ * NOTE: some callers currently pass NIL for pathkeys because they
+ * can't conveniently supply the sort keys. In this case, it will fallback to
+ * simple comparison cost estimate.
+ */
+static Cost
+compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
+ Cost comparison_cost, double tuples, double output_tuples,
+ bool heapSort)
+{
+ Cost per_tuple_cost = 0.0;
+ ListCell *lc;
+ List *pathkeyExprs = NIL;
+ double tuplesPerPrevGroup = tuples;
+ double totalFuncCost = 1.0;
+ bool has_fake_var = false;
+ int i = 0;
+ Oid prev_datatype = InvalidOid;
+ List *cache_varinfos = NIL;
+
+ /* fallback if pathkeys is unknown */
+ if (list_length(pathkeys) == 0)
+ {
+ /*
+ * If we'll use a bounded heap-sort keeping just K tuples in memory,
+ * for a total number of tuple comparisons of N log2 K; but the
+ * constant factor is a bit higher than for quicksort. Tweak it so
+ * that the cost curve is continuous at the crossover point.
+ */
+ output_tuples = (heapSort) ? 2.0 * output_tuples : tuples;
+ per_tuple_cost += 2.0 * cpu_operator_cost * LOG2(output_tuples);
+
+ /* add cost provided by caller */
+ per_tuple_cost += comparison_cost;
+
+ return per_tuple_cost * tuples;
+ }
+
+ /*
+ * Computing total cost of sorting takes into account the per-column
+ * comparison function cost. We try to compute the needed number of
+ * comparisons per column.
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(lc);
+ EquivalenceMember *em;
+ double nGroups,
+ correctedNGroups;
+ Cost funcCost = 1.0;
+
+ /*
+ * We believe that equivalence members aren't very different, so, to
+ * estimate cost we consider just the first member.
+ */
+ em = (EquivalenceMember *) linitial(pathkey->pk_eclass->ec_members);
+
+ if (em->em_datatype != InvalidOid)
+ {
+ /* do not lookup funcCost if the data type is the same */
+ if (prev_datatype != em->em_datatype)
+ {
+ Oid sortop;
+ QualCost cost;
+
+ sortop = get_opfamily_member(pathkey->pk_opfamily,
+ em->em_datatype, em->em_datatype,
+ pathkey->pk_strategy);
+
+ cost.startup = 0;
+ cost.per_tuple = 0;
+ add_function_cost(root, get_opcode(sortop), NULL, &cost);
+
+ /*
+ * add_function_cost returns the product of cpu_operator_cost
+ * and procost, but we need just procost, co undo that.
+ */
+ funcCost = cost.per_tuple / cpu_operator_cost;
+
+ prev_datatype = em->em_datatype;
+ }
+ }
+
+ /* factor in the width of the values in this column */
+ funcCost *= get_width_cost_multiplier(root, em->em_expr);
+
+ /* now we have per-key cost, so add to the running total */
+ totalFuncCost += funcCost;
+
+ /* remember if we have found a fake Var in pathkeys */
+ has_fake_var |= is_fake_var(em->em_expr);
+ pathkeyExprs = lappend(pathkeyExprs, em->em_expr);
+
+ /*
+ * We need to calculate the number of comparisons for this column,
+ * which requires knowing the group size. So we estimate the number of
+ * groups by calling estimate_num_groups_incremental(), which
+ * estimates the group size for "new" pathkeys.
+ *
+ * Note: estimate_num_groups_incremental does not handle fake Vars, so
+ * use a default estimate otherwise.
+ */
+ if (!has_fake_var)
+ nGroups = estimate_num_groups_incremental(root, pathkeyExprs,
+ tuplesPerPrevGroup, NULL, NULL,
+ &cache_varinfos,
+ list_length(pathkeyExprs) - 1);
+ else if (tuples > 4.0)
+
+ /*
+ * Use geometric mean as estimation if there are no stats.
+ *
+ * We don't use DEFAULT_NUM_DISTINCT here, because that's used for
+ * a single column, but here we're dealing with multiple columns.
+ */
+ nGroups = ceil(2.0 + sqrt(tuples) * (i + 1) / list_length(pathkeys));
+ else
+ nGroups = tuples;
+
+ /*
+ * Presorted keys are not considered in the cost above, but we still
+ * do have to compare them in the qsort comparator. So make sure to
+ * factor in the cost in that case.
+ */
+ if (i >= nPresortedKeys)
+ {
+ if (heapSort)
+ {
+ /*
+ * have to keep at least one group, and a multiple of group
+ * size
+ */
+ correctedNGroups = ceil(output_tuples / tuplesPerPrevGroup);
+ }
+ else
+ /* all groups in the input */
+ correctedNGroups = nGroups;
+
+ correctedNGroups = Max(1.0, ceil(correctedNGroups));
+
+ per_tuple_cost += totalFuncCost * LOG2(correctedNGroups);
+ }
+
+ i++;
+
+ /*
+ * Once we get single-row group, it means tuples in the group are
+ * unique and we can skip all remaining columns.
+ */
+ if (tuplesPerPrevGroup <= nGroups)
+ break;
+
+ /*
+ * Uniform distributions with all groups being of the same size are
+ * the best case, with nice smooth behavior. Real-world distributions
+ * tend not to be uniform, though, and we don't have any reliable
+ * easy-to-use information. As a basic defense against skewed
+ * distributions, we use a 1.5 factor to make the expected group a bit
+ * larger, but we need to be careful not to make the group larger than
+ * in the preceding step.
+ */
+ tuplesPerPrevGroup = Min(tuplesPerPrevGroup,
+ ceil(1.5 * tuplesPerPrevGroup / nGroups));
+ }
+
+ list_free(pathkeyExprs);
+
+ /* per_tuple_cost is in cpu_operator_cost units */
+ per_tuple_cost *= cpu_operator_cost;
+
+ /*
+ * Accordingly to "Introduction to algorithms", Thomas H. Cormen, Charles
+ * E. Leiserson, Ronald L. Rivest, ISBN 0-07-013143-0, quicksort
+ * estimation formula has additional term proportional to number of tuples
+ * (see Chapter 8.2 and Theorem 4.1). That affects cases with a low number
+ * of tuples, approximately less than 1e4. We could implement it as an
+ * additional multiplier under the logarithm, but we use a bit more
+ * complex formula which takes into account the number of unique tuples
+ * and it's not clear how to combine the multiplier with the number of
+ * groups. Estimate it as 10 cpu_operator_cost units.
+ */
+ per_tuple_cost += 10 * cpu_operator_cost;
+
+ per_tuple_cost += comparison_cost;
+
+ return tuples * per_tuple_cost;
+}
+
+/*
+ * simple wrapper just to estimate best sort path
+ */
+Cost
+cost_sort_estimate(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
+ double tuples)
+{
+ return compute_cpu_sort_cost(root, pathkeys, nPresortedKeys,
+ 0, tuples, tuples, false);
+}
+
/*
* cost_tuplesort
* Determines and returns the cost of sorting a relation using tuplesort,
@@ -1832,7 +2153,7 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
* number of initial runs formed and M is the merge order used by tuplesort.c.
* Since the average initial run should be about sort_mem, we have
* disk traffic = 2 * relsize * ceil(logM(p / sort_mem))
- * cpu = comparison_cost * t * log2(t)
+ * and cpu cost (computed by compute_cpu_sort_cost()).
*
* If the sort is bounded (i.e., only the first k result tuples are needed)
* and k tuples can fit into sort_mem, we use a heap method that keeps only
@@ -1851,9 +2172,11 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
* 'comparison_cost' is the extra cost per comparison, if any
* 'sort_mem' is the number of kilobytes of work memory allowed for the sort
* 'limit_tuples' is the bound on the number of output tuples; -1 if no bound
+ * 'startup_cost' is expected to be 0 at input. If there is "input cost" it should
+ * be added by caller later
*/
static void
-cost_tuplesort(Cost *startup_cost, Cost *run_cost,
+cost_tuplesort(PlannerInfo *root, List *pathkeys, Cost *startup_cost, Cost *run_cost,
double tuples, int width,
Cost comparison_cost, int sort_mem,
double limit_tuples)
@@ -1870,9 +2193,6 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
if (tuples < 2.0)
tuples = 2.0;
- /* Include the default cost-per-comparison */
- comparison_cost += 2.0 * cpu_operator_cost;
-
/* Do we have a useful LIMIT? */
if (limit_tuples > 0 && limit_tuples < tuples)
{
@@ -1896,12 +2216,10 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
double log_runs;
double npageaccesses;
- /*
- * CPU costs
- *
- * Assume about N log2 N comparisons
- */
- *startup_cost = comparison_cost * tuples * LOG2(tuples);
+ /* CPU costs */
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ tuples, false);
/* Disk costs */
@@ -1917,18 +2235,17 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
}
else if (tuples > 2 * output_tuples || input_bytes > sort_mem_bytes)
{
- /*
- * We'll use a bounded heap-sort keeping just K tuples in memory, for
- * a total number of tuple comparisons of N log2 K; but the constant
- * factor is a bit higher than for quicksort. Tweak it so that the
- * cost curve is continuous at the crossover point.
- */
- *startup_cost = comparison_cost * tuples * LOG2(2.0 * output_tuples);
+ /* We'll use a bounded heap-sort keeping just K tuples in memory. */
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ output_tuples, true);
}
else
{
/* We'll use plain quicksort on all the input tuples */
- *startup_cost = comparison_cost * tuples * LOG2(tuples);
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ tuples, false);
}
/*
@@ -2042,7 +2359,7 @@ cost_incremental_sort(Path *path,
* Estimate the average cost of sorting of one group where presorted keys
* are equal.
*/
- cost_tuplesort(&group_startup_cost, &group_run_cost,
+ cost_tuplesort(root, pathkeys, &group_startup_cost, &group_run_cost,
group_tuples, width, comparison_cost, sort_mem,
limit_tuples);
@@ -2102,7 +2419,7 @@ cost_sort(Path *path, PlannerInfo *root,
Cost startup_cost;
Cost run_cost;
- cost_tuplesort(&startup_cost, &run_cost,
+ cost_tuplesort(root, pathkeys, &startup_cost, &run_cost,
tuples, width,
comparison_cost, sort_mem,
limit_tuples);
@@ -2200,7 +2517,7 @@ append_nonpartial_cost(List *subpaths, int numpaths, int parallel_workers)
* Determines and returns the cost of an Append node.
*/
void
-cost_append(AppendPath *apath)
+cost_append(AppendPath *apath, PlannerInfo *root)
{
ListCell *l;
@@ -2268,7 +2585,7 @@ cost_append(AppendPath *apath)
* any child.
*/
cost_sort(&sort_path,
- NULL, /* doesn't currently need root */
+ root,
pathkeys,
subpath->total_cost,
subpath->rows,
diff --git a/src/backend/optimizer/path/equivclass.c b/src/backend/optimizer/path/equivclass.c
index 7fa502d6e2..07edd4f38e 100644
--- a/src/backend/optimizer/path/equivclass.c
+++ b/src/backend/optimizer/path/equivclass.c
@@ -652,7 +652,18 @@ get_eclass_for_sort_expr(PlannerInfo *root,
if (opcintype == cur_em->em_datatype &&
equal(expr, cur_em->em_expr))
- return cur_ec; /* Match! */
+ {
+ /*
+ * Match!
+ *
+ * Copy the sortref if it wasn't set yet. That may happen if
+ * the ec was constructed from WHERE clause, i.e. it doesn't
+ * have a target reference at all.
+ */
+ if (cur_ec->ec_sortref == 0 && sortref > 0)
+ cur_ec->ec_sortref = sortref;
+ return cur_ec;
+ }
}
}
diff --git a/src/backend/optimizer/path/pathkeys.c b/src/backend/optimizer/path/pathkeys.c
index e53ea84224..2143b39873 100644
--- a/src/backend/optimizer/path/pathkeys.c
+++ b/src/backend/optimizer/path/pathkeys.c
@@ -17,17 +17,24 @@
*/
#include "postgres.h"
+#include <float.h>
+
+#include "miscadmin.h"
#include "access/stratnum.h"
#include "catalog/pg_opfamily.h"
#include "nodes/makefuncs.h"
#include "nodes/nodeFuncs.h"
#include "nodes/plannodes.h"
+#include "optimizer/cost.h"
#include "optimizer/optimizer.h"
#include "optimizer/pathnode.h"
#include "optimizer/paths.h"
#include "partitioning/partbounds.h"
#include "utils/lsyscache.h"
+#include "utils/selfuncs.h"
+/* Consider reordering of GROUP BY keys? */
+bool enable_group_by_reordering = true;
static bool pathkey_is_redundant(PathKey *new_pathkey, List *pathkeys);
static bool matches_boolean_partition_clause(RestrictInfo *rinfo,
@@ -350,6 +357,548 @@ pathkeys_contained_in(List *keys1, List *keys2)
return false;
}
+/*
+ * group_keys_reorder_by_pathkeys
+ * Reorder GROUP BY keys to match pathkeys of input path.
+ *
+ * Function returns new lists (pathkeys and clauses), original GROUP BY lists
+ * stay untouched.
+ *
+ * Returns the number of GROUP BY keys with a matching pathkey.
+ */
+int
+group_keys_reorder_by_pathkeys(List *pathkeys, List **group_pathkeys,
+ List **group_clauses,
+ int num_groupby_pathkeys)
+{
+ List *new_group_pathkeys = NIL,
+ *new_group_clauses = NIL;
+ ListCell *lc;
+ int n;
+
+ if (pathkeys == NIL || *group_pathkeys == NIL)
+ return 0;
+
+ /*
+ * Walk the pathkeys (determining ordering of the input path) and see if
+ * there's a matching GROUP BY key. If we find one, we append it to the
+ * list, and do the same for the clauses.
+ *
+ * Once we find the first pathkey without a matching GROUP BY key, the
+ * rest of the pathkeys are useless and can't be used to evaluate the
+ * grouping, so we abort the loop and ignore the remaining pathkeys.
+ *
+ * XXX Pathkeys are built in a way to allow simply comparing pointers.
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(lc);
+ SortGroupClause *sgc;
+
+ if (foreach_current_index(lc) >= num_groupby_pathkeys ||
+ !list_member_ptr(*group_pathkeys, pathkey))
+ /* */
+ break;
+
+ /* abort on first mismatch */
+ sgc = get_sortgroupref_clause_noerr(pathkey->pk_eclass->ec_sortref,
+ *group_clauses);
+ if (!sgc)
+ /* The grouping clause is not cover this paathkey */
+ break;
+
+ new_group_pathkeys = lappend(new_group_pathkeys, pathkey);
+ new_group_clauses = lappend(new_group_clauses, sgc);
+ }
+
+ /* remember the number of pathkeys with a matching GROUP BY key */
+ n = list_length(new_group_pathkeys);
+
+ /* append the remaining group pathkeys (will be treated as not sorted) */
+ *group_pathkeys = list_concat_unique_ptr(new_group_pathkeys,
+ *group_pathkeys);
+ *group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+
+ return n;
+}
+
+/*
+ * Used to generate all permutations of a pathkey list.
+ */
+typedef struct PathkeyMutatorState
+{
+ List *elemsList;
+ ListCell **elemCells;
+ void **elems;
+ int *positions;
+ int mutatorNColumns;
+ int count;
+} PathkeyMutatorState;
+
+
+/*
+ * PathkeyMutatorInit
+ * Initialize state of the permutation generator.
+ *
+ * We want to generate permutations of elements in the "elems" list. We may want
+ * to skip some number of elements at the beginning (when treating as presorted)
+ * or at the end (we only permute a limited number of group keys).
+ *
+ * The list is decomposed into elements, and we also keep pointers to individual
+ * cells. This allows us to build the permuted list quickly and cheaply, without
+ * creating any copies.
+ */
+static void
+PathkeyMutatorInit(PathkeyMutatorState *state, List *elems, int start, int end)
+{
+ int i;
+ int n = end - start;
+ ListCell *lc;
+
+ memset(state, 0, sizeof(*state));
+
+ state->mutatorNColumns = n;
+
+ state->elemsList = list_copy(elems);
+
+ state->elems = palloc(sizeof(void *) * n);
+ state->elemCells = palloc(sizeof(ListCell *) * n);
+ state->positions = palloc(sizeof(int) * n);
+
+ i = 0;
+ for_each_cell(lc, state->elemsList, list_nth_cell(state->elemsList, start))
+ {
+ state->elemCells[i] = lc;
+ state->elems[i] = lfirst(lc);
+ state->positions[i] = i + 1;
+ i++;
+
+ if (i >= n)
+ break;
+ }
+}
+
+/* Swap two elements of an array. */
+static void
+PathkeyMutatorSwap(int *a, int i, int j)
+{
+ int s = a[i];
+
+ a[i] = a[j];
+ a[j] = s;
+}
+
+/*
+ * Generate the next permutation of elements.
+ */
+static bool
+PathkeyMutatorNextSet(int *a, int n)
+{
+ int j,
+ k,
+ l,
+ r;
+
+ j = n - 2;
+
+ while (j >= 0 && a[j] >= a[j + 1])
+ j--;
+
+ if (j < 0)
+ return false;
+
+ k = n - 1;
+
+ while (k >= 0 && a[j] >= a[k])
+ k--;
+
+ PathkeyMutatorSwap(a, j, k);
+
+ l = j + 1;
+ r = n - 1;
+
+ while (l < r)
+ PathkeyMutatorSwap(a, l++, r--);
+
+ return true;
+}
+
+/*
+ * PathkeyMutatorNext
+ * Generate the next permutation of list of elements.
+ *
+ * Returns the next permutation (as a list of elements) or NIL if there are no
+ * more permutations.
+ */
+static List *
+PathkeyMutatorNext(PathkeyMutatorState *state)
+{
+ int i;
+
+ state->count++;
+
+ /* first permutation is original list */
+ if (state->count == 1)
+ return state->elemsList;
+
+ /* when there are no more permutations, return NIL */
+ if (!PathkeyMutatorNextSet(state->positions, state->mutatorNColumns))
+ {
+ pfree(state->elems);
+ pfree(state->elemCells);
+ pfree(state->positions);
+
+ list_free(state->elemsList);
+
+ return NIL;
+ }
+
+ /* update the list cells to point to the right elements */
+ for (i = 0; i < state->mutatorNColumns; i++)
+ lfirst(state->elemCells[i]) =
+ (void *) state->elems[state->positions[i] - 1];
+
+ return state->elemsList;
+}
+
+/*
+ * Cost of comparing pathkeys.
+ */
+typedef struct PathkeySortCost
+{
+ Cost cost;
+ PathKey *pathkey;
+} PathkeySortCost;
+
+static int
+pathkey_sort_cost_comparator(const void *_a, const void *_b)
+{
+ const PathkeySortCost *a = (PathkeySortCost *) _a;
+ const PathkeySortCost *b = (PathkeySortCost *) _b;
+
+ if (a->cost < b->cost)
+ return -1;
+ else if (a->cost == b->cost)
+ return 0;
+ return 1;
+}
+
+/*
+ * get_cheapest_group_keys_order
+ * Reorders the group pathkeys / clauses to minimize the comparison cost.
+ *
+ * Given the list of pathkeys in '*group_pathkeys', we try to arrange these
+ * in an order that minimizes the sort costs that will be incurred by the
+ * GROUP BY. The costs mainly depend on the cost of the sort comparator
+ * function(s) and the number of distinct values in each column of the GROUP
+ * BY clause (*group_clauses). Sorting on subsequent columns is only required
+ * for tiebreak situations where two values sort equally.
+ *
+ * In case the input is partially sorted, only the remaining pathkeys are
+ * considered. 'n_preordered' denotes how many of the leading *group_pathkeys
+ * the input is presorted by.
+ *
+ * Returns true and sets *group_pathkeys and *group_clauses to the newly
+ * ordered versions of the lists that were passed in via these parameters.
+ * If no reordering was deemed necessary then we return false, in which case
+ * the *group_pathkeys and *group_clauses lists are left untouched. The
+ * original *group_pathkeys and *group_clauses parameter values are never
+ * destructively modified in place.
+ */
+static bool
+get_cheapest_group_keys_order(PlannerInfo *root, double nrows,
+ List **group_pathkeys, List **group_clauses,
+ int n_preordered)
+{
+ List *new_group_pathkeys = NIL;
+ List *new_group_clauses = NIL;
+ List *var_group_pathkeys;
+ ListCell *cell;
+ PathkeyMutatorState mstate;
+ double cheapest_sort_cost = DBL_MAX;
+ int nFreeKeys;
+ int nToPermute;
+
+ /* If there are less than 2 unsorted pathkeys, we're done. */
+ if (root->num_groupby_pathkeys - n_preordered < 2)
+ return false;
+
+ /*
+ * We could exhaustively cost all possible orderings of the pathkeys, but
+ * for a large number of pathkeys it might be prohibitively expensive. So
+ * we try to apply simple cheap heuristics first - we sort the pathkeys by
+ * sort cost (as if the pathkey was sorted independently) and then check
+ * only the four cheapest pathkeys. The remaining pathkeys are kept
+ * ordered by cost.
+ *
+ * XXX This is a very simple heuristics, but likely to work fine for most
+ * cases (because the number of GROUP BY clauses tends to be lower than
+ * 4). But it ignores how the number of distinct values in each pathkey
+ * affects the following steps. It might be better to use "more expensive"
+ * pathkey first if it has many distinct values, because it then limits
+ * the number of comparisons for the remaining pathkeys. But evaluating
+ * that is likely quite the expensive.
+ */
+ nFreeKeys = root->num_groupby_pathkeys - n_preordered;
+ nToPermute = 4;
+ if (nFreeKeys > nToPermute)
+ {
+ PathkeySortCost *costs = palloc(sizeof(PathkeySortCost) * nFreeKeys);
+ PathkeySortCost *cost = costs;
+ int i = 0;
+
+ /*
+ * Estimate cost for sorting individual pathkeys skipping the
+ * pre-ordered pathkeys.
+ */
+ for_each_from(cell, *group_pathkeys, n_preordered)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(cell);
+ List *to_cost;
+
+ if (foreach_current_index(cell) >= root->num_groupby_pathkeys)
+ break;
+
+ i++;
+ to_cost = list_make1(pathkey);
+ cost->pathkey = pathkey;
+ cost->cost = cost_sort_estimate(root, to_cost, 0, nrows);
+ cost++;
+
+ list_free(to_cost);
+ }
+
+ Assert(i == nFreeKeys);
+ /* sort the pathkeys by sort cost in ascending order */
+ qsort(costs, nFreeKeys, sizeof(*costs), pathkey_sort_cost_comparator);
+
+ /*
+ * Rebuild the list of pathkeys - first the preordered ones, then the
+ * rest ordered by cost.
+ */
+ new_group_pathkeys = list_copy_head(*group_pathkeys, n_preordered);
+
+ for (int i = 0; i < nFreeKeys; i++)
+ new_group_pathkeys = lappend(new_group_pathkeys, costs[i].pathkey);
+
+ new_group_pathkeys = list_copy_tail(*group_pathkeys,
+ root->num_groupby_pathkeys);
+ pfree(costs);
+ }
+ else
+ {
+ /* Copy the list, so that we can free the new list by list_free. */
+ new_group_pathkeys = list_copy(*group_pathkeys);
+ nToPermute = nFreeKeys;
+ }
+
+ Assert(list_length(new_group_pathkeys) == list_length(*group_pathkeys));
+
+ /*
+ * Generate pathkey lists with permutations of the first nToPermute
+ * pathkeys.
+ *
+ * XXX We simply calculate sort cost for each individual pathkey list, but
+ * there's room for two dynamic programming optimizations here. Firstly,
+ * we may pass the current "best" cost to cost_sort_estimate so that it
+ * can "abort" if the estimated pathkeys list exceeds it. Secondly, it
+ * could pass the return information about the position when it exceeded
+ * the cost, and we could skip all permutations with the same prefix.
+ *
+ * Imagine we've already found ordering with cost C1, and we're evaluating
+ * another ordering - cost_sort_estimate() calculates cost by adding the
+ * pathkeys one by one (more or less), and the cost only grows. If at any
+ * point it exceeds C1, it can't possibly be "better" so we can discard
+ * it. But we also know that we can discard all ordering with the same
+ * prefix, because if we're estimating (a,b,c,d) and we exceed C1 at (a,b)
+ * then the same thing will happen for any ordering with this prefix.
+ */
+ PathkeyMutatorInit(&mstate, new_group_pathkeys, n_preordered,
+ n_preordered + nToPermute);
+
+ while ((var_group_pathkeys = PathkeyMutatorNext(&mstate)) != NIL)
+ {
+ Cost cost;
+
+ cost = cost_sort_estimate(root, var_group_pathkeys, n_preordered, nrows);
+
+ if (cost < cheapest_sort_cost)
+ {
+ list_free(new_group_pathkeys);
+ new_group_pathkeys = list_copy(var_group_pathkeys);
+ cheapest_sort_cost = cost;
+ }
+ }
+
+ Assert(list_length(new_group_pathkeys) == list_length(root->group_pathkeys));
+
+ /* Reorder the group clauses according to the reordered pathkeys. */
+ foreach(cell, new_group_pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(cell);
+
+ if (foreach_current_index(cell) >= root->num_groupby_pathkeys)
+ break;
+
+ new_group_clauses =
+ lappend(new_group_clauses,
+ get_sortgroupref_clause(pathkey->pk_eclass->ec_sortref,
+ *group_clauses));
+ }
+
+ /* Just append the rest GROUP BY clauses */
+ new_group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+ *group_pathkeys = new_group_pathkeys;
+ *group_clauses = new_group_clauses;
+ Assert(list_length(new_group_clauses) == list_length(root->processed_groupClause));
+
+ return true;
+}
+
+/*
+ * get_useful_group_keys_orderings
+ * Determine which orderings of GROUP BY keys are potentially interesting.
+ *
+ * Returns list of PathKeyInfo items, each representing an interesting ordering
+ * of GROUP BY keys. Each item stores pathkeys and clauses in matching order.
+ *
+ * The function considers (and keeps) multiple group by orderings:
+ *
+ * - the original ordering, as specified by the GROUP BY clause
+ *
+ * - GROUP BY keys reordered to minimize the sort cost
+ *
+ * - GROUP BY keys reordered to match path ordering (as much as possible), with
+ * the tail reordered to minimize the sort cost
+ *
+ * - GROUP BY keys to match target ORDER BY clause (as much as possible), with
+ * the tail reordered to minimize the sort cost
+ *
+ * There are other potentially interesting orderings (e.g. it might be best to
+ * match the first ORDER BY key, order the remaining keys differently and then
+ * rely on the incremental sort to fix this), but we ignore those for now. To
+ * make this work we'd have to pretty much generate all possible permutations.
+ */
+List *
+get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys)
+{
+ Query *parse = root->parse;
+ List *infos = NIL;
+ PathKeyInfo *info;
+ int n_preordered = 0;
+
+ List *pathkeys = root->group_pathkeys;
+ List *clauses = root->processed_groupClause;
+
+ /* always return at least the original pathkeys/clauses */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+ infos = lappend(infos, info);
+
+ /*
+ * Should we try generating alternative orderings of the group keys? If
+ * not, we produce only the order specified in the query, i.e. the
+ * optimization is effectively disabled.
+ */
+ if (!enable_group_by_reordering)
+ return infos;
+
+ /* for grouping sets we can't do any reordering */
+ if (parse->groupingSets)
+ return infos;
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost, ignoring both the
+ * target ordering (ORDER BY) and ordering of the input path.
+ */
+ if (get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered))
+ {
+ Assert(list_length(pathkeys) == list_length(clauses));
+
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * If the path is sorted in some way, try reordering the group keys to
+ * match as much of the ordering as possible - we get this sort for free
+ * (mostly).
+ *
+ * We must not do this when there are no grouping sets, because those use
+ * more complex logic to decide the ordering.
+ *
+ * XXX Isn't this somewhat redundant with presorted_keys? Actually, it's
+ * more a complement, because it allows benefiting from incremental sort
+ * as much as possible.
+ *
+ * XXX This does nothing if (n_preordered == 0). We shouldn't create the
+ * info in this case.
+ */
+ if (path_pathkeys)
+ {
+ n_preordered = group_keys_reorder_by_pathkeys(path_pathkeys,
+ &pathkeys,
+ &clauses,
+ root->num_groupby_pathkeys);
+
+ /* reorder the tail to minimize sort cost */
+ get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, in
+ * which case get_cheapest_group_keys_order returns false. But we
+ * still want to keep the keys reordered to path_pathkeys.
+ */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost (this time consider
+ * the ORDER BY clause, but only if set debug_group_by_match_order_by).
+ */
+ if (root->sort_pathkeys)
+ {
+ n_preordered = group_keys_reorder_by_pathkeys(root->sort_pathkeys,
+ &pathkeys,
+ &clauses,
+ root->num_groupby_pathkeys);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, in
+ * which case get_cheapest_group_keys_order returns false. But we
+ * still want to keep the keys reordered to sort_pathkeys.
+ */
+ get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered);
+
+ /* keep the group keys reordered to match ordering of input path */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ return infos;
+}
+
/*
* pathkeys_count_contained_in
* Same as pathkeys_contained_in, but also sets length of longest
@@ -1937,6 +2486,54 @@ pathkeys_useful_for_ordering(PlannerInfo *root, List *pathkeys)
return n_common_pathkeys;
}
+/*
+ * pathkeys_useful_for_grouping
+ * Count the number of pathkeys that are useful for grouping (instead of
+ * explicit sort)
+ *
+ * Group pathkeys could be reordered to benefit from the ordering. The
+ * ordering may not be "complete" and may require incremental sort, but that's
+ * fine. So we simply count prefix pathkeys with a matching group key, and
+ * stop once we find the first pathkey without a match.
+ *
+ * So e.g. with pathkeys (a,b,c) and group keys (a,b,e) this determines (a,b)
+ * pathkeys are useful for grouping, and we might do incremental sort to get
+ * path ordered by (a,b,e).
+ *
+ * This logic is necessary to retain paths with ordering not matching grouping
+ * keys directly, without the reordering.
+ *
+ * Returns the length of pathkey prefix with matching group keys.
+ */
+static int
+pathkeys_useful_for_grouping(PlannerInfo *root, List *pathkeys)
+{
+ ListCell *key;
+ int n = 0;
+
+ /* no special ordering requested for grouping */
+ if (root->group_pathkeys == NIL)
+ return 0;
+
+ /* unordered path */
+ if (pathkeys == NIL)
+ return 0;
+
+ /* walk the pathkeys and search for matching group key */
+ foreach(key, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(key);
+
+ /* no matching group key, we're done */
+ if (!list_member_ptr(root->group_pathkeys, pathkey))
+ break;
+
+ n++;
+ }
+
+ return n;
+}
+
/*
* truncate_useless_pathkeys
* Shorten the given pathkey list to just the useful pathkeys.
@@ -1951,6 +2548,9 @@ truncate_useless_pathkeys(PlannerInfo *root,
nuseful = pathkeys_useful_for_merging(root, rel, pathkeys);
nuseful2 = pathkeys_useful_for_ordering(root, pathkeys);
+ if (nuseful2 > nuseful)
+ nuseful = nuseful2;
+ nuseful2 = pathkeys_useful_for_grouping(root, pathkeys);
if (nuseful2 > nuseful)
nuseful = nuseful2;
@@ -1986,6 +2586,8 @@ has_useful_pathkeys(PlannerInfo *root, RelOptInfo *rel)
{
if (rel->joininfo != NIL || rel->has_eclass_joins)
return true; /* might be able to use pathkeys for merging */
+ if (root->group_pathkeys != NIL)
+ return true; /* might be able to use pathkeys for grouping */
if (root->query_pathkeys != NIL)
return true; /* might be able to use them for ordering */
return false; /* definitely useless */
diff --git a/src/backend/optimizer/plan/planner.c b/src/backend/optimizer/plan/planner.c
index 44efb1f4eb..e4d3273f80 100644
--- a/src/backend/optimizer/plan/planner.c
+++ b/src/backend/optimizer/plan/planner.c
@@ -2810,8 +2810,9 @@ remove_useless_groupby_columns(PlannerInfo *root)
*
* In principle it might be interesting to consider other orderings of the
* GROUP BY elements, which could match the sort ordering of other
- * possible plans (eg an indexscan) and thereby reduce cost. We don't
- * bother with that, though. Hashed grouping will frequently win anyway.
+ * possible plans (eg an indexscan) and thereby reduce cost. However, we
+ * don't yet have sufficient information to do that here, so that's left until
+ * later in planning. See get_useful_group_keys_orderings().
*
* Note: we need no comparable processing of the distinctClause because
* the parser already enforced that that matches ORDER BY.
@@ -3468,6 +3469,7 @@ standard_qp_callback(PlannerInfo *root, void *extra)
tlist,
true,
&sortable);
+
if (!sortable)
{
/* Can't sort; no point in considering aggregate ordering either */
@@ -6842,90 +6844,108 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
+
+ Assert(list_length(pathkey_orderings) > 0);
- if (!is_sorted)
+ foreach(lc2, pathkey_orderings)
{
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially sorted
- * already (no need to deal with paths which have presorted
- * keys when incremental sort is disabled unless it's the
- * cheapest input path).
- */
- if (path != cheapest_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ bool is_sorted;
+ int presorted_keys;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- /*
- * We've no need to consider both a sort and incremental sort.
- * We'll just do a sort if there are no presorted keys and an
- * incremental sort when there are presorted keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
- /* Now decide what to stick atop it */
- if (parse->groupingSets)
- {
- consider_groupingsets_paths(root, grouped_rel,
- path, true, can_hash,
- gd, agg_costs, dNumGroups);
- }
- else if (parse->hasAggs)
- {
- /*
- * We have aggregation, possibly with plain GROUP BY. Make an
- * AggPath.
- */
- add_path(grouped_rel, (Path *)
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (!is_sorted)
+ {
+ /*
+ * Try at least sorting the cheapest path and also try
+ * incrementally sorting any path which is partially sorted
+ * already (no need to deal with paths which have presorted
+ * keys when incremental sort is disabled unless it's the
+ * cheapest input path).
+ */
+ if (path != cheapest_path &&
+ (presorted_keys == 0 || !enable_incremental_sort))
+ continue;
+
+ /*
+ * We've no need to consider both a sort and incremental sort.
+ * We'll just do a sort if there are no presorted keys and an
+ * incremental sort when there are presorted keys.
+ */
+ if (presorted_keys == 0 || !enable_incremental_sort)
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ else
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+ }
+
+ /* Now decide what to stick atop it */
+ if (parse->groupingSets)
+ {
+ consider_groupingsets_paths(root, grouped_rel,
+ path, true, can_hash,
+ gd, agg_costs, dNumGroups);
+ }
+ else if (parse->hasAggs)
+ {
+ /*
+ * We have aggregation, possibly with plain GROUP BY. Make an
+ * AggPath.
+ */
+ add_path(grouped_rel, (Path *)
create_agg_path(root,
grouped_rel,
path,
grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_SIMPLE,
- root->processed_groupClause,
+ info->clauses,
havingQual,
agg_costs,
dNumGroups));
- }
- else if (parse->groupClause)
- {
- /*
- * We have GROUP BY without aggregation or grouping sets. Make
- * a GroupPath.
- */
- add_path(grouped_rel, (Path *)
+ }
+ else if (parse->groupClause)
+ {
+ /*
+ * We have GROUP BY without aggregation or grouping sets. Make
+ * a GroupPath.
+ */
+ add_path(grouped_rel, (Path *)
create_group_path(root,
grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
havingQual,
dNumGroups));
- }
- else
- {
- /* Other cases should have been handled above */
- Assert(false);
+ }
+ else
+ {
+ /* Other cases should have been handled above */
+ Assert(false);
+ }
}
}
@@ -6937,69 +6957,84 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
{
foreach(lc, partially_grouped_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
+ List *pathkey_orderings = NIL;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
- if (!is_sorted)
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach(lc2, pathkey_orderings)
{
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially
- * sorted already (no need to deal with paths which have
- * presorted keys when incremental sort is disabled unless
- * it's the cheapest input path).
- */
- if (path != partially_grouped_rel->cheapest_total_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ bool is_sorted;
+ int presorted_keys;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- /*
- * We've no need to consider both a sort and incremental
- * sort. We'll just do a sort if there are no pre-sorted
- * keys and an incremental sort when there are presorted
- * keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
+ path->pathkeys,
+ &presorted_keys);
- if (parse->hasAggs)
- add_path(grouped_rel, (Path *)
+ if (!is_sorted)
+ {
+ /*
+ * Try at least sorting the cheapest path and also try
+ * incrementally sorting any path which is partially
+ * sorted already (no need to deal with paths which have
+ * presorted keys when incremental sort is disabled unless
+ * it's the cheapest input path).
+ */
+ if (path != partially_grouped_rel->cheapest_total_path &&
+ (presorted_keys == 0 || !enable_incremental_sort))
+ continue;
+
+ /*
+ * We've no need to consider both a sort and incremental
+ * sort. We'll just do a sort if there are no pre-sorted
+ * keys and an incremental sort when there are presorted
+ * keys.
+ */
+ if (presorted_keys == 0 || !enable_incremental_sort)
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ else
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_path(grouped_rel, (Path *)
create_agg_path(root,
grouped_rel,
path,
grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_FINAL_DESERIAL,
- root->processed_groupClause,
+ info->clauses,
havingQual,
agg_final_costs,
dNumGroups));
- else
- add_path(grouped_rel, (Path *)
+ else
+ add_path(grouped_rel, (Path *)
create_group_path(root,
grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
havingQual,
dNumGroups));
+ }
}
}
}
@@ -7202,66 +7237,85 @@ create_partial_grouping_paths(PlannerInfo *root,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
- if (!is_sorted)
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach(lc2, pathkey_orderings)
{
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially sorted
- * already (no need to deal with paths which have presorted
- * keys when incremental sort is disabled unless it's the
- * cheapest input path).
- */
- if (path != cheapest_total_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ bool is_sorted;
+ int presorted_keys;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- /*
- * We've no need to consider both a sort and incremental sort.
- * We'll just do a sort if there are no presorted keys and an
- * incremental sort when there are presorted keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
- if (parse->hasAggs)
- add_path(partially_grouped_rel, (Path *)
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+ if (!is_sorted)
+ {
+ /*
+ * Try at least sorting the cheapest path and also try
+ * incrementally sorting any path which is partially sorted
+ * already (no need to deal with paths which have presorted
+ * keys when incremental sort is disabled unless it's the
+ * cheapest input path).
+ */
+ if (path != cheapest_total_path &&
+ (presorted_keys == 0 || !enable_incremental_sort))
+ continue;
+
+ /*
+ * We've no need to consider both a sort and incremental sort.
+ * We'll just do a sort if there are no presorted keys and an
+ * incremental sort when there are presorted keys.
+ */
+ if (presorted_keys == 0 || !enable_incremental_sort)
+ path = (Path *) create_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ else
+ path = (Path *) create_incremental_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_path(partially_grouped_rel, (Path *)
create_agg_path(root,
partially_grouped_rel,
path,
partially_grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_INITIAL_SERIAL,
- root->processed_groupClause,
+ info->clauses,
NIL,
agg_partial_costs,
dNumPartialGroups));
- else
- add_path(partially_grouped_rel, (Path *)
- create_group_path(root,
- partially_grouped_rel,
- path,
- root->processed_groupClause,
- NIL,
- dNumPartialGroups));
+ else
+ add_path(partially_grouped_rel, (Path *)
+ create_group_path(root,
+ partially_grouped_rel,
+ path,
+ info->clauses,
+ NIL,
+ dNumPartialGroups));
+ }
}
}
@@ -7270,67 +7324,86 @@ create_partial_grouping_paths(PlannerInfo *root,
/* Similar to above logic, but for partial paths. */
foreach(lc, input_rel->partial_pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
- if (!is_sorted)
- {
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially sorted
- * already (no need to deal with paths which have presorted
- * keys when incremental sort is disabled unless it's the
- * cheapest input path).
- */
- if (path != cheapest_partial_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ Assert(list_length(pathkey_orderings) > 0);
- /*
- * We've no need to consider both a sort and incremental sort.
- * We'll just do a sort if there are no presorted keys and an
- * incremental sort when there are presorted keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ /* process all potentially interesting grouping reorderings */
+ foreach(lc2, pathkey_orderings)
+ {
+ bool is_sorted;
+ int presorted_keys;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
- if (parse->hasAggs)
- add_partial_path(partially_grouped_rel, (Path *)
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (!is_sorted)
+ {
+ /*
+ * Try at least sorting the cheapest path and also try
+ * incrementally sorting any path which is partially sorted
+ * already (no need to deal with paths which have presorted
+ * keys when incremental sort is disabled unless it's the
+ * cheapest input path).
+ */
+ if (path != cheapest_partial_path &&
+ (presorted_keys == 0 || !enable_incremental_sort))
+ continue;
+
+ /*
+ * We've no need to consider both a sort and incremental sort.
+ * We'll just do a sort if there are no presorted keys and an
+ * incremental sort when there are presorted keys.
+ */
+ if (presorted_keys == 0 || !enable_incremental_sort)
+ path = (Path *) create_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ else
+ path = (Path *) create_incremental_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_partial_path(partially_grouped_rel, (Path *)
create_agg_path(root,
partially_grouped_rel,
path,
partially_grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_INITIAL_SERIAL,
- root->processed_groupClause,
+ info->clauses,
NIL,
agg_partial_costs,
dNumPartialPartialGroups));
- else
- add_partial_path(partially_grouped_rel, (Path *)
+ else
+ add_partial_path(partially_grouped_rel, (Path *)
create_group_path(root,
partially_grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
NIL,
dNumPartialPartialGroups));
+ }
}
}
@@ -7444,6 +7517,8 @@ gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
* We can also skip the entire loop when we only have a single-item
* group_pathkeys because then we can't possibly have a presorted prefix
* of the list without having the list be fully sorted.
+ *
+ * XXX Shouldn't this also consider the group-key-reordering?
*/
if (!enable_incremental_sort || list_length(root->group_pathkeys) == 1)
return;
diff --git a/src/backend/optimizer/util/pathnode.c b/src/backend/optimizer/util/pathnode.c
index f123fcb41e..c72ea08447 100644
--- a/src/backend/optimizer/util/pathnode.c
+++ b/src/backend/optimizer/util/pathnode.c
@@ -1346,12 +1346,12 @@ create_append_path(PlannerInfo *root,
pathnode->path.total_cost = child->total_cost;
}
else
- cost_append(pathnode);
+ cost_append(pathnode, root);
/* Must do this last, else cost_append complains */
pathnode->path.pathkeys = child->pathkeys;
}
else
- cost_append(pathnode);
+ cost_append(pathnode, root);
/* If the caller provided a row estimate, override the computed value. */
if (rows >= 0)
diff --git a/src/backend/utils/adt/selfuncs.c b/src/backend/utils/adt/selfuncs.c
index c4fcd0076e..cb7e67099e 100644
--- a/src/backend/utils/adt/selfuncs.c
+++ b/src/backend/utils/adt/selfuncs.c
@@ -3386,11 +3386,28 @@ double
estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
List **pgset, EstimationInfo *estinfo)
{
- List *varinfos = NIL;
+ return estimate_num_groups_incremental(root, groupExprs,
+ input_rows, pgset, estinfo,
+ NULL, 0);
+}
+
+/*
+ * estimate_num_groups_incremental
+ * An estimate_num_groups variant, optimized for cases that are adding the
+ * expressions incrementally (e.g. one by one).
+ */
+double
+estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
+ double input_rows,
+ List **pgset, EstimationInfo *estinfo,
+ List **cache_varinfos, int prevNExprs)
+{
+ List *varinfos = (cache_varinfos) ? *cache_varinfos : NIL;
double srf_multiplier = 1.0;
double numdistinct;
ListCell *l;
- int i;
+ int i,
+ j;
/* Zero the estinfo output parameter, if non-NULL */
if (estinfo != NULL)
@@ -3421,7 +3438,7 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
*/
numdistinct = 1.0;
- i = 0;
+ i = j = 0;
foreach(l, groupExprs)
{
Node *groupexpr = (Node *) lfirst(l);
@@ -3430,6 +3447,14 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
List *varshere;
ListCell *l2;
+ /* was done on previous call */
+ if (cache_varinfos && j++ < prevNExprs)
+ {
+ if (pgset)
+ i++; /* to keep in sync with lines below */
+ continue;
+ }
+
/* is expression in this grouping set? */
if (pgset && !list_member_int(*pgset, i++))
continue;
@@ -3499,7 +3524,11 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
if (varshere == NIL)
{
if (contain_volatile_functions(groupexpr))
+ {
+ if (cache_varinfos)
+ *cache_varinfos = varinfos;
return input_rows;
+ }
continue;
}
@@ -3516,6 +3545,9 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
}
}
+ if (cache_varinfos)
+ *cache_varinfos = varinfos;
+
/*
* If now no Vars, we must have an all-constant or all-boolean GROUP BY
* list.
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index f9dba43b8c..1c10ab5a7d 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -1025,6 +1025,16 @@ struct config_bool ConfigureNamesBool[] =
true,
NULL, NULL, NULL
},
+ {
+ {"enable_group_by_reordering", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("Enables reordering of GROUP BY keys."),
+ NULL,
+ GUC_EXPLAIN
+ },
+ &enable_group_by_reordering,
+ true,
+ NULL, NULL, NULL
+ },
{
{"geqo", PGC_USERSET, QUERY_TUNING_GEQO,
gettext_noop("Enables genetic query optimization."),
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index c768af9a73..671b88fc60 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -393,6 +393,7 @@
#enable_seqscan = on
#enable_sort = on
#enable_tidscan = on
+#enable_group_by_reordering = on
# - Planner Cost Constants -
diff --git a/src/include/nodes/pathnodes.h b/src/include/nodes/pathnodes.h
index c17b53f7ad..eeeafd8e76 100644
--- a/src/include/nodes/pathnodes.h
+++ b/src/include/nodes/pathnodes.h
@@ -1456,6 +1456,16 @@ typedef struct PathKey
bool pk_nulls_first; /* do NULLs come before normal values? */
} PathKey;
+/*
+ * Combines information about pathkeys and the associated clauses.
+ */
+typedef struct PathKeyInfo
+{
+ NodeTag type;
+ List *pathkeys;
+ List *clauses;
+} PathKeyInfo;
+
/*
* VolatileFunctionStatus -- allows nodes to cache their
* contain_volatile_functions properties. VOLATILITY_UNKNOWN means not yet
diff --git a/src/include/optimizer/cost.h b/src/include/optimizer/cost.h
index 6cf49705d3..28175d3472 100644
--- a/src/include/optimizer/cost.h
+++ b/src/include/optimizer/cost.h
@@ -116,7 +116,9 @@ extern void cost_incremental_sort(Path *path,
Cost input_startup_cost, Cost input_total_cost,
double input_tuples, int width, Cost comparison_cost, int sort_mem,
double limit_tuples);
-extern void cost_append(AppendPath *apath);
+extern Cost cost_sort_estimate(PlannerInfo *root, List *pathkeys,
+ int nPresortedKeys, double tuples);
+extern void cost_append(AppendPath *path, PlannerInfo *root);
extern void cost_merge_append(Path *path, PlannerInfo *root,
List *pathkeys, int n_streams,
Cost input_startup_cost, Cost input_total_cost,
diff --git a/src/include/optimizer/paths.h b/src/include/optimizer/paths.h
index 50bc3b503a..e4ff16d0c4 100644
--- a/src/include/optimizer/paths.h
+++ b/src/include/optimizer/paths.h
@@ -24,6 +24,7 @@ extern PGDLLIMPORT bool enable_geqo;
extern PGDLLIMPORT int geqo_threshold;
extern PGDLLIMPORT int min_parallel_table_scan_size;
extern PGDLLIMPORT int min_parallel_index_scan_size;
+extern PGDLLIMPORT bool enable_group_by_reordering;
/* Hook for plugins to get control in set_rel_pathlist() */
typedef void (*set_rel_pathlist_hook_type) (PlannerInfo *root,
@@ -205,6 +206,12 @@ typedef enum
extern PathKeysComparison compare_pathkeys(List *keys1, List *keys2);
extern bool pathkeys_contained_in(List *keys1, List *keys2);
extern bool pathkeys_count_contained_in(List *keys1, List *keys2, int *n_common);
+extern int group_keys_reorder_by_pathkeys(List *pathkeys,
+ List **group_pathkeys,
+ List **group_clauses,
+ int num_groupby_pathkeys);
+extern List *get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys);
extern Path *get_cheapest_path_for_pathkeys(List *paths, List *pathkeys,
Relids required_outer,
CostSelector cost_criterion,
diff --git a/src/include/utils/selfuncs.h b/src/include/utils/selfuncs.h
index 2f76c473db..2b21c06cc9 100644
--- a/src/include/utils/selfuncs.h
+++ b/src/include/utils/selfuncs.h
@@ -214,6 +214,11 @@ extern double estimate_num_groups(PlannerInfo *root, List *groupExprs,
double input_rows, List **pgset,
EstimationInfo *estinfo);
+extern double estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
+ double input_rows, List **pgset,
+ EstimationInfo *estinfo,
+ List **cache_varinfos, int prevNExprs);
+
extern void estimate_hash_bucket_stats(PlannerInfo *root,
Node *hashkey, double nbuckets,
Selectivity *mcv_freq,
diff --git a/src/test/regress/expected/aggregates.out b/src/test/regress/expected/aggregates.out
index d8271da4d1..205be28b3d 100644
--- a/src/test/regress/expected/aggregates.out
+++ b/src/test/regress/expected/aggregates.out
@@ -1234,7 +1234,8 @@ explain (costs off)
select distinct min(f1), max(f1) from minmaxtest;
QUERY PLAN
---------------------------------------------------------------------------------------------
- Unique
+ HashAggregate
+ Group Key: $0, $1
InitPlan 1 (returns $0)
-> Limit
-> Merge Append
@@ -1257,10 +1258,8 @@ explain (costs off)
-> Index Only Scan using minmaxtest2i on minmaxtest2 minmaxtest_8
Index Cond: (f1 IS NOT NULL)
-> Index Only Scan Backward using minmaxtest3i on minmaxtest3 minmaxtest_9
- -> Sort
- Sort Key: ($0), ($1)
- -> Result
-(26 rows)
+ -> Result
+(25 rows)
select distinct min(f1), max(f1) from minmaxtest;
min | max
@@ -2715,6 +2714,241 @@ SELECT balk(hundred) FROM tenk1;
(1 row)
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+SELECT
+ i AS id,
+ i/2 AS p,
+ format('%60s', i%2) AS v,
+ i/4 AS c,
+ i/8 AS d,
+ (random() * (10000/8))::int as e --the same as d but no correlation with p
+ INTO btg
+FROM
+ generate_series(1, 10000) i;
+VACUUM btg;
+ANALYZE btg;
+-- GROUP BY optimization by reorder columns by frequency
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Sort
+ Sort Key: p, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Sort
+ Sort Key: p, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, c, v
+ -> Sort
+ Sort Key: p, c, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: v, p, c
+ -> Sort
+ Sort Key: v, p, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: p, d, v, c
+ -> Sort
+ Sort Key: p, d, v, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: v, p, d, c
+ -> Sort
+ Sort Key: v, p, d, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: p, v, d, c
+ -> Sort
+ Sort Key: p, v, d, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, d, e
+ -> Sort
+ Sort Key: p, d, e
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, e, d
+ -> Sort
+ Sort Key: p, e, d
+ -> Seq Scan on btg
+(5 rows)
+
+CREATE STATISTICS btg_dep ON d, e, p FROM btg;
+ANALYZE btg;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, d, e
+ -> Sort
+ Sort Key: p, d, e
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, e, d
+ -> Sort
+ Sort Key: p, e, d
+ -> Seq Scan on btg
+(5 rows)
+
+-- GROUP BY optimization by reorder columns by index scan
+CREATE INDEX ON btg(p, v);
+SET enable_seqscan=off;
+SET enable_bitmapscan=off;
+VACUUM btg;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v ORDER BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p ORDER BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, c, v
+ -> Incremental Sort
+ Sort Key: p, c, v
+ Presorted Key: p
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY p, v;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, v, c
+ -> Incremental Sort
+ Sort Key: p, v, c
+ Presorted Key: p, v
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, c, v, d
+ -> Incremental Sort
+ Sort Key: p, c, v, d
+ Presorted Key: p
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d ORDER BY p, v;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, v, c, d
+ -> Incremental Sort
+ Sort Key: p, v, c, d
+ Presorted Key: p, v
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+DROP TABLE btg;
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+RESET enable_seqscan;
+RESET enable_bitmapscan;
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/expected/incremental_sort.out b/src/test/regress/expected/incremental_sort.out
index 7fdb685313..b8a0968b06 100644
--- a/src/test/regress/expected/incremental_sort.out
+++ b/src/test/regress/expected/incremental_sort.out
@@ -1426,7 +1426,7 @@ set parallel_setup_cost = 0;
set parallel_tuple_cost = 0;
set max_parallel_workers_per_gather = 2;
create table t (a int, b int, c int);
-insert into t select mod(i,10),mod(i,10),i from generate_series(1,10000) s(i);
+insert into t select mod(i,10),mod(i,10),i from generate_series(1,60000) s(i);
create index on t (a);
analyze t;
set enable_incremental_sort = off;
diff --git a/src/test/regress/expected/join.out b/src/test/regress/expected/join.out
index 9b8638f286..27684bfce2 100644
--- a/src/test/regress/expected/join.out
+++ b/src/test/regress/expected/join.out
@@ -1984,8 +1984,8 @@ USING (name);
------+----+----
bb | 12 | 13
cc | 22 | 23
- dd | | 33
ee | 42 |
+ dd | | 33
(4 rows)
-- Cases with non-nullable expressions in subquery results;
@@ -2019,8 +2019,8 @@ NATURAL FULL JOIN
------+------+------+------+------
bb | 12 | 2 | 13 | 3
cc | 22 | 2 | 23 | 3
- dd | | | 33 | 3
ee | 42 | 2 | |
+ dd | | | 33 | 3
(4 rows)
SELECT * FROM
@@ -5645,18 +5645,20 @@ select d.* from d left join (select * from b group by b.id, b.c_id) s
explain (costs off)
select d.* from d left join (select distinct * from b) s
on d.a = s.id;
- QUERY PLAN
---------------------------------------
- Merge Right Join
- Merge Cond: (b.id = d.a)
- -> Unique
- -> Sort
- Sort Key: b.id, b.c_id
- -> Seq Scan on b
+ QUERY PLAN
+---------------------------------------------
+ Merge Left Join
+ Merge Cond: (d.a = s.id)
-> Sort
Sort Key: d.a
-> Seq Scan on d
-(9 rows)
+ -> Sort
+ Sort Key: s.id
+ -> Subquery Scan on s
+ -> HashAggregate
+ Group Key: b.id, b.c_id
+ -> Seq Scan on b
+(11 rows)
-- join removal is not possible here
explain (costs off)
@@ -7597,44 +7599,39 @@ select * from j1 natural join j2;
explain (verbose, costs off)
select * from j1
inner join (select distinct id from j3) j3 on j1.id = j3.id;
- QUERY PLAN
------------------------------------------
+ QUERY PLAN
+-----------------------------------
Nested Loop
Output: j1.id, j3.id
Inner Unique: true
Join Filter: (j1.id = j3.id)
- -> Unique
+ -> HashAggregate
Output: j3.id
- -> Sort
+ Group Key: j3.id
+ -> Seq Scan on public.j3
Output: j3.id
- Sort Key: j3.id
- -> Seq Scan on public.j3
- Output: j3.id
-> Seq Scan on public.j1
Output: j1.id
-(13 rows)
+(11 rows)
-- ensure group by clause allows the inner to become unique
explain (verbose, costs off)
select * from j1
inner join (select id from j3 group by id) j3 on j1.id = j3.id;
- QUERY PLAN
------------------------------------------
+ QUERY PLAN
+-----------------------------------
Nested Loop
Output: j1.id, j3.id
Inner Unique: true
Join Filter: (j1.id = j3.id)
- -> Group
+ -> HashAggregate
Output: j3.id
Group Key: j3.id
- -> Sort
+ -> Seq Scan on public.j3
Output: j3.id
- Sort Key: j3.id
- -> Seq Scan on public.j3
- Output: j3.id
-> Seq Scan on public.j1
Output: j1.id
-(14 rows)
+(11 rows)
drop table j1;
drop table j2;
diff --git a/src/test/regress/expected/merge.out b/src/test/regress/expected/merge.out
index 133d42117c..f7c4cc4833 100644
--- a/src/test/regress/expected/merge.out
+++ b/src/test/regress/expected/merge.out
@@ -1459,18 +1459,15 @@ WHEN MATCHED AND t.a < 10 THEN
explain_merge
--------------------------------------------------------------------
Merge on ex_mtarget t (actual rows=0 loops=1)
- -> Merge Join (actual rows=0 loops=1)
- Merge Cond: (t.a = s.a)
- -> Sort (actual rows=0 loops=1)
- Sort Key: t.a
- Sort Method: quicksort Memory: xxx
+ -> Hash Join (actual rows=0 loops=1)
+ Hash Cond: (s.a = t.a)
+ -> Seq Scan on ex_msource s (actual rows=1 loops=1)
+ -> Hash (actual rows=0 loops=1)
+ Buckets: xxx Batches: xxx Memory Usage: xxx
-> Seq Scan on ex_mtarget t (actual rows=0 loops=1)
Filter: (a < '-1000'::integer)
Rows Removed by Filter: 54
- -> Sort (never executed)
- Sort Key: s.a
- -> Seq Scan on ex_msource s (never executed)
-(12 rows)
+(9 rows)
DROP TABLE ex_msource, ex_mtarget;
DROP FUNCTION explain_merge(text);
diff --git a/src/test/regress/expected/partition_aggregate.out b/src/test/regress/expected/partition_aggregate.out
index 1b900fddf8..800114dfde 100644
--- a/src/test/regress/expected/partition_aggregate.out
+++ b/src/test/regress/expected/partition_aggregate.out
@@ -948,12 +948,12 @@ SET parallel_setup_cost = 0;
-- is not partial agg safe.
EXPLAIN (COSTS OFF)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3 ORDER BY 1, 2, 3;
- QUERY PLAN
---------------------------------------------------------------------------------------
- Sort
- Sort Key: pagg_tab_ml.a, (sum(pagg_tab_ml.b)), (array_agg(DISTINCT pagg_tab_ml.c))
- -> Gather
- Workers Planned: 2
+ QUERY PLAN
+--------------------------------------------------------------------------------------------
+ Gather Merge
+ Workers Planned: 2
+ -> Sort
+ Sort Key: pagg_tab_ml.a, (sum(pagg_tab_ml.b)), (array_agg(DISTINCT pagg_tab_ml.c))
-> Parallel Append
-> GroupAggregate
Group Key: pagg_tab_ml.a
@@ -1380,28 +1380,26 @@ SELECT x, sum(y), avg(y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) <
-- When GROUP BY clause does not match; partial aggregation is performed for each partition.
EXPLAIN (COSTS OFF)
SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) < 12 ORDER BY 1, 2, 3;
- QUERY PLAN
--------------------------------------------------------------------------------------------
+ QUERY PLAN
+-------------------------------------------------------------------------------------
Sort
Sort Key: pagg_tab_para.y, (sum(pagg_tab_para.x)), (avg(pagg_tab_para.x))
- -> Finalize GroupAggregate
+ -> Finalize HashAggregate
Group Key: pagg_tab_para.y
Filter: (avg(pagg_tab_para.x) < '12'::numeric)
- -> Gather Merge
+ -> Gather
Workers Planned: 2
- -> Sort
- Sort Key: pagg_tab_para.y
- -> Parallel Append
- -> Partial HashAggregate
- Group Key: pagg_tab_para.y
- -> Parallel Seq Scan on pagg_tab_para_p1 pagg_tab_para
- -> Partial HashAggregate
- Group Key: pagg_tab_para_1.y
- -> Parallel Seq Scan on pagg_tab_para_p2 pagg_tab_para_1
- -> Partial HashAggregate
- Group Key: pagg_tab_para_2.y
- -> Parallel Seq Scan on pagg_tab_para_p3 pagg_tab_para_2
-(19 rows)
+ -> Parallel Append
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para.y
+ -> Parallel Seq Scan on pagg_tab_para_p1 pagg_tab_para
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para_1.y
+ -> Parallel Seq Scan on pagg_tab_para_p2 pagg_tab_para_1
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para_2.y
+ -> Parallel Seq Scan on pagg_tab_para_p3 pagg_tab_para_2
+(17 rows)
SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) < 12 ORDER BY 1, 2, 3;
y | sum | avg | count
diff --git a/src/test/regress/expected/partition_join.out b/src/test/regress/expected/partition_join.out
index 762cc8e53b..a1add421f5 100644
--- a/src/test/regress/expected/partition_join.out
+++ b/src/test/regress/expected/partition_join.out
@@ -466,52 +466,41 @@ EXPLAIN (COSTS OFF)
SELECT a, b FROM prt1 FULL JOIN prt2 p2(b,a,c) USING(a,b)
WHERE a BETWEEN 490 AND 510
GROUP BY 1, 2 ORDER BY 1, 2;
- QUERY PLAN
------------------------------------------------------------------------------------------------------------------
+ QUERY PLAN
+-----------------------------------------------------------------------------------------------------------
Group
Group Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Merge Append
+ -> Sort
Sort Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Group
- Group Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Sort
- Sort Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Merge Full Join
- Merge Cond: ((prt1.a = p2.a) AND (prt1.b = p2.b))
- Filter: ((COALESCE(prt1.a, p2.a) >= 490) AND (COALESCE(prt1.a, p2.a) <= 510))
- -> Sort
- Sort Key: prt1.a, prt1.b
- -> Seq Scan on prt1_p1 prt1
- -> Sort
- Sort Key: p2.a, p2.b
- -> Seq Scan on prt2_p1 p2
- -> Group
- Group Key: (COALESCE(prt1_1.a, p2_1.a)), (COALESCE(prt1_1.b, p2_1.b))
- -> Sort
- Sort Key: (COALESCE(prt1_1.a, p2_1.a)), (COALESCE(prt1_1.b, p2_1.b))
- -> Merge Full Join
- Merge Cond: ((prt1_1.a = p2_1.a) AND (prt1_1.b = p2_1.b))
- Filter: ((COALESCE(prt1_1.a, p2_1.a) >= 490) AND (COALESCE(prt1_1.a, p2_1.a) <= 510))
- -> Sort
- Sort Key: prt1_1.a, prt1_1.b
- -> Seq Scan on prt1_p2 prt1_1
- -> Sort
- Sort Key: p2_1.a, p2_1.b
- -> Seq Scan on prt2_p2 p2_1
- -> Group
- Group Key: (COALESCE(prt1_2.a, p2_2.a)), (COALESCE(prt1_2.b, p2_2.b))
- -> Sort
- Sort Key: (COALESCE(prt1_2.a, p2_2.a)), (COALESCE(prt1_2.b, p2_2.b))
- -> Merge Full Join
- Merge Cond: ((prt1_2.a = p2_2.a) AND (prt1_2.b = p2_2.b))
- Filter: ((COALESCE(prt1_2.a, p2_2.a) >= 490) AND (COALESCE(prt1_2.a, p2_2.a) <= 510))
- -> Sort
- Sort Key: prt1_2.a, prt1_2.b
- -> Seq Scan on prt1_p3 prt1_2
- -> Sort
- Sort Key: p2_2.a, p2_2.b
- -> Seq Scan on prt2_p3 p2_2
-(43 rows)
+ -> Append
+ -> Merge Full Join
+ Merge Cond: ((prt1_1.a = p2_1.a) AND (prt1_1.b = p2_1.b))
+ Filter: ((COALESCE(prt1_1.a, p2_1.a) >= 490) AND (COALESCE(prt1_1.a, p2_1.a) <= 510))
+ -> Sort
+ Sort Key: prt1_1.a, prt1_1.b
+ -> Seq Scan on prt1_p1 prt1_1
+ -> Sort
+ Sort Key: p2_1.a, p2_1.b
+ -> Seq Scan on prt2_p1 p2_1
+ -> Merge Full Join
+ Merge Cond: ((prt1_2.a = p2_2.a) AND (prt1_2.b = p2_2.b))
+ Filter: ((COALESCE(prt1_2.a, p2_2.a) >= 490) AND (COALESCE(prt1_2.a, p2_2.a) <= 510))
+ -> Sort
+ Sort Key: prt1_2.a, prt1_2.b
+ -> Seq Scan on prt1_p2 prt1_2
+ -> Sort
+ Sort Key: p2_2.a, p2_2.b
+ -> Seq Scan on prt2_p2 p2_2
+ -> Merge Full Join
+ Merge Cond: ((prt1_3.a = p2_3.a) AND (prt1_3.b = p2_3.b))
+ Filter: ((COALESCE(prt1_3.a, p2_3.a) >= 490) AND (COALESCE(prt1_3.a, p2_3.a) <= 510))
+ -> Sort
+ Sort Key: prt1_3.a, prt1_3.b
+ -> Seq Scan on prt1_p3 prt1_3
+ -> Sort
+ Sort Key: p2_3.a, p2_3.b
+ -> Seq Scan on prt2_p3 p2_3
+(32 rows)
SELECT a, b FROM prt1 FULL JOIN prt2 p2(b,a,c) USING(a,b)
WHERE a BETWEEN 490 AND 510
diff --git a/src/test/regress/expected/sysviews.out b/src/test/regress/expected/sysviews.out
index 001c6e7eb9..d5e9a35096 100644
--- a/src/test/regress/expected/sysviews.out
+++ b/src/test/regress/expected/sysviews.out
@@ -114,6 +114,7 @@ select name, setting from pg_settings where name like 'enable%';
enable_async_append | on
enable_bitmapscan | on
enable_gathermerge | on
+ enable_group_by_reordering | on
enable_hashagg | on
enable_hashjoin | on
enable_incremental_sort | on
@@ -132,7 +133,7 @@ select name, setting from pg_settings where name like 'enable%';
enable_seqscan | on
enable_sort | on
enable_tidscan | on
-(21 rows)
+(22 rows)
-- Test that the pg_timezone_names and pg_timezone_abbrevs views are
-- more-or-less working. We can't test their contents in any great detail
diff --git a/src/test/regress/expected/union.out b/src/test/regress/expected/union.out
index e2613d6777..220f0c1629 100644
--- a/src/test/regress/expected/union.out
+++ b/src/test/regress/expected/union.out
@@ -1303,24 +1303,22 @@ select distinct q1 from
union all
select distinct * from int8_tbl i82) ss
where q2 = q2;
- QUERY PLAN
-----------------------------------------------------------
- Unique
- -> Merge Append
- Sort Key: "*SELECT* 1".q1
+ QUERY PLAN
+----------------------------------------------------
+ HashAggregate
+ Group Key: "*SELECT* 1".q1
+ -> Append
-> Subquery Scan on "*SELECT* 1"
- -> Unique
- -> Sort
- Sort Key: i81.q1, i81.q2
- -> Seq Scan on int8_tbl i81
- Filter: (q2 IS NOT NULL)
+ -> HashAggregate
+ Group Key: i81.q1, i81.q2
+ -> Seq Scan on int8_tbl i81
+ Filter: (q2 IS NOT NULL)
-> Subquery Scan on "*SELECT* 2"
- -> Unique
- -> Sort
- Sort Key: i82.q1, i82.q2
- -> Seq Scan on int8_tbl i82
- Filter: (q2 IS NOT NULL)
-(15 rows)
+ -> HashAggregate
+ Group Key: i82.q1, i82.q2
+ -> Seq Scan on int8_tbl i82
+ Filter: (q2 IS NOT NULL)
+(13 rows)
select distinct q1 from
(select distinct * from int8_tbl i81
@@ -1339,24 +1337,22 @@ select distinct q1 from
union all
select distinct * from int8_tbl i82) ss
where -q1 = q2;
- QUERY PLAN
---------------------------------------------------------
- Unique
- -> Merge Append
- Sort Key: "*SELECT* 1".q1
+ QUERY PLAN
+--------------------------------------------------
+ HashAggregate
+ Group Key: "*SELECT* 1".q1
+ -> Append
-> Subquery Scan on "*SELECT* 1"
- -> Unique
- -> Sort
- Sort Key: i81.q1, i81.q2
- -> Seq Scan on int8_tbl i81
- Filter: ((- q1) = q2)
+ -> HashAggregate
+ Group Key: i81.q1, i81.q2
+ -> Seq Scan on int8_tbl i81
+ Filter: ((- q1) = q2)
-> Subquery Scan on "*SELECT* 2"
- -> Unique
- -> Sort
- Sort Key: i82.q1, i82.q2
- -> Seq Scan on int8_tbl i82
- Filter: ((- q1) = q2)
-(15 rows)
+ -> HashAggregate
+ Group Key: i82.q1, i82.q2
+ -> Seq Scan on int8_tbl i82
+ Filter: ((- q1) = q2)
+(13 rows)
select distinct q1 from
(select distinct * from int8_tbl i81
diff --git a/src/test/regress/sql/aggregates.sql b/src/test/regress/sql/aggregates.sql
index 75c78be640..5c4e768825 100644
--- a/src/test/regress/sql/aggregates.sql
+++ b/src/test/regress/sql/aggregates.sql
@@ -1172,6 +1172,105 @@ SELECT balk(hundred) FROM tenk1;
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+
+SELECT
+ i AS id,
+ i/2 AS p,
+ format('%60s', i%2) AS v,
+ i/4 AS c,
+ i/8 AS d,
+ (random() * (10000/8))::int as e --the same as d but no correlation with p
+ INTO btg
+FROM
+ generate_series(1, 10000) i;
+
+VACUUM btg;
+ANALYZE btg;
+
+-- GROUP BY optimization by reorder columns by frequency
+
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+
+CREATE STATISTICS btg_dep ON d, e, p FROM btg;
+ANALYZE btg;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+
+
+-- GROUP BY optimization by reorder columns by index scan
+
+CREATE INDEX ON btg(p, v);
+SET enable_seqscan=off;
+SET enable_bitmapscan=off;
+VACUUM btg;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d ORDER BY p, v;
+
+DROP TABLE btg;
+
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+RESET enable_seqscan;
+RESET enable_bitmapscan;
+
+
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/sql/incremental_sort.sql b/src/test/regress/sql/incremental_sort.sql
index ab471bdfff..70f311a6f2 100644
--- a/src/test/regress/sql/incremental_sort.sql
+++ b/src/test/regress/sql/incremental_sort.sql
@@ -208,7 +208,7 @@ set parallel_tuple_cost = 0;
set max_parallel_workers_per_gather = 2;
create table t (a int, b int, c int);
-insert into t select mod(i,10),mod(i,10),i from generate_series(1,10000) s(i);
+insert into t select mod(i,10),mod(i,10),i from generate_series(1,60000) s(i);
create index on t (a);
analyze t;
--
2.41.0
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2023-07-20 11:46 Tomas Vondra <[email protected]>
parent: Andrey Lepikhov <[email protected]>
0 siblings, 4 replies; 38+ messages in thread
From: Tomas Vondra @ 2023-07-20 11:46 UTC (permalink / raw)
To: Andrey Lepikhov <[email protected]>; PostgreSQL Developers <[email protected]>; Tom Lane <[email protected]>; +Cc: Zhihong Yu <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; David Rowley <[email protected]>; a.rybakina <[email protected]>; Белялов Дамир Наилевич <[email protected]>
On 7/20/23 08:37, Andrey Lepikhov wrote:
> On 3/10/2022 21:56, Tom Lane wrote:
>> Revert "Optimize order of GROUP BY keys".
>>
>> This reverts commit db0d67db2401eb6238ccc04c6407a4fd4f985832 and
>> several follow-on fixes.
>> ...
>> Since we're hard up against the release deadline for v15, let's
>> revert these changes for now. We can always try again later.
>
> It may be time to restart the project. As a first step, I rebased the
> patch on the current master. It wasn't trivial because of some latest
> optimizations (a29eab, 1349d27 and 8d83a5d).
> Now, Let's repeat the review and rewrite the current path according to
> the reasons uttered in the revert commit.
I think the fundamental task is to make the costing more reliable, and
the commit message 443df6e2db points out a couple challenges in this
area. Not sure how feasible it is to address enough of them ...
1) procost = 1.0 - I guess we could make this more realistic by doing
some microbenchmarks and tuning the costs for the most expensive cases.
2) estimating quicksort comparisons - This relies on ndistinct
estimates, and I'm not sure how much more reliable we can make those.
Probably not much :-( Not sure what to do about this, the only thing I
can think of is to track "reliability" of the estimates and only do the
reordering if we have high confidence in the estimates. That means we'll
miss some optimization opportunities, but it should limit the risk.
regards
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2023-07-21 04:21 Andrey Lepikhov <[email protected]>
parent: Tomas Vondra <[email protected]>
3 siblings, 0 replies; 38+ messages in thread
From: Andrey Lepikhov @ 2023-07-21 04:21 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>; Tom Lane <[email protected]>; +Cc: Zhihong Yu <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; David Rowley <[email protected]>; a.rybakina <[email protected]>; Белялов Дамир Наилевич <[email protected]>
On 20/7/2023 18:46, Tomas Vondra wrote:
> On 7/20/23 08:37, Andrey Lepikhov wrote:
>> On 3/10/2022 21:56, Tom Lane wrote:
>>> Revert "Optimize order of GROUP BY keys".
>>>
>>> This reverts commit db0d67db2401eb6238ccc04c6407a4fd4f985832 and
>>> several follow-on fixes.
>>> ...
>>> Since we're hard up against the release deadline for v15, let's
>>> revert these changes for now. We can always try again later.
>>
>> It may be time to restart the project. As a first step, I rebased the
>> patch on the current master. It wasn't trivial because of some latest
>> optimizations (a29eab, 1349d27 and 8d83a5d).
>> Now, Let's repeat the review and rewrite the current path according to
>> the reasons uttered in the revert commit.
>
> I think the fundamental task is to make the costing more reliable, and
> the commit message 443df6e2db points out a couple challenges in this
> area. Not sure how feasible it is to address enough of them ...
>
> 1) procost = 1.0 - I guess we could make this more realistic by doing
> some microbenchmarks and tuning the costs for the most expensive cases.
>
> 2) estimating quicksort comparisons - This relies on ndistinct
> estimates, and I'm not sure how much more reliable we can make those.
> Probably not much :-( Not sure what to do about this, the only thing I
> can think of is to track "reliability" of the estimates and only do the
> reordering if we have high confidence in the estimates. That means we'll
> miss some optimization opportunities, but it should limit the risk.
For me personally, the most challenging issue is:
3. Imbalance, induced by the cost_sort() changes. It may increase or
decrease the contribution of sorting to the total cost compared to other
factors and change choice of sorted paths.
--
regards,
Andrey Lepikhov
Postgres Professional
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2023-07-24 02:10 Andrey Lepikhov <[email protected]>
parent: Tomas Vondra <[email protected]>
3 siblings, 1 reply; 38+ messages in thread
From: Andrey Lepikhov @ 2023-07-24 02:10 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>; Tom Lane <[email protected]>; +Cc: Zhihong Yu <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; David Rowley <[email protected]>; a.rybakina <[email protected]>; Белялов Дамир Наилевич <[email protected]>
On 20/7/2023 18:46, Tomas Vondra wrote:
> On 7/20/23 08:37, Andrey Lepikhov wrote:
>> On 3/10/2022 21:56, Tom Lane wrote:
>>> Revert "Optimize order of GROUP BY keys".
>>>
>>> This reverts commit db0d67db2401eb6238ccc04c6407a4fd4f985832 and
>>> several follow-on fixes.
>>> ...
>>> Since we're hard up against the release deadline for v15, let's
>>> revert these changes for now. We can always try again later.
>>
>> It may be time to restart the project. As a first step, I rebased the
>> patch on the current master. It wasn't trivial because of some latest
>> optimizations (a29eab, 1349d27 and 8d83a5d).
>> Now, Let's repeat the review and rewrite the current path according to
>> the reasons uttered in the revert commit.
>
> I think the fundamental task is to make the costing more reliable, and
> the commit message 443df6e2db points out a couple challenges in this
> area. Not sure how feasible it is to address enough of them ...
>
> 1) procost = 1.0 - I guess we could make this more realistic by doing
> some microbenchmarks and tuning the costs for the most expensive cases.
>
> 2) estimating quicksort comparisons - This relies on ndistinct
> estimates, and I'm not sure how much more reliable we can make those.
> Probably not much :-( Not sure what to do about this, the only thing I
> can think of is to track "reliability" of the estimates and only do the
> reordering if we have high confidence in the estimates. That means we'll
> miss some optimization opportunities, but it should limit the risk.
I read up on the history of this thread.
As I see, all the problems mentioned above can be beaten by excluding
the new cost model at all. We can sort GROUP BY columns according to the
'ndistinct' value.
I see the reason for introducing the cost model in [1]. The main
supporting point here is that with this patch, people couldn't optimize
the query by themselves, organizing the order of the columns in a more
optimal way. But now we have at least the GUC to switch off the
behaviour introduced here. Also, some extensions, like the well-known
pg_hint_plan, can help with automation.
So, how about committing of the undoubted part of the feature and
working on the cost model in a new thread?
[1]
https://www.postgresql.org/message-id/6d1e0cdb-dde3-f62a-43e2-e90bbd9b0f42%402ndquadrant.com
--
regards,
Andrey Lepikhov
Postgres Professional
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2023-07-24 09:56 Tomas Vondra <[email protected]>
parent: Andrey Lepikhov <[email protected]>
0 siblings, 1 reply; 38+ messages in thread
From: Tomas Vondra @ 2023-07-24 09:56 UTC (permalink / raw)
To: Andrey Lepikhov <[email protected]>; PostgreSQL Developers <[email protected]>; Tom Lane <[email protected]>; +Cc: Zhihong Yu <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; David Rowley <[email protected]>; a.rybakina <[email protected]>; Белялов Дамир Наилевич <[email protected]>
On 7/24/23 04:10, Andrey Lepikhov wrote:
> On 20/7/2023 18:46, Tomas Vondra wrote:
>> On 7/20/23 08:37, Andrey Lepikhov wrote:
>>> On 3/10/2022 21:56, Tom Lane wrote:
>>>> Revert "Optimize order of GROUP BY keys".
>>>>
>>>> This reverts commit db0d67db2401eb6238ccc04c6407a4fd4f985832 and
>>>> several follow-on fixes.
>>>> ...
>>>> Since we're hard up against the release deadline for v15, let's
>>>> revert these changes for now. We can always try again later.
>>>
>>> It may be time to restart the project. As a first step, I rebased the
>>> patch on the current master. It wasn't trivial because of some latest
>>> optimizations (a29eab, 1349d27 and 8d83a5d).
>>> Now, Let's repeat the review and rewrite the current path according to
>>> the reasons uttered in the revert commit.
>>
>> I think the fundamental task is to make the costing more reliable, and
>> the commit message 443df6e2db points out a couple challenges in this
>> area. Not sure how feasible it is to address enough of them ...
>>
>> 1) procost = 1.0 - I guess we could make this more realistic by doing
>> some microbenchmarks and tuning the costs for the most expensive cases.
>>
>> 2) estimating quicksort comparisons - This relies on ndistinct
>> estimates, and I'm not sure how much more reliable we can make those.
>> Probably not much :-( Not sure what to do about this, the only thing I
>> can think of is to track "reliability" of the estimates and only do the
>> reordering if we have high confidence in the estimates. That means we'll
>> miss some optimization opportunities, but it should limit the risk.
> I read up on the history of this thread.
> As I see, all the problems mentioned above can be beaten by excluding
> the new cost model at all. We can sort GROUP BY columns according to the
> 'ndistinct' value.
> I see the reason for introducing the cost model in [1]. The main
> supporting point here is that with this patch, people couldn't optimize
> the query by themselves, organizing the order of the columns in a more
> optimal way. But now we have at least the GUC to switch off the
> behaviour introduced here. Also, some extensions, like the well-known
> pg_hint_plan, can help with automation.
I think the main concern is that if we reorder the group keys and get it
wrong, it's a regression. If that happens often (due to costing based on
poor stats), it's a problem. Yes, there's a GUC, but that's a rather
blunt instrument, unfortunately.
> So, how about committing of the undoubted part of the feature and
> working on the cost model in a new thread?
>
But Tom's commit message says this:
Worse, to arrive at estimates of the number of calls made to the
lower-order-column comparison functions, the code needs to make
estimates of the numbers of distinct values of multiple columns,
which are necessarily even less trustworthy than per-column stats.
so I'm not sure this really counts as "undoubted".
regards
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2023-07-24 12:04 Andrey Lepikhov <[email protected]>
parent: Tomas Vondra <[email protected]>
0 siblings, 0 replies; 38+ messages in thread
From: Andrey Lepikhov @ 2023-07-24 12:04 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>; Tom Lane <[email protected]>; +Cc: Zhihong Yu <[email protected]>; Teodor Sigaev <[email protected]>; Tomas Vondra <[email protected]>; David Rowley <[email protected]>; a.rybakina <[email protected]>; Белялов Дамир Наилевич <[email protected]>
On 24/7/2023 16:56, Tomas Vondra wrote:
> On 7/24/23 04:10, Andrey Lepikhov wrote:
>> On 20/7/2023 18:46, Tomas Vondra wrote:
>>> On 7/20/23 08:37, Andrey Lepikhov wrote:
>>>> On 3/10/2022 21:56, Tom Lane wrote:
>>>>> Revert "Optimize order of GROUP BY keys".
>>>>>
>>>>> This reverts commit db0d67db2401eb6238ccc04c6407a4fd4f985832 and
>>>>> several follow-on fixes.
>>>>> ...
>>>>> Since we're hard up against the release deadline for v15, let's
>>>>> revert these changes for now. We can always try again later.
>>>>
>>>> It may be time to restart the project. As a first step, I rebased the
>>>> patch on the current master. It wasn't trivial because of some latest
>>>> optimizations (a29eab, 1349d27 and 8d83a5d).
>>>> Now, Let's repeat the review and rewrite the current path according to
>>>> the reasons uttered in the revert commit.
>>>
>>> I think the fundamental task is to make the costing more reliable, and
>>> the commit message 443df6e2db points out a couple challenges in this
>>> area. Not sure how feasible it is to address enough of them ...
>>>
>>> 1) procost = 1.0 - I guess we could make this more realistic by doing
>>> some microbenchmarks and tuning the costs for the most expensive cases.
>>>
>>> 2) estimating quicksort comparisons - This relies on ndistinct
>>> estimates, and I'm not sure how much more reliable we can make those.
>>> Probably not much :-( Not sure what to do about this, the only thing I
>>> can think of is to track "reliability" of the estimates and only do the
>>> reordering if we have high confidence in the estimates. That means we'll
>>> miss some optimization opportunities, but it should limit the risk.
>> I read up on the history of this thread.
>> As I see, all the problems mentioned above can be beaten by excluding
>> the new cost model at all. We can sort GROUP BY columns according to the
>> 'ndistinct' value.
>> I see the reason for introducing the cost model in [1]. The main
>> supporting point here is that with this patch, people couldn't optimize
>> the query by themselves, organizing the order of the columns in a more
>> optimal way. But now we have at least the GUC to switch off the
>> behaviour introduced here. Also, some extensions, like the well-known
>> pg_hint_plan, can help with automation.
>
> I think the main concern is that if we reorder the group keys and get it
> wrong, it's a regression. If that happens often (due to costing based on
> poor stats), it's a problem. Yes, there's a GUC, but that's a rather
> blunt instrument, unfortunately.
I see. My point here is if the ndistinct of one column is much more than
the ndistinct of another, it is more probable that this correlation will
be kept in the grouping phase. Here we can get some regression, which
can be overweighed by the possibility below.
>
>> So, how about committing of the undoubted part of the feature and
>> working on the cost model in a new thread?
>>
>
> But Tom's commit message says this:
>
> Worse, to arrive at estimates of the number of calls made to the
> lower-order-column comparison functions, the code needs to make
> estimates of the numbers of distinct values of multiple columns,
> which are necessarily even less trustworthy than per-column stats.
>
> so I'm not sure this really counts as "undoubted".
Don't try to estimate multiple columns - just sort columns according to
the value of ndistinct as a heuristic.
I think we should estimate the number of values of multiple columns only
if we have extended statistics on these columns. And this can extend the
applicability of extended statistics.
The suggestion above is just an attempt to gather low-hanging fruit
right now. If it is not applicable, we will go a long way.
--
regards,
Andrey Lepikhov
Postgres Professional
^ permalink raw reply [nested|flat] 38+ messages in thread
* [PATCH v10 4/4] Remove open-coded binary heap in pg_dump_sort.c.
@ 2023-07-25 18:18 Nathan Bossart <[email protected]>
0 siblings, 0 replies; 38+ messages in thread
From: Nathan Bossart @ 2023-07-25 18:18 UTC (permalink / raw)
Thanks to commit XXXXXXXXXX, binaryheap is available to frontend
code. This commit replaces the open-coded heap implementation in
pg_dump_sort.c with a binaryheap, saving a few lines of code.
Reviewed-by: Tom Lane
Discussion: https://postgr.es/m/3612876.1689443232%40sss.pgh.pa.us
---
src/bin/pg_dump/pg_dump_sort.c | 105 +++++++--------------------------
1 file changed, 22 insertions(+), 83 deletions(-)
diff --git a/src/bin/pg_dump/pg_dump_sort.c b/src/bin/pg_dump/pg_dump_sort.c
index 523a19c155..8436c1d757 100644
--- a/src/bin/pg_dump/pg_dump_sort.c
+++ b/src/bin/pg_dump/pg_dump_sort.c
@@ -16,6 +16,7 @@
#include "postgres_fe.h"
#include "catalog/pg_class_d.h"
+#include "lib/binaryheap.h"
#include "pg_backup_archiver.h"
#include "pg_backup_utils.h"
#include "pg_dump.h"
@@ -161,8 +162,6 @@ static bool TopoSort(DumpableObject **objs,
int numObjs,
DumpableObject **ordering,
int *nOrdering);
-static void addHeapElement(int val, int *heap, int heapLength);
-static int removeHeapElement(int *heap, int heapLength);
static void findDependencyLoops(DumpableObject **objs, int nObjs, int totObjs);
static int findLoop(DumpableObject *obj,
DumpId startPoint,
@@ -174,6 +173,7 @@ static void repairDependencyLoop(DumpableObject **loop,
int nLoop);
static void describeDumpableObject(DumpableObject *obj,
char *buf, int bufsize);
+static int int_cmp(void *a, void *b, void *arg);
/*
@@ -374,11 +374,10 @@ TopoSort(DumpableObject **objs,
int *nOrdering) /* output argument */
{
DumpId maxDumpId = getMaxDumpId();
- int *pendingHeap;
+ binaryheap *pendingHeap;
int *beforeConstraints;
int *idMap;
DumpableObject *obj;
- int heapLength;
int i,
j,
k;
@@ -403,7 +402,7 @@ TopoSort(DumpableObject **objs,
return true;
/* Create workspace for the above-described heap */
- pendingHeap = (int *) pg_malloc(numObjs * sizeof(int));
+ pendingHeap = binaryheap_allocate(numObjs, int_cmp, NULL);
/*
* Scan the constraints, and for each item in the input, generate a count
@@ -434,19 +433,16 @@ TopoSort(DumpableObject **objs,
* Now initialize the heap of items-ready-to-output by filling it with the
* indexes of items that already have beforeConstraints[id] == 0.
*
- * The essential property of a heap is heap[(j-1)/2] >= heap[j] for each j
- * in the range 1..heapLength-1 (note we are using 0-based subscripts
- * here, while the discussion in Knuth assumes 1-based subscripts). So, if
- * we simply enter the indexes into pendingHeap[] in decreasing order, we
- * a-fortiori have the heap invariant satisfied at completion of this
- * loop, and don't need to do any sift-up comparisons.
+ * We enter the objects into pendingHeap in decreasing order so that the
+ * heap invariant is satisfied at the completion of this loop. This
+ * reduces the amount of work that binaryheap_build() must do.
*/
- heapLength = 0;
for (i = numObjs; --i >= 0;)
{
if (beforeConstraints[objs[i]->dumpId] == 0)
- pendingHeap[heapLength++] = i;
+ binaryheap_add_unordered(pendingHeap, (void *) (intptr_t) i);
}
+ binaryheap_build(pendingHeap);
/*--------------------
* Now emit objects, working backwards in the output list. At each step,
@@ -464,10 +460,10 @@ TopoSort(DumpableObject **objs,
*--------------------
*/
i = numObjs;
- while (heapLength > 0)
+ while (!binaryheap_empty(pendingHeap))
{
/* Select object to output by removing largest heap member */
- j = removeHeapElement(pendingHeap, heapLength--);
+ j = (int) (intptr_t) binaryheap_remove_first(pendingHeap);
obj = objs[j];
/* Output candidate to ordering[] */
ordering[--i] = obj;
@@ -477,7 +473,7 @@ TopoSort(DumpableObject **objs,
int id = obj->dependencies[k];
if ((--beforeConstraints[id]) == 0)
- addHeapElement(idMap[id], pendingHeap, heapLength++);
+ binaryheap_add(pendingHeap, (void *) (intptr_t) idMap[id]);
}
}
@@ -497,79 +493,13 @@ TopoSort(DumpableObject **objs,
}
/* Done */
- free(pendingHeap);
+ binaryheap_free(pendingHeap);
free(beforeConstraints);
free(idMap);
return (i == 0);
}
-/*
- * Add an item to a heap (priority queue)
- *
- * heapLength is the current heap size; caller is responsible for increasing
- * its value after the call. There must be sufficient storage at *heap.
- */
-static void
-addHeapElement(int val, int *heap, int heapLength)
-{
- int j;
-
- /*
- * Sift-up the new entry, per Knuth 5.2.3 exercise 16. Note that Knuth is
- * using 1-based array indexes, not 0-based.
- */
- j = heapLength;
- while (j > 0)
- {
- int i = (j - 1) >> 1;
-
- if (val <= heap[i])
- break;
- heap[j] = heap[i];
- j = i;
- }
- heap[j] = val;
-}
-
-/*
- * Remove the largest item present in a heap (priority queue)
- *
- * heapLength is the current heap size; caller is responsible for decreasing
- * its value after the call.
- *
- * We remove and return heap[0], which is always the largest element of
- * the heap, and then "sift up" to maintain the heap invariant.
- */
-static int
-removeHeapElement(int *heap, int heapLength)
-{
- int result = heap[0];
- int val;
- int i;
-
- if (--heapLength <= 0)
- return result;
- val = heap[heapLength]; /* value that must be reinserted */
- i = 0; /* i is where the "hole" is */
- for (;;)
- {
- int j = 2 * i + 1;
-
- if (j >= heapLength)
- break;
- if (j + 1 < heapLength &&
- heap[j] < heap[j + 1])
- j++;
- if (val >= heap[j])
- break;
- heap[i] = heap[j];
- i = j;
- }
- heap[i] = val;
- return result;
-}
-
/*
* findDependencyLoops - identify loops in TopoSort's failure output,
* and pass each such loop to repairDependencyLoop() for action
@@ -1559,3 +1489,12 @@ describeDumpableObject(DumpableObject *obj, char *buf, int bufsize)
(int) obj->objType,
obj->dumpId, obj->catId.oid);
}
+
+/*
+ * Compares "a" and "b" as integers.
+ */
+static int
+int_cmp(void *a, void *b, void *arg)
+{
+ return (int) (intptr_t) a - (int) (intptr_t) b;
+}
--
2.25.1
--y0ulUmNC+osPPQO6--
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2023-09-19 04:42 Andrey Lepikhov <[email protected]>
parent: Tomas Vondra <[email protected]>
3 siblings, 0 replies; 38+ messages in thread
From: Andrey Lepikhov @ 2023-09-19 04:42 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; +Cc: Teodor Sigaev <[email protected]>; PostgreSQL Developers <[email protected]>; David Rowley <[email protected]>; a.rybakina <[email protected]>; Tom Lane <[email protected]>; Белялов Дамир Наилевич <[email protected]>
On 20/7/2023 18:46, Tomas Vondra wrote:
> On 7/20/23 08:37, Andrey Lepikhov wrote:
>> On 3/10/2022 21:56, Tom Lane wrote:
>>> Revert "Optimize order of GROUP BY keys".
>>>
>>> This reverts commit db0d67db2401eb6238ccc04c6407a4fd4f985832 and
>>> several follow-on fixes.
>>> ...
>>> Since we're hard up against the release deadline for v15, let's
>>> revert these changes for now. We can always try again later.
>>
>> It may be time to restart the project. As a first step, I rebased the
>> patch on the current master. It wasn't trivial because of some latest
>> optimizations (a29eab, 1349d27 and 8d83a5d).
>> Now, Let's repeat the review and rewrite the current path according to
>> the reasons uttered in the revert commit.
> 1) procost = 1.0 - I guess we could make this more realistic by doing
> some microbenchmarks and tuning the costs for the most expensive cases.
Ok, some thoughts on this part of the task. As I see, we have not so
many different operators: 26 with fixed width and 16 with variable width:
SELECT o.oid,oprcode,typname,typlen FROM pg_operator o
JOIN pg_type t ON (oprleft = t.oid)
WHERE (oprname='<') AND oprleft=oprright AND typlen>0
ORDER BY o.oid;
SELECT o.oid,oprcode,typname,typlen FROM pg_operator o
JOIN pg_type t ON (oprleft = t.oid)
WHERE (oprname='<') AND oprleft=oprright AND typlen<0
ORDER BY o.oid;
Benchmarking procedure of types with fixed length can be something like
below:
CREATE OR REPLACE FUNCTION pass_sort(typ regtype) RETURNS TABLE (
nrows integer,
exec_time float
) AS $$
DECLARE
data json;
step integer;
BEGIN
SET work_mem='1GB';
FOR step IN 0..3 LOOP
SELECT pow(100, step)::integer INTO nrows;
DROP TABLE IF EXISTS test CASCADE;
EXECUTE format('CREATE TABLE test AS SELECT gs::%s AS x
FROM generate_series(1,%s) AS gs;', typ, nrows);
EXPLAIN (ANALYZE, COSTS OFF, TIMING OFF, FORMAT JSON)
SELECT * FROM test ORDER BY (x) DESC INTO data;
SELECT data->0->'Execution Time' INTO exec_time;
RETURN NEXT;
END LOOP;
END;
$$ LANGUAGE plpgsql VOLATILE;
Execution of SELECT * FROM pass_sort('integer'); shows quite linear grow
of the execution time. So, using '2.0 * cpu_operator_cost' as a cost for
the integer type (as a basis) we can calculate costs for other
operators. Variable-width types, i think, could require more complex
technique to check dependency on the length.
Does this way look worthwhile?
--
regards,
Andrey Lepikhov
Postgres Professional
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2023-09-26 05:51 Andrey Lepikhov <[email protected]>
parent: Tomas Vondra <[email protected]>
3 siblings, 1 reply; 38+ messages in thread
From: Andrey Lepikhov @ 2023-09-26 05:51 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; PostgreSQL Developers <[email protected]>; Tom Lane <[email protected]>; +Cc: Teodor Sigaev <[email protected]>; David Rowley <[email protected]>; a.rybakina <[email protected]>; Белялов Дамир Наилевич <[email protected]>
On 20/7/2023 18:46, Tomas Vondra wrote:
> 2) estimating quicksort comparisons - This relies on ndistinct
> estimates, and I'm not sure how much more reliable we can make those.
> Probably not much :-( Not sure what to do about this, the only thing I
> can think of is to track "reliability" of the estimates and only do the
> reordering if we have high confidence in the estimates. That means we'll
> miss some optimization opportunities, but it should limit the risk.
According to this issue, I see two options:
1. Go through the grouping column list and find the most reliable one.
If we have a unique column or column with statistics on the number of
distinct values, which is significantly more than ndistincts for other
grouping columns, we can place this column as the first in the grouping.
It should guarantee the reliability of such a decision, isn't it?
2. If we have extended statistics on distinct values and these
statistics cover some set of first columns in the grouping list, we can
optimize these positions. It also looks reliable.
Any thoughts?
--
regards,
Andrey Lepikhov
Postgres Professional
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2023-10-01 06:32 Andrei Lepikhov <[email protected]>
parent: Andrey Lepikhov <[email protected]>
0 siblings, 1 reply; 38+ messages in thread
From: Andrei Lepikhov @ 2023-10-01 06:32 UTC (permalink / raw)
To: PostgreSQL Developers <[email protected]>; Tomas Vondra <[email protected]>; +Cc: Teodor Sigaev <[email protected]>; David Rowley <[email protected]>; a.rybakina <[email protected]>; Белялов Дамир Наилевич <[email protected]>; Tom Lane <[email protected]>
New version of the patch. Fixed minor inconsistencies and rebased onto
current master.
--
regards,
Andrey Lepikhov
Postgres Professional
From 2f5a42c8a53286f830e3376ff4d3f8b7d4215b4b Mon Sep 17 00:00:00 2001
From: "Andrey V. Lepikhov" <[email protected]>
Date: Wed, 13 Sep 2023 11:20:03 +0700
Subject: [PATCH] Explore alternative orderings of group-by pathkeys during
optimization.
When evaluating a query with a multi-column GROUP BY clause using sort,
the cost may depend heavily on the order in which the keys are compared when
building the groups. Grouping does not imply any ordering, so we can compare
the keys in arbitrary order, and a Hash Agg leverages this. But for Group Agg,
we simply compared keys in the order specified in the query. This commit
explores alternative ordering of the keys, trying to find a cheaper one.
In principle, we might generate grouping paths for all permutations of the keys
and leave the rest to the optimizer. But that might get very expensive, so we
try to pick only a couple interesting orderings based on both local and global
information.
When planning the grouping path, we explore statistics (number of distinct
values, cost of the comparison function) for the keys and reorder them
to minimize comparison costs. Intuitively, it may be better to perform more
expensive comparisons (for complex data types, etc.) last because maybe
the cheaper comparisons will be enough. Similarly, the higher the cardinality
of a key, the lower the probability we'll need to compare more keys. The patch
generates and costs various orderings, picking the cheapest ones.
The ordering of group keys may interact with other parts of the query, some of
which may not be known while planning the grouping. For example, there may be
an explicit ORDER BY clause or some other ordering-dependent operation higher up
in the query, and using the same ordering may allow using either incremental
sort or even eliminating the sort entirely.
The patch generates orderings and picks those, minimizing the comparison cost
(for various path keys), and then adds orderings that might be useful for
operations higher up in the plan (ORDER BY, etc.). Finally, it always keeps
the ordering specified in the query, assuming the user might have additional
insights.
This introduces a new GUC enable_group_by_reordering so that the optimization
may be disabled if needed.
---
.../postgres_fdw/expected/postgres_fdw.out | 36 +-
src/backend/optimizer/path/costsize.c | 363 ++++++++++-
src/backend/optimizer/path/equivclass.c | 13 +-
src/backend/optimizer/path/pathkeys.c | 600 ++++++++++++++++++
src/backend/optimizer/plan/planner.c | 465 ++++++++------
src/backend/optimizer/util/pathnode.c | 4 +-
src/backend/utils/adt/selfuncs.c | 38 +-
src/backend/utils/misc/guc_tables.c | 10 +
src/backend/utils/misc/postgresql.conf.sample | 1 +
src/include/nodes/pathnodes.h | 10 +
src/include/optimizer/cost.h | 4 +-
src/include/optimizer/paths.h | 7 +
src/include/utils/selfuncs.h | 5 +
src/test/regress/expected/aggregates.out | 244 ++++++-
.../regress/expected/incremental_sort.out | 2 +-
src/test/regress/expected/join.out | 51 +-
src/test/regress/expected/merge.out | 15 +-
.../regress/expected/partition_aggregate.out | 44 +-
src/test/regress/expected/partition_join.out | 75 +--
src/test/regress/expected/sysviews.out | 3 +-
src/test/regress/expected/union.out | 60 +-
src/test/regress/sql/aggregates.sql | 99 +++
src/test/regress/sql/incremental_sort.sql | 2 +-
23 files changed, 1769 insertions(+), 382 deletions(-)
diff --git a/contrib/postgres_fdw/expected/postgres_fdw.out b/contrib/postgres_fdw/expected/postgres_fdw.out
index 144c114d0f..63af7feabe 100644
--- a/contrib/postgres_fdw/expected/postgres_fdw.out
+++ b/contrib/postgres_fdw/expected/postgres_fdw.out
@@ -2319,18 +2319,21 @@ SELECT t1."C 1" FROM "S 1"."T 1" t1, LATERAL (SELECT DISTINCT t2.c1, t3.c1 FROM
-- join with pseudoconstant quals
EXPLAIN (VERBOSE, COSTS OFF)
SELECT t1.c1, t2.c1 FROM ft1 t1 JOIN ft2 t2 ON (t1.c1 = t2.c1 AND CURRENT_USER = SESSION_USER) ORDER BY t1.c3, t1.c1 OFFSET 100 LIMIT 10;
- QUERY PLAN
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+ QUERY PLAN
+-----------------------------------------------------------------------------------------------------------------------------------------------
Limit
Output: t1.c1, t2.c1, t1.c3
- -> Result
+ -> Sort
Output: t1.c1, t2.c1, t1.c3
- One-Time Filter: (CURRENT_USER = SESSION_USER)
- -> Foreign Scan
- Output: t1.c1, t1.c3, t2.c1
- Relations: (public.ft1 t1) INNER JOIN (public.ft2 t2)
- Remote SQL: SELECT r1."C 1", r2."C 1", r1.c3 FROM ("S 1"."T 1" r1 INNER JOIN "S 1"."T 1" r2 ON (((r2."C 1" = r1."C 1")))) ORDER BY r1.c3 ASC NULLS LAST, r1."C 1" ASC NULLS LAST
-(9 rows)
+ Sort Key: t1.c3, t1.c1
+ -> Result
+ Output: t1.c1, t2.c1, t1.c3
+ One-Time Filter: (CURRENT_USER = SESSION_USER)
+ -> Foreign Scan
+ Output: t1.c1, t1.c3, t2.c1
+ Relations: (public.ft1 t1) INNER JOIN (public.ft2 t2)
+ Remote SQL: SELECT r1."C 1", r2."C 1", r1.c3 FROM ("S 1"."T 1" r1 INNER JOIN "S 1"."T 1" r2 ON (((r2."C 1" = r1."C 1"))))
+(12 rows)
-- non-Var items in targetlist of the nullable rel of a join preventing
-- push-down in some cases
@@ -9609,13 +9612,16 @@ SELECT t1.a,t2.b,t3.c FROM fprt1 t1 INNER JOIN fprt2 t2 ON (t1.a = t2.b) INNER J
-- left outer join + nullable clause
EXPLAIN (VERBOSE, COSTS OFF)
SELECT t1.a,t2.b,t2.c FROM fprt1 t1 LEFT JOIN (SELECT * FROM fprt2 WHERE a < 10) t2 ON (t1.a = t2.b and t1.b = t2.a) WHERE t1.a < 10 ORDER BY 1,2,3;
- QUERY PLAN
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
- Foreign Scan
+ QUERY PLAN
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+ Sort
Output: t1.a, fprt2.b, fprt2.c
- Relations: (public.ftprt1_p1 t1) LEFT JOIN (public.ftprt2_p1 fprt2)
- Remote SQL: SELECT r5.a, r6.b, r6.c FROM (public.fprt1_p1 r5 LEFT JOIN public.fprt2_p1 r6 ON (((r5.a = r6.b)) AND ((r5.b = r6.a)) AND ((r6.a < 10)))) WHERE ((r5.a < 10)) ORDER BY r5.a ASC NULLS LAST, r6.b ASC NULLS LAST, r6.c ASC NULLS LAST
-(4 rows)
+ Sort Key: t1.a, fprt2.b, fprt2.c
+ -> Foreign Scan
+ Output: t1.a, fprt2.b, fprt2.c
+ Relations: (public.ftprt1_p1 t1) LEFT JOIN (public.ftprt2_p1 fprt2)
+ Remote SQL: SELECT r5.a, r6.b, r6.c FROM (public.fprt1_p1 r5 LEFT JOIN public.fprt2_p1 r6 ON (((r5.a = r6.b)) AND ((r5.b = r6.a)) AND ((r6.a < 10)))) WHERE ((r5.a < 10))
+(7 rows)
SELECT t1.a,t2.b,t2.c FROM fprt1 t1 LEFT JOIN (SELECT * FROM fprt2 WHERE a < 10) t2 ON (t1.a = t2.b and t1.b = t2.a) WHERE t1.a < 10 ORDER BY 1,2,3;
a | b | c
diff --git a/src/backend/optimizer/path/costsize.c b/src/backend/optimizer/path/costsize.c
index d6ceafd51c..6f11f679ba 100644
--- a/src/backend/optimizer/path/costsize.c
+++ b/src/backend/optimizer/path/costsize.c
@@ -1816,6 +1816,327 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
rterm->pathtarget->width);
}
+/*
+ * is_fake_var
+ * Workaround for generate_append_tlist() which generates fake Vars with
+ * varno == 0, that will cause a fail of estimate_num_group() call
+ *
+ * XXX Ummm, why would estimate_num_group fail with this?
+ */
+static bool
+is_fake_var(Expr *expr)
+{
+ if (IsA(expr, RelabelType))
+ expr = (Expr *) ((RelabelType *) expr)->arg;
+
+ return (IsA(expr, Var) && ((Var *) expr)->varno == 0);
+}
+
+/*
+ * get_width_cost_multiplier
+ * Returns relative complexity of comparing two values based on its width.
+ * The idea behind is that the comparison becomes more expensive the longer the
+ * value is. Return value is in cpu_operator_cost units.
+ */
+static double
+get_width_cost_multiplier(PlannerInfo *root, Expr *expr)
+{
+ double width = -1.0; /* fake value */
+
+ if (IsA(expr, RelabelType))
+ expr = (Expr *) ((RelabelType *) expr)->arg;
+
+ /* Try to find actual stat in corresponding relation */
+ if (IsA(expr, Var))
+ {
+ Var *var = (Var *) expr;
+
+ if (var->varno > 0 && var->varno < root->simple_rel_array_size)
+ {
+ RelOptInfo *rel = root->simple_rel_array[var->varno];
+
+ if (rel != NULL &&
+ var->varattno >= rel->min_attr &&
+ var->varattno <= rel->max_attr)
+ {
+ int ndx = var->varattno - rel->min_attr;
+
+ if (rel->attr_widths[ndx] > 0)
+ width = rel->attr_widths[ndx];
+ }
+ }
+ }
+
+ /* Didn't find any actual stats, try using type width instead. */
+ if (width < 0.0)
+ {
+ Node *node = (Node *) expr;
+
+ width = get_typavgwidth(exprType(node), exprTypmod(node));
+ }
+
+ /*
+ * Values are passed as Datum type, so comparisons can't be cheaper than
+ * comparing a Datum value.
+ *
+ * FIXME I find this reasoning questionable. We may pass int2, and
+ * comparing it is probably a bit cheaper than comparing a bigint.
+ */
+ if (width <= sizeof(Datum))
+ return 1.0;
+
+ /*
+ * We consider the cost of a comparison not to be directly proportional to
+ * width of the argument, because widths of the arguments could be
+ * slightly different (we only know the average width for the whole
+ * column). So we use log16(width) as an estimate.
+ */
+ return 1.0 + 0.125 * LOG2(width / sizeof(Datum));
+}
+
+/*
+ * compute_cpu_sort_cost
+ * compute CPU cost of sort (i.e. in-memory)
+ *
+ * The main thing we need to calculate to estimate sort CPU costs is the number
+ * of calls to the comparator functions. The difficulty is that for multi-column
+ * sorts there may be different data types involved (for some of which the calls
+ * may be much more expensive). Furthermore, columns may have a very different
+ * number of distinct values - the higher the number, the fewer comparisons will
+ * be needed for the following columns.
+ *
+ * The algorithm is incremental - we add pathkeys one by one, and at each step we
+ * estimate the number of necessary comparisons (based on the number of distinct
+ * groups in the current pathkey prefix and the new pathkey), and the comparison
+ * costs (which is data type specific).
+ *
+ * Estimation of the number of comparisons is based on ideas from:
+ *
+ * "Quicksort Is Optimal", Robert Sedgewick, Jon Bentley, 2002
+ * [https://www.cs.princeton.edu/~rs/talks/QuicksortIsOptimal.pdf]
+ *
+ * In term of that paper, let N - number of tuples, Xi - number of identical
+ * tuples with value Ki, then the estimate of number of comparisons is:
+ *
+ * log(N! / (X1! * X2! * ..)) ~ sum(Xi * log(N/Xi))
+ *
+ * We assume all Xi the same because now we don't have any estimation of
+ * group sizes, we have only know the estimate of number of groups (distinct
+ * values). In that case, formula becomes:
+ *
+ * N * log(NumberOfGroups)
+ *
+ * For multi-column sorts we need to estimate the number of comparisons for
+ * each individual column - for example with columns (c1, c2, ..., ck) we
+ * can estimate that number of comparisons on ck is roughly
+ *
+ * ncomparisons(c1, c2, ..., ck) / ncomparisons(c1, c2, ..., c(k-1))
+ *
+ * Let k be a column number, Gk - number of groups defined by k columns, and Fk
+ * the cost of the comparison is
+ *
+ * N * sum( Fk * log(Gk) )
+ *
+ * Note: We also consider column width, not just the comparator cost.
+ *
+ * NOTE: some callers currently pass NIL for pathkeys because they
+ * can't conveniently supply the sort keys. In this case, it will fallback to
+ * simple comparison cost estimate.
+ */
+static Cost
+compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
+ Cost comparison_cost, double tuples, double output_tuples,
+ bool heapSort)
+{
+ Cost per_tuple_cost = 0.0;
+ ListCell *lc;
+ List *pathkeyExprs = NIL;
+ double tuplesPerPrevGroup = tuples;
+ double totalFuncCost = 1.0;
+ bool has_fake_var = false;
+ int i = 0;
+ Oid prev_datatype = InvalidOid;
+ List *cache_varinfos = NIL;
+
+ /* fallback if pathkeys is unknown */
+ if (list_length(pathkeys) == 0)
+ {
+ /*
+ * If we'll use a bounded heap-sort keeping just K tuples in memory,
+ * for a total number of tuple comparisons of N log2 K; but the
+ * constant factor is a bit higher than for quicksort. Tweak it so
+ * that the cost curve is continuous at the crossover point.
+ */
+ output_tuples = (heapSort) ? 2.0 * output_tuples : tuples;
+ per_tuple_cost += 2.0 * cpu_operator_cost * LOG2(output_tuples);
+
+ /* add cost provided by caller */
+ per_tuple_cost += comparison_cost;
+
+ return per_tuple_cost * tuples;
+ }
+
+ /*
+ * Computing total cost of sorting takes into account the per-column
+ * comparison function cost. We try to compute the needed number of
+ * comparisons per column.
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(lc);
+ EquivalenceMember *em;
+ double nGroups,
+ correctedNGroups;
+ Cost funcCost = 1.0;
+
+ /*
+ * We believe that equivalence members aren't very different, so, to
+ * estimate cost we consider just the first member.
+ */
+ em = (EquivalenceMember *) linitial(pathkey->pk_eclass->ec_members);
+
+ if (em->em_datatype != InvalidOid)
+ {
+ /* do not lookup funcCost if the data type is the same */
+ if (prev_datatype != em->em_datatype)
+ {
+ Oid sortop;
+ QualCost cost;
+
+ sortop = get_opfamily_member(pathkey->pk_opfamily,
+ em->em_datatype, em->em_datatype,
+ pathkey->pk_strategy);
+
+ cost.startup = 0;
+ cost.per_tuple = 0;
+ add_function_cost(root, get_opcode(sortop), NULL, &cost);
+
+ /*
+ * add_function_cost returns the product of cpu_operator_cost
+ * and procost, but we need just procost, co undo that.
+ */
+ funcCost = cost.per_tuple / cpu_operator_cost;
+
+ prev_datatype = em->em_datatype;
+ }
+ }
+
+ /* factor in the width of the values in this column */
+ funcCost *= get_width_cost_multiplier(root, em->em_expr);
+
+ /* now we have per-key cost, so add to the running total */
+ totalFuncCost += funcCost;
+
+ /* remember if we have found a fake Var in pathkeys */
+ has_fake_var |= is_fake_var(em->em_expr);
+ pathkeyExprs = lappend(pathkeyExprs, em->em_expr);
+
+ /*
+ * We need to calculate the number of comparisons for this column,
+ * which requires knowing the group size. So we estimate the number of
+ * groups by calling estimate_num_groups_incremental(), which
+ * estimates the group size for "new" pathkeys.
+ *
+ * Note: estimate_num_groups_incremental does not handle fake Vars, so
+ * use a default estimate otherwise.
+ */
+ if (!has_fake_var)
+ nGroups = estimate_num_groups_incremental(root, pathkeyExprs,
+ tuplesPerPrevGroup, NULL, NULL,
+ &cache_varinfos,
+ list_length(pathkeyExprs) - 1);
+ else if (tuples > 4.0)
+
+ /*
+ * Use geometric mean as estimation if there are no stats.
+ *
+ * We don't use DEFAULT_NUM_DISTINCT here, because that's used for
+ * a single column, but here we're dealing with multiple columns.
+ */
+ nGroups = ceil(2.0 + sqrt(tuples) * (i + 1) / list_length(pathkeys));
+ else
+ nGroups = tuples;
+
+ /*
+ * Presorted keys are not considered in the cost above, but we still
+ * do have to compare them in the qsort comparator. So make sure to
+ * factor in the cost in that case.
+ */
+ if (i >= nPresortedKeys)
+ {
+ if (heapSort)
+ {
+ /*
+ * have to keep at least one group, and a multiple of group
+ * size
+ */
+ correctedNGroups = ceil(output_tuples / tuplesPerPrevGroup);
+ }
+ else
+ /* all groups in the input */
+ correctedNGroups = nGroups;
+
+ correctedNGroups = Max(1.0, ceil(correctedNGroups));
+
+ per_tuple_cost += totalFuncCost * LOG2(correctedNGroups);
+ }
+
+ i++;
+
+ /*
+ * Once we get single-row group, it means tuples in the group are
+ * unique and we can skip all remaining columns.
+ */
+ if (tuplesPerPrevGroup <= nGroups)
+ break;
+
+ /*
+ * Uniform distributions with all groups being of the same size are
+ * the best case, with nice smooth behavior. Real-world distributions
+ * tend not to be uniform, though, and we don't have any reliable
+ * easy-to-use information. As a basic defense against skewed
+ * distributions, we use a 1.5 factor to make the expected group a bit
+ * larger, but we need to be careful not to make the group larger than
+ * in the preceding step.
+ */
+ tuplesPerPrevGroup = Min(tuplesPerPrevGroup,
+ ceil(1.5 * tuplesPerPrevGroup / nGroups));
+ }
+
+ list_free(pathkeyExprs);
+
+ /* per_tuple_cost is in cpu_operator_cost units */
+ per_tuple_cost *= cpu_operator_cost;
+
+ /*
+ * Accordingly to "Introduction to algorithms", Thomas H. Cormen, Charles
+ * E. Leiserson, Ronald L. Rivest, ISBN 0-07-013143-0, quicksort
+ * estimation formula has additional term proportional to number of tuples
+ * (see Chapter 8.2 and Theorem 4.1). That affects cases with a low number
+ * of tuples, approximately less than 1e4. We could implement it as an
+ * additional multiplier under the logarithm, but we use a bit more
+ * complex formula which takes into account the number of unique tuples
+ * and it's not clear how to combine the multiplier with the number of
+ * groups. Estimate it as 10 cpu_operator_cost units.
+ */
+ per_tuple_cost += 10 * cpu_operator_cost;
+
+ per_tuple_cost += comparison_cost;
+
+ return tuples * per_tuple_cost;
+}
+
+/*
+ * simple wrapper just to estimate best sort path
+ */
+Cost
+cost_sort_estimate(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
+ double tuples)
+{
+ return compute_cpu_sort_cost(root, pathkeys, nPresortedKeys,
+ 0, tuples, tuples, false);
+}
+
/*
* cost_tuplesort
* Determines and returns the cost of sorting a relation using tuplesort,
@@ -1832,7 +2153,7 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
* number of initial runs formed and M is the merge order used by tuplesort.c.
* Since the average initial run should be about sort_mem, we have
* disk traffic = 2 * relsize * ceil(logM(p / sort_mem))
- * cpu = comparison_cost * t * log2(t)
+ * and cpu cost (computed by compute_cpu_sort_cost()).
*
* If the sort is bounded (i.e., only the first k result tuples are needed)
* and k tuples can fit into sort_mem, we use a heap method that keeps only
@@ -1851,9 +2172,11 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
* 'comparison_cost' is the extra cost per comparison, if any
* 'sort_mem' is the number of kilobytes of work memory allowed for the sort
* 'limit_tuples' is the bound on the number of output tuples; -1 if no bound
+ * 'startup_cost' is expected to be 0 at input. If there is "input cost" it should
+ * be added by caller later
*/
static void
-cost_tuplesort(Cost *startup_cost, Cost *run_cost,
+cost_tuplesort(PlannerInfo *root, List *pathkeys, Cost *startup_cost, Cost *run_cost,
double tuples, int width,
Cost comparison_cost, int sort_mem,
double limit_tuples)
@@ -1870,9 +2193,6 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
if (tuples < 2.0)
tuples = 2.0;
- /* Include the default cost-per-comparison */
- comparison_cost += 2.0 * cpu_operator_cost;
-
/* Do we have a useful LIMIT? */
if (limit_tuples > 0 && limit_tuples < tuples)
{
@@ -1896,12 +2216,10 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
double log_runs;
double npageaccesses;
- /*
- * CPU costs
- *
- * Assume about N log2 N comparisons
- */
- *startup_cost = comparison_cost * tuples * LOG2(tuples);
+ /* CPU costs */
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ tuples, false);
/* Disk costs */
@@ -1917,18 +2235,17 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
}
else if (tuples > 2 * output_tuples || input_bytes > sort_mem_bytes)
{
- /*
- * We'll use a bounded heap-sort keeping just K tuples in memory, for
- * a total number of tuple comparisons of N log2 K; but the constant
- * factor is a bit higher than for quicksort. Tweak it so that the
- * cost curve is continuous at the crossover point.
- */
- *startup_cost = comparison_cost * tuples * LOG2(2.0 * output_tuples);
+ /* We'll use a bounded heap-sort keeping just K tuples in memory. */
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ output_tuples, true);
}
else
{
/* We'll use plain quicksort on all the input tuples */
- *startup_cost = comparison_cost * tuples * LOG2(tuples);
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ tuples, false);
}
/*
@@ -2042,7 +2359,7 @@ cost_incremental_sort(Path *path,
* Estimate the average cost of sorting of one group where presorted keys
* are equal.
*/
- cost_tuplesort(&group_startup_cost, &group_run_cost,
+ cost_tuplesort(root, pathkeys, &group_startup_cost, &group_run_cost,
group_tuples, width, comparison_cost, sort_mem,
limit_tuples);
@@ -2102,7 +2419,7 @@ cost_sort(Path *path, PlannerInfo *root,
Cost startup_cost;
Cost run_cost;
- cost_tuplesort(&startup_cost, &run_cost,
+ cost_tuplesort(root, pathkeys, &startup_cost, &run_cost,
tuples, width,
comparison_cost, sort_mem,
limit_tuples);
@@ -2200,7 +2517,7 @@ append_nonpartial_cost(List *subpaths, int numpaths, int parallel_workers)
* Determines and returns the cost of an Append node.
*/
void
-cost_append(AppendPath *apath)
+cost_append(AppendPath *apath, PlannerInfo *root)
{
ListCell *l;
@@ -2268,7 +2585,7 @@ cost_append(AppendPath *apath)
* any child.
*/
cost_sort(&sort_path,
- NULL, /* doesn't currently need root */
+ root,
pathkeys,
subpath->total_cost,
subpath->rows,
diff --git a/src/backend/optimizer/path/equivclass.c b/src/backend/optimizer/path/equivclass.c
index 7fa502d6e2..07edd4f38e 100644
--- a/src/backend/optimizer/path/equivclass.c
+++ b/src/backend/optimizer/path/equivclass.c
@@ -652,7 +652,18 @@ get_eclass_for_sort_expr(PlannerInfo *root,
if (opcintype == cur_em->em_datatype &&
equal(expr, cur_em->em_expr))
- return cur_ec; /* Match! */
+ {
+ /*
+ * Match!
+ *
+ * Copy the sortref if it wasn't set yet. That may happen if
+ * the ec was constructed from WHERE clause, i.e. it doesn't
+ * have a target reference at all.
+ */
+ if (cur_ec->ec_sortref == 0 && sortref > 0)
+ cur_ec->ec_sortref = sortref;
+ return cur_ec;
+ }
}
}
diff --git a/src/backend/optimizer/path/pathkeys.c b/src/backend/optimizer/path/pathkeys.c
index fdb60aaa8d..ac3e13ddd4 100644
--- a/src/backend/optimizer/path/pathkeys.c
+++ b/src/backend/optimizer/path/pathkeys.c
@@ -17,17 +17,24 @@
*/
#include "postgres.h"
+#include <float.h>
+
+#include "miscadmin.h"
#include "access/stratnum.h"
#include "catalog/pg_opfamily.h"
#include "nodes/makefuncs.h"
#include "nodes/nodeFuncs.h"
#include "nodes/plannodes.h"
+#include "optimizer/cost.h"
#include "optimizer/optimizer.h"
#include "optimizer/pathnode.h"
#include "optimizer/paths.h"
#include "partitioning/partbounds.h"
#include "utils/lsyscache.h"
+#include "utils/selfuncs.h"
+/* Consider reordering of GROUP BY keys? */
+bool enable_group_by_reordering = true;
static bool pathkey_is_redundant(PathKey *new_pathkey, List *pathkeys);
static bool matches_boolean_partition_clause(RestrictInfo *rinfo,
@@ -350,6 +357,546 @@ pathkeys_contained_in(List *keys1, List *keys2)
return false;
}
+/*
+ * group_keys_reorder_by_pathkeys
+ * Reorder GROUP BY keys to match pathkeys of input path.
+ *
+ * Function returns new lists (pathkeys and clauses), original GROUP BY lists
+ * stay untouched.
+ *
+ * Returns the number of GROUP BY keys with a matching pathkey.
+ */
+int
+group_keys_reorder_by_pathkeys(List *pathkeys, List **group_pathkeys,
+ List **group_clauses,
+ int num_groupby_pathkeys)
+{
+ List *new_group_pathkeys = NIL,
+ *new_group_clauses = NIL;
+ ListCell *lc;
+ int n;
+
+ if (pathkeys == NIL || *group_pathkeys == NIL)
+ return 0;
+
+ /*
+ * Walk the pathkeys (determining ordering of the input path) and see if
+ * there's a matching GROUP BY key. If we find one, we append it to the
+ * list, and do the same for the clauses.
+ *
+ * Once we find the first pathkey without a matching GROUP BY key, the
+ * rest of the pathkeys are useless and can't be used to evaluate the
+ * grouping, so we abort the loop and ignore the remaining pathkeys.
+ *
+ * XXX Pathkeys are built in a way to allow simply comparing pointers.
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(lc);
+ SortGroupClause *sgc;
+
+ if (foreach_current_index(lc) >= num_groupby_pathkeys ||
+ !list_member_ptr(*group_pathkeys, pathkey))
+ /* */
+ break;
+
+ /* abort on first mismatch */
+ sgc = get_sortgroupref_clause_noerr(pathkey->pk_eclass->ec_sortref,
+ *group_clauses);
+ if (!sgc)
+ /* The grouping clause is not cover this paathkey */
+ break;
+
+ new_group_pathkeys = lappend(new_group_pathkeys, pathkey);
+ new_group_clauses = lappend(new_group_clauses, sgc);
+ }
+
+ /* remember the number of pathkeys with a matching GROUP BY key */
+ n = list_length(new_group_pathkeys);
+
+ /* append the remaining group pathkeys (will be treated as not sorted) */
+ *group_pathkeys = list_concat_unique_ptr(new_group_pathkeys,
+ *group_pathkeys);
+ *group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+
+ return n;
+}
+
+/*
+ * Used to generate all permutations of a pathkey list.
+ */
+typedef struct PathkeyMutatorState
+{
+ List *elemsList;
+ ListCell **elemCells;
+ void **elems;
+ int *positions;
+ int mutatorNColumns;
+ int count;
+} PathkeyMutatorState;
+
+
+/*
+ * PathkeyMutatorInit
+ * Initialize state of the permutation generator.
+ *
+ * We want to generate permutations of elements in the "elems" list. We may want
+ * to skip some number of elements at the beginning (when treating as presorted)
+ * or at the end (we only permute a limited number of group keys).
+ *
+ * The list is decomposed into elements, and we also keep pointers to individual
+ * cells. This allows us to build the permuted list quickly and cheaply, without
+ * creating any copies.
+ */
+static void
+PathkeyMutatorInit(PathkeyMutatorState *state, List *elems, int start, int end)
+{
+ int i;
+ int n = end - start;
+ ListCell *lc;
+
+ memset(state, 0, sizeof(*state));
+
+ state->mutatorNColumns = n;
+
+ state->elemsList = list_copy(elems);
+
+ state->elems = palloc(sizeof(void *) * n);
+ state->elemCells = palloc(sizeof(ListCell *) * n);
+ state->positions = palloc(sizeof(int) * n);
+
+ i = 0;
+ for_each_cell(lc, state->elemsList, list_nth_cell(state->elemsList, start))
+ {
+ state->elemCells[i] = lc;
+ state->elems[i] = lfirst(lc);
+ state->positions[i] = i + 1;
+ i++;
+
+ if (i >= n)
+ break;
+ }
+}
+
+/* Swap two elements of an array. */
+static void
+PathkeyMutatorSwap(int *a, int i, int j)
+{
+ int s = a[i];
+
+ a[i] = a[j];
+ a[j] = s;
+}
+
+/*
+ * Generate the next permutation of elements.
+ */
+static bool
+PathkeyMutatorNextSet(int *a, int n)
+{
+ int j,
+ k,
+ l,
+ r;
+
+ j = n - 2;
+
+ while (j >= 0 && a[j] >= a[j + 1])
+ j--;
+
+ if (j < 0)
+ return false;
+
+ k = n - 1;
+
+ while (k > 0 && a[j] >= a[k])
+ k--;
+
+ PathkeyMutatorSwap(a, j, k);
+
+ l = j + 1;
+ r = n - 1;
+
+ while (l < r)
+ PathkeyMutatorSwap(a, l++, r--);
+
+ return true;
+}
+
+/*
+ * PathkeyMutatorNext
+ * Generate the next permutation of list of elements.
+ *
+ * Returns the next permutation (as a list of elements) or NIL if there are no
+ * more permutations.
+ */
+static List *
+PathkeyMutatorNext(PathkeyMutatorState *state)
+{
+ int i;
+
+ state->count++;
+
+ /* first permutation is original list */
+ if (state->count == 1)
+ return state->elemsList;
+
+ /* when there are no more permutations, return NIL */
+ if (!PathkeyMutatorNextSet(state->positions, state->mutatorNColumns))
+ {
+ pfree(state->elems);
+ pfree(state->elemCells);
+ pfree(state->positions);
+
+ list_free(state->elemsList);
+
+ return NIL;
+ }
+
+ /* update the list cells to point to the right elements */
+ for (i = 0; i < state->mutatorNColumns; i++)
+ lfirst(state->elemCells[i]) =
+ (void *) state->elems[state->positions[i] - 1];
+
+ return state->elemsList;
+}
+
+/*
+ * Cost of comparing pathkeys.
+ */
+typedef struct PathkeySortCost
+{
+ Cost cost;
+ PathKey *pathkey;
+} PathkeySortCost;
+
+static int
+pathkey_sort_cost_comparator(const void *_a, const void *_b)
+{
+ const PathkeySortCost *a = (PathkeySortCost *) _a;
+ const PathkeySortCost *b = (PathkeySortCost *) _b;
+
+ if (a->cost < b->cost)
+ return -1;
+ else if (a->cost == b->cost)
+ return 0;
+ return 1;
+}
+
+/*
+ * get_cheapest_group_keys_order
+ * Reorders the group pathkeys / clauses to minimize the comparison cost.
+ *
+ * Given the list of pathkeys in '*group_pathkeys', we try to arrange these
+ * in an order that minimizes the sort costs that will be incurred by the
+ * GROUP BY. The costs mainly depend on the cost of the sort comparator
+ * function(s) and the number of distinct values in each column of the GROUP
+ * BY clause (*group_clauses). Sorting on subsequent columns is only required
+ * for tiebreak situations where two values sort equally.
+ *
+ * In case the input is partially sorted, only the remaining pathkeys are
+ * considered. 'n_preordered' denotes how many of the leading *group_pathkeys
+ * the input is presorted by.
+ *
+ * Returns true and sets *group_pathkeys and *group_clauses to the newly
+ * ordered versions of the lists that were passed in via these parameters.
+ * If no reordering was deemed necessary then we return false, in which case
+ * the *group_pathkeys and *group_clauses lists are left untouched. The
+ * original *group_pathkeys and *group_clauses parameter values are never
+ * destructively modified in place.
+ */
+static bool
+get_cheapest_group_keys_order(PlannerInfo *root, double nrows,
+ List **group_pathkeys, List **group_clauses,
+ int n_preordered)
+{
+ List *new_group_pathkeys = NIL;
+ List *new_group_clauses = NIL;
+ List *var_group_pathkeys;
+ ListCell *cell;
+ PathkeyMutatorState mstate;
+ double cheapest_sort_cost = DBL_MAX;
+ int nFreeKeys;
+ int nToPermute;
+
+ /* If there are less than 2 unsorted pathkeys, we're done. */
+ if (root->num_groupby_pathkeys - n_preordered < 2)
+ return false;
+
+ /*
+ * We could exhaustively cost all possible orderings of the pathkeys, but
+ * for a large number of pathkeys it might be prohibitively expensive. So
+ * we try to apply simple cheap heuristics first - we sort the pathkeys by
+ * sort cost (as if the pathkey was sorted independently) and then check
+ * only the four cheapest pathkeys. The remaining pathkeys are kept
+ * ordered by cost.
+ *
+ * XXX This is a very simple heuristics, but likely to work fine for most
+ * cases (because the number of GROUP BY clauses tends to be lower than
+ * 4). But it ignores how the number of distinct values in each pathkey
+ * affects the following steps. It might be better to use "more expensive"
+ * pathkey first if it has many distinct values, because it then limits
+ * the number of comparisons for the remaining pathkeys. But evaluating
+ * that is likely quite the expensive.
+ */
+ nFreeKeys = root->num_groupby_pathkeys - n_preordered;
+ nToPermute = 4;
+ if (nFreeKeys > nToPermute)
+ {
+ PathkeySortCost *costs = palloc(sizeof(PathkeySortCost) * nFreeKeys);
+ PathkeySortCost *cost = costs;
+ int i;
+
+ /*
+ * Estimate cost for sorting individual pathkeys skipping the
+ * pre-ordered pathkeys.
+ */
+ for_each_from(cell, *group_pathkeys, n_preordered)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(cell);
+ List *to_cost;
+
+ if (foreach_current_index(cell) >= root->num_groupby_pathkeys)
+ break;
+
+ to_cost = list_make1(pathkey);
+ cost->pathkey = pathkey;
+ cost->cost = cost_sort_estimate(root, to_cost, 0, nrows);
+ cost++;
+
+ list_free(to_cost);
+ }
+
+ /* sort the pathkeys by sort cost in ascending order */
+ qsort(costs, nFreeKeys, sizeof(*costs), pathkey_sort_cost_comparator);
+
+ /*
+ * Rebuild the list of pathkeys - first the preordered ones, then the
+ * rest ordered by cost.
+ */
+ new_group_pathkeys = list_copy_head(*group_pathkeys, n_preordered);
+
+ for (i = 0; i < nFreeKeys; i++)
+ new_group_pathkeys = lappend(new_group_pathkeys, costs[i].pathkey);
+
+ new_group_pathkeys = list_copy_tail(*group_pathkeys,
+ root->num_groupby_pathkeys);
+ pfree(costs);
+ }
+ else
+ {
+ /* Copy the list, so that we can free the new list by list_free. */
+ new_group_pathkeys = list_copy(*group_pathkeys);
+ nToPermute = nFreeKeys;
+ }
+
+ Assert(list_length(new_group_pathkeys) == list_length(*group_pathkeys));
+
+ /*
+ * Generate pathkey lists with permutations of the first nToPermute
+ * pathkeys.
+ *
+ * XXX We simply calculate sort cost for each individual pathkey list, but
+ * there's room for two dynamic programming optimizations here. Firstly,
+ * we may pass the current "best" cost to cost_sort_estimate so that it
+ * can "abort" if the estimated pathkeys list exceeds it. Secondly, it
+ * could pass the return information about the position when it exceeded
+ * the cost, and we could skip all permutations with the same prefix.
+ *
+ * Imagine we've already found ordering with cost C1, and we're evaluating
+ * another ordering - cost_sort_estimate() calculates cost by adding the
+ * pathkeys one by one (more or less), and the cost only grows. If at any
+ * point it exceeds C1, it can't possibly be "better" so we can discard
+ * it. But we also know that we can discard all ordering with the same
+ * prefix, because if we're estimating (a,b,c,d) and we exceed C1 at (a,b)
+ * then the same thing will happen for any ordering with this prefix.
+ */
+ PathkeyMutatorInit(&mstate, new_group_pathkeys, n_preordered,
+ n_preordered + nToPermute);
+
+ while ((var_group_pathkeys = PathkeyMutatorNext(&mstate)) != NIL)
+ {
+ Cost cost;
+
+ cost = cost_sort_estimate(root, var_group_pathkeys, n_preordered, nrows);
+
+ if (cost < cheapest_sort_cost)
+ {
+ list_free(new_group_pathkeys);
+ new_group_pathkeys = list_copy(var_group_pathkeys);
+ cheapest_sort_cost = cost;
+ }
+ }
+
+ Assert(list_length(new_group_pathkeys) == list_length(root->group_pathkeys));
+
+ /* Reorder the group clauses according to the reordered pathkeys. */
+ foreach(cell, new_group_pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(cell);
+
+ if (foreach_current_index(cell) >= root->num_groupby_pathkeys)
+ break;
+
+ new_group_clauses =
+ lappend(new_group_clauses,
+ get_sortgroupref_clause(pathkey->pk_eclass->ec_sortref,
+ *group_clauses));
+ }
+
+ /* Just append the rest GROUP BY clauses */
+ new_group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+ *group_pathkeys = new_group_pathkeys;
+ *group_clauses = new_group_clauses;
+ Assert(list_length(new_group_clauses) == list_length(root->processed_groupClause));
+
+ return true;
+}
+
+/*
+ * get_useful_group_keys_orderings
+ * Determine which orderings of GROUP BY keys are potentially interesting.
+ *
+ * Returns list of PathKeyInfo items, each representing an interesting ordering
+ * of GROUP BY keys. Each item stores pathkeys and clauses in matching order.
+ *
+ * The function considers (and keeps) multiple group by orderings:
+ *
+ * - the original ordering, as specified by the GROUP BY clause
+ *
+ * - GROUP BY keys reordered to minimize the sort cost
+ *
+ * - GROUP BY keys reordered to match path ordering (as much as possible), with
+ * the tail reordered to minimize the sort cost
+ *
+ * - GROUP BY keys to match target ORDER BY clause (as much as possible), with
+ * the tail reordered to minimize the sort cost
+ *
+ * There are other potentially interesting orderings (e.g. it might be best to
+ * match the first ORDER BY key, order the remaining keys differently and then
+ * rely on the incremental sort to fix this), but we ignore those for now. To
+ * make this work we'd have to pretty much generate all possible permutations.
+ */
+List *
+get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys)
+{
+ Query *parse = root->parse;
+ List *infos = NIL;
+ PathKeyInfo *info;
+ int n_preordered = 0;
+
+ List *pathkeys = root->group_pathkeys;
+ List *clauses = root->processed_groupClause;
+
+ /* always return at least the original pathkeys/clauses */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+ infos = lappend(infos, info);
+
+ /*
+ * Should we try generating alternative orderings of the group keys? If
+ * not, we produce only the order specified in the query, i.e. the
+ * optimization is effectively disabled.
+ */
+ if (!enable_group_by_reordering)
+ return infos;
+
+ /* for grouping sets we can't do any reordering */
+ if (parse->groupingSets)
+ return infos;
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost, ignoring both the
+ * target ordering (ORDER BY) and ordering of the input path.
+ */
+ if (get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered))
+ {
+ Assert(list_length(pathkeys) == list_length(clauses));
+
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * If the path is sorted in some way, try reordering the group keys to
+ * match as much of the ordering as possible - we get this sort for free
+ * (mostly).
+ *
+ * We must not do this when there are no grouping sets, because those use
+ * more complex logic to decide the ordering.
+ *
+ * XXX Isn't this somewhat redundant with presorted_keys? Actually, it's
+ * more a complement, because it allows benefiting from incremental sort
+ * as much as possible.
+ *
+ * XXX This does nothing if (n_preordered == 0). We shouldn't create the
+ * info in this case.
+ */
+ if (path_pathkeys)
+ {
+ n_preordered = group_keys_reorder_by_pathkeys(path_pathkeys,
+ &pathkeys,
+ &clauses,
+ root->num_groupby_pathkeys);
+
+ /* reorder the tail to minimize sort cost */
+ get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, in
+ * which case get_cheapest_group_keys_order returns false. But we
+ * still want to keep the keys reordered to path_pathkeys.
+ */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost (this time consider
+ * the ORDER BY clause, but only if set debug_group_by_match_order_by).
+ */
+ if (root->sort_pathkeys)
+ {
+ n_preordered = group_keys_reorder_by_pathkeys(root->sort_pathkeys,
+ &pathkeys,
+ &clauses,
+ root->num_groupby_pathkeys);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, in
+ * which case get_cheapest_group_keys_order returns false. But we
+ * still want to keep the keys reordered to sort_pathkeys.
+ */
+ get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered);
+
+ /* keep the group keys reordered to match ordering of input path */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ return infos;
+}
+
/*
* pathkeys_count_contained_in
* Same as pathkeys_contained_in, but also sets length of longest
@@ -1939,6 +2486,54 @@ pathkeys_useful_for_ordering(PlannerInfo *root, List *pathkeys)
return n_common_pathkeys;
}
+/*
+ * pathkeys_useful_for_grouping
+ * Count the number of pathkeys that are useful for grouping (instead of
+ * explicit sort)
+ *
+ * Group pathkeys could be reordered to benefit from the ordering. The
+ * ordering may not be "complete" and may require incremental sort, but that's
+ * fine. So we simply count prefix pathkeys with a matching group key, and
+ * stop once we find the first pathkey without a match.
+ *
+ * So e.g. with pathkeys (a,b,c) and group keys (a,b,e) this determines (a,b)
+ * pathkeys are useful for grouping, and we might do incremental sort to get
+ * path ordered by (a,b,e).
+ *
+ * This logic is necessary to retain paths with ordering not matching grouping
+ * keys directly, without the reordering.
+ *
+ * Returns the length of pathkey prefix with matching group keys.
+ */
+static int
+pathkeys_useful_for_grouping(PlannerInfo *root, List *pathkeys)
+{
+ ListCell *key;
+ int n = 0;
+
+ /* no special ordering requested for grouping */
+ if (root->group_pathkeys == NIL)
+ return 0;
+
+ /* unordered path */
+ if (pathkeys == NIL)
+ return 0;
+
+ /* walk the pathkeys and search for matching group key */
+ foreach(key, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(key);
+
+ /* no matching group key, we're done */
+ if (!list_member_ptr(root->group_pathkeys, pathkey))
+ break;
+
+ n++;
+ }
+
+ return n;
+}
+
/*
* truncate_useless_pathkeys
* Shorten the given pathkey list to just the useful pathkeys.
@@ -1953,6 +2548,9 @@ truncate_useless_pathkeys(PlannerInfo *root,
nuseful = pathkeys_useful_for_merging(root, rel, pathkeys);
nuseful2 = pathkeys_useful_for_ordering(root, pathkeys);
+ if (nuseful2 > nuseful)
+ nuseful = nuseful2;
+ nuseful2 = pathkeys_useful_for_grouping(root, pathkeys);
if (nuseful2 > nuseful)
nuseful = nuseful2;
@@ -1988,6 +2586,8 @@ has_useful_pathkeys(PlannerInfo *root, RelOptInfo *rel)
{
if (rel->joininfo != NIL || rel->has_eclass_joins)
return true; /* might be able to use pathkeys for merging */
+ if (root->group_pathkeys != NIL)
+ return true; /* might be able to use pathkeys for grouping */
if (root->query_pathkeys != NIL)
return true; /* might be able to use them for ordering */
return false; /* definitely useless */
diff --git a/src/backend/optimizer/plan/planner.c b/src/backend/optimizer/plan/planner.c
index aafef24cf1..4727a5a157 100644
--- a/src/backend/optimizer/plan/planner.c
+++ b/src/backend/optimizer/plan/planner.c
@@ -2811,8 +2811,9 @@ remove_useless_groupby_columns(PlannerInfo *root)
*
* In principle it might be interesting to consider other orderings of the
* GROUP BY elements, which could match the sort ordering of other
- * possible plans (eg an indexscan) and thereby reduce cost. We don't
- * bother with that, though. Hashed grouping will frequently win anyway.
+ * possible plans (eg an indexscan) and thereby reduce cost. However, we
+ * don't yet have sufficient information to do that here, so that's left until
+ * later in planning. See get_useful_group_keys_orderings().
*
* Note: we need no comparable processing of the distinctClause because
* the parser already enforced that that matches ORDER BY.
@@ -3469,6 +3470,7 @@ standard_qp_callback(PlannerInfo *root, void *extra)
tlist,
true,
&sortable);
+
if (!sortable)
{
/* Can't sort; no point in considering aggregate ordering either */
@@ -6843,90 +6845,108 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
+
+ Assert(list_length(pathkey_orderings) > 0);
- if (!is_sorted)
+ foreach(lc2, pathkey_orderings)
{
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially sorted
- * already (no need to deal with paths which have presorted
- * keys when incremental sort is disabled unless it's the
- * cheapest input path).
- */
- if (path != cheapest_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ bool is_sorted;
+ int presorted_keys;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- /*
- * We've no need to consider both a sort and incremental sort.
- * We'll just do a sort if there are no presorted keys and an
- * incremental sort when there are presorted keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
- /* Now decide what to stick atop it */
- if (parse->groupingSets)
- {
- consider_groupingsets_paths(root, grouped_rel,
- path, true, can_hash,
- gd, agg_costs, dNumGroups);
- }
- else if (parse->hasAggs)
- {
- /*
- * We have aggregation, possibly with plain GROUP BY. Make an
- * AggPath.
- */
- add_path(grouped_rel, (Path *)
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (!is_sorted)
+ {
+ /*
+ * Try at least sorting the cheapest path and also try
+ * incrementally sorting any path which is partially sorted
+ * already (no need to deal with paths which have presorted
+ * keys when incremental sort is disabled unless it's the
+ * cheapest input path).
+ */
+ if (path != cheapest_path &&
+ (presorted_keys == 0 || !enable_incremental_sort))
+ continue;
+
+ /*
+ * We've no need to consider both a sort and incremental sort.
+ * We'll just do a sort if there are no presorted keys and an
+ * incremental sort when there are presorted keys.
+ */
+ if (presorted_keys == 0 || !enable_incremental_sort)
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ else
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+ }
+
+ /* Now decide what to stick atop it */
+ if (parse->groupingSets)
+ {
+ consider_groupingsets_paths(root, grouped_rel,
+ path, true, can_hash,
+ gd, agg_costs, dNumGroups);
+ }
+ else if (parse->hasAggs)
+ {
+ /*
+ * We have aggregation, possibly with plain GROUP BY. Make an
+ * AggPath.
+ */
+ add_path(grouped_rel, (Path *)
create_agg_path(root,
grouped_rel,
path,
grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_SIMPLE,
- root->processed_groupClause,
+ info->clauses,
havingQual,
agg_costs,
dNumGroups));
- }
- else if (parse->groupClause)
- {
- /*
- * We have GROUP BY without aggregation or grouping sets. Make
- * a GroupPath.
- */
- add_path(grouped_rel, (Path *)
+ }
+ else if (parse->groupClause)
+ {
+ /*
+ * We have GROUP BY without aggregation or grouping sets. Make
+ * a GroupPath.
+ */
+ add_path(grouped_rel, (Path *)
create_group_path(root,
grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
havingQual,
dNumGroups));
- }
- else
- {
- /* Other cases should have been handled above */
- Assert(false);
+ }
+ else
+ {
+ /* Other cases should have been handled above */
+ Assert(false);
+ }
}
}
@@ -6938,69 +6958,84 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
{
foreach(lc, partially_grouped_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
+ List *pathkey_orderings = NIL;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
- if (!is_sorted)
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach(lc2, pathkey_orderings)
{
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially
- * sorted already (no need to deal with paths which have
- * presorted keys when incremental sort is disabled unless
- * it's the cheapest input path).
- */
- if (path != partially_grouped_rel->cheapest_total_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ bool is_sorted;
+ int presorted_keys;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- /*
- * We've no need to consider both a sort and incremental
- * sort. We'll just do a sort if there are no pre-sorted
- * keys and an incremental sort when there are presorted
- * keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
+ path->pathkeys,
+ &presorted_keys);
- if (parse->hasAggs)
- add_path(grouped_rel, (Path *)
+ if (!is_sorted)
+ {
+ /*
+ * Try at least sorting the cheapest path and also try
+ * incrementally sorting any path which is partially
+ * sorted already (no need to deal with paths which have
+ * presorted keys when incremental sort is disabled unless
+ * it's the cheapest input path).
+ */
+ if (path != partially_grouped_rel->cheapest_total_path &&
+ (presorted_keys == 0 || !enable_incremental_sort))
+ continue;
+
+ /*
+ * We've no need to consider both a sort and incremental
+ * sort. We'll just do a sort if there are no pre-sorted
+ * keys and an incremental sort when there are presorted
+ * keys.
+ */
+ if (presorted_keys == 0 || !enable_incremental_sort)
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ else
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_path(grouped_rel, (Path *)
create_agg_path(root,
grouped_rel,
path,
grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_FINAL_DESERIAL,
- root->processed_groupClause,
+ info->clauses,
havingQual,
agg_final_costs,
dNumGroups));
- else
- add_path(grouped_rel, (Path *)
+ else
+ add_path(grouped_rel, (Path *)
create_group_path(root,
grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
havingQual,
dNumGroups));
+ }
}
}
}
@@ -7203,66 +7238,85 @@ create_partial_grouping_paths(PlannerInfo *root,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
- if (!is_sorted)
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach(lc2, pathkey_orderings)
{
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially sorted
- * already (no need to deal with paths which have presorted
- * keys when incremental sort is disabled unless it's the
- * cheapest input path).
- */
- if (path != cheapest_total_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ bool is_sorted;
+ int presorted_keys;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- /*
- * We've no need to consider both a sort and incremental sort.
- * We'll just do a sort if there are no presorted keys and an
- * incremental sort when there are presorted keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
- if (parse->hasAggs)
- add_path(partially_grouped_rel, (Path *)
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+ if (!is_sorted)
+ {
+ /*
+ * Try at least sorting the cheapest path and also try
+ * incrementally sorting any path which is partially sorted
+ * already (no need to deal with paths which have presorted
+ * keys when incremental sort is disabled unless it's the
+ * cheapest input path).
+ */
+ if (path != cheapest_total_path &&
+ (presorted_keys == 0 || !enable_incremental_sort))
+ continue;
+
+ /*
+ * We've no need to consider both a sort and incremental sort.
+ * We'll just do a sort if there are no presorted keys and an
+ * incremental sort when there are presorted keys.
+ */
+ if (presorted_keys == 0 || !enable_incremental_sort)
+ path = (Path *) create_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ else
+ path = (Path *) create_incremental_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_path(partially_grouped_rel, (Path *)
create_agg_path(root,
partially_grouped_rel,
path,
partially_grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_INITIAL_SERIAL,
- root->processed_groupClause,
+ info->clauses,
NIL,
agg_partial_costs,
dNumPartialGroups));
- else
- add_path(partially_grouped_rel, (Path *)
- create_group_path(root,
- partially_grouped_rel,
- path,
- root->processed_groupClause,
- NIL,
- dNumPartialGroups));
+ else
+ add_path(partially_grouped_rel, (Path *)
+ create_group_path(root,
+ partially_grouped_rel,
+ path,
+ info->clauses,
+ NIL,
+ dNumPartialGroups));
+ }
}
}
@@ -7271,67 +7325,86 @@ create_partial_grouping_paths(PlannerInfo *root,
/* Similar to above logic, but for partial paths. */
foreach(lc, input_rel->partial_pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
+
+ Assert(list_length(pathkey_orderings) > 0);
- if (!is_sorted)
+ /* process all potentially interesting grouping reorderings */
+ foreach(lc2, pathkey_orderings)
{
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially sorted
- * already (no need to deal with paths which have presorted
- * keys when incremental sort is disabled unless it's the
- * cheapest input path).
- */
- if (path != cheapest_partial_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ bool is_sorted;
+ int presorted_keys;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- /*
- * We've no need to consider both a sort and incremental sort.
- * We'll just do a sort if there are no presorted keys and an
- * incremental sort when there are presorted keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
- if (parse->hasAggs)
- add_partial_path(partially_grouped_rel, (Path *)
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (!is_sorted)
+ {
+ /*
+ * Try at least sorting the cheapest path and also try
+ * incrementally sorting any path which is partially sorted
+ * already (no need to deal with paths which have presorted
+ * keys when incremental sort is disabled unless it's the
+ * cheapest input path).
+ */
+ if (path != cheapest_partial_path &&
+ (presorted_keys == 0 || !enable_incremental_sort))
+ continue;
+
+ /*
+ * We've no need to consider both a sort and incremental sort.
+ * We'll just do a sort if there are no presorted keys and an
+ * incremental sort when there are presorted keys.
+ */
+ if (presorted_keys == 0 || !enable_incremental_sort)
+ path = (Path *) create_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ else
+ path = (Path *) create_incremental_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_partial_path(partially_grouped_rel, (Path *)
create_agg_path(root,
partially_grouped_rel,
path,
partially_grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_INITIAL_SERIAL,
- root->processed_groupClause,
+ info->clauses,
NIL,
agg_partial_costs,
dNumPartialPartialGroups));
- else
- add_partial_path(partially_grouped_rel, (Path *)
+ else
+ add_partial_path(partially_grouped_rel, (Path *)
create_group_path(root,
partially_grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
NIL,
dNumPartialPartialGroups));
+ }
}
}
@@ -7445,6 +7518,8 @@ gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
* We can also skip the entire loop when we only have a single-item
* group_pathkeys because then we can't possibly have a presorted prefix
* of the list without having the list be fully sorted.
+ *
+ * XXX Shouldn't this also consider the group-key-reordering?
*/
if (!enable_incremental_sort || list_length(root->group_pathkeys) == 1)
return;
diff --git a/src/backend/optimizer/util/pathnode.c b/src/backend/optimizer/util/pathnode.c
index 211ba65389..7b8d0cd294 100644
--- a/src/backend/optimizer/util/pathnode.c
+++ b/src/backend/optimizer/util/pathnode.c
@@ -1346,12 +1346,12 @@ create_append_path(PlannerInfo *root,
pathnode->path.total_cost = child->total_cost;
}
else
- cost_append(pathnode);
+ cost_append(pathnode, root);
/* Must do this last, else cost_append complains */
pathnode->path.pathkeys = child->pathkeys;
}
else
- cost_append(pathnode);
+ cost_append(pathnode, root);
/* If the caller provided a row estimate, override the computed value. */
if (rows >= 0)
diff --git a/src/backend/utils/adt/selfuncs.c b/src/backend/utils/adt/selfuncs.c
index c4fcd0076e..cb7e67099e 100644
--- a/src/backend/utils/adt/selfuncs.c
+++ b/src/backend/utils/adt/selfuncs.c
@@ -3386,11 +3386,28 @@ double
estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
List **pgset, EstimationInfo *estinfo)
{
- List *varinfos = NIL;
+ return estimate_num_groups_incremental(root, groupExprs,
+ input_rows, pgset, estinfo,
+ NULL, 0);
+}
+
+/*
+ * estimate_num_groups_incremental
+ * An estimate_num_groups variant, optimized for cases that are adding the
+ * expressions incrementally (e.g. one by one).
+ */
+double
+estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
+ double input_rows,
+ List **pgset, EstimationInfo *estinfo,
+ List **cache_varinfos, int prevNExprs)
+{
+ List *varinfos = (cache_varinfos) ? *cache_varinfos : NIL;
double srf_multiplier = 1.0;
double numdistinct;
ListCell *l;
- int i;
+ int i,
+ j;
/* Zero the estinfo output parameter, if non-NULL */
if (estinfo != NULL)
@@ -3421,7 +3438,7 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
*/
numdistinct = 1.0;
- i = 0;
+ i = j = 0;
foreach(l, groupExprs)
{
Node *groupexpr = (Node *) lfirst(l);
@@ -3430,6 +3447,14 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
List *varshere;
ListCell *l2;
+ /* was done on previous call */
+ if (cache_varinfos && j++ < prevNExprs)
+ {
+ if (pgset)
+ i++; /* to keep in sync with lines below */
+ continue;
+ }
+
/* is expression in this grouping set? */
if (pgset && !list_member_int(*pgset, i++))
continue;
@@ -3499,7 +3524,11 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
if (varshere == NIL)
{
if (contain_volatile_functions(groupexpr))
+ {
+ if (cache_varinfos)
+ *cache_varinfos = varinfos;
return input_rows;
+ }
continue;
}
@@ -3516,6 +3545,9 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
}
}
+ if (cache_varinfos)
+ *cache_varinfos = varinfos;
+
/*
* If now no Vars, we must have an all-constant or all-boolean GROUP BY
* list.
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 16ec6c5ef0..9701d6c70f 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -1027,6 +1027,16 @@ struct config_bool ConfigureNamesBool[] =
true,
NULL, NULL, NULL
},
+ {
+ {"enable_group_by_reordering", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("Enables reordering of GROUP BY keys."),
+ NULL,
+ GUC_EXPLAIN
+ },
+ &enable_group_by_reordering,
+ true,
+ NULL, NULL, NULL
+ },
{
{"geqo", PGC_USERSET, QUERY_TUNING_GEQO,
gettext_noop("Enables genetic query optimization."),
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index d08d55c3fe..0fcd023c38 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -391,6 +391,7 @@
#enable_seqscan = on
#enable_sort = on
#enable_tidscan = on
+#enable_group_by_reordering = on
# - Planner Cost Constants -
diff --git a/src/include/nodes/pathnodes.h b/src/include/nodes/pathnodes.h
index 5702fbba60..15ba2f8218 100644
--- a/src/include/nodes/pathnodes.h
+++ b/src/include/nodes/pathnodes.h
@@ -1456,6 +1456,16 @@ typedef struct PathKey
bool pk_nulls_first; /* do NULLs come before normal values? */
} PathKey;
+/*
+ * Combines information about pathkeys and the associated clauses.
+ */
+typedef struct PathKeyInfo
+{
+ NodeTag type;
+ List *pathkeys;
+ List *clauses;
+} PathKeyInfo;
+
/*
* VolatileFunctionStatus -- allows nodes to cache their
* contain_volatile_functions properties. VOLATILITY_UNKNOWN means not yet
diff --git a/src/include/optimizer/cost.h b/src/include/optimizer/cost.h
index bee090ffc2..bbb1132473 100644
--- a/src/include/optimizer/cost.h
+++ b/src/include/optimizer/cost.h
@@ -116,7 +116,9 @@ extern void cost_incremental_sort(Path *path,
Cost input_startup_cost, Cost input_total_cost,
double input_tuples, int width, Cost comparison_cost, int sort_mem,
double limit_tuples);
-extern void cost_append(AppendPath *apath);
+extern Cost cost_sort_estimate(PlannerInfo *root, List *pathkeys,
+ int nPresortedKeys, double tuples);
+extern void cost_append(AppendPath *path, PlannerInfo *root);
extern void cost_merge_append(Path *path, PlannerInfo *root,
List *pathkeys, int n_streams,
Cost input_startup_cost, Cost input_total_cost,
diff --git a/src/include/optimizer/paths.h b/src/include/optimizer/paths.h
index 50bc3b503a..e4ff16d0c4 100644
--- a/src/include/optimizer/paths.h
+++ b/src/include/optimizer/paths.h
@@ -24,6 +24,7 @@ extern PGDLLIMPORT bool enable_geqo;
extern PGDLLIMPORT int geqo_threshold;
extern PGDLLIMPORT int min_parallel_table_scan_size;
extern PGDLLIMPORT int min_parallel_index_scan_size;
+extern PGDLLIMPORT bool enable_group_by_reordering;
/* Hook for plugins to get control in set_rel_pathlist() */
typedef void (*set_rel_pathlist_hook_type) (PlannerInfo *root,
@@ -205,6 +206,12 @@ typedef enum
extern PathKeysComparison compare_pathkeys(List *keys1, List *keys2);
extern bool pathkeys_contained_in(List *keys1, List *keys2);
extern bool pathkeys_count_contained_in(List *keys1, List *keys2, int *n_common);
+extern int group_keys_reorder_by_pathkeys(List *pathkeys,
+ List **group_pathkeys,
+ List **group_clauses,
+ int num_groupby_pathkeys);
+extern List *get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys);
extern Path *get_cheapest_path_for_pathkeys(List *paths, List *pathkeys,
Relids required_outer,
CostSelector cost_criterion,
diff --git a/src/include/utils/selfuncs.h b/src/include/utils/selfuncs.h
index 2f76c473db..2b21c06cc9 100644
--- a/src/include/utils/selfuncs.h
+++ b/src/include/utils/selfuncs.h
@@ -214,6 +214,11 @@ extern double estimate_num_groups(PlannerInfo *root, List *groupExprs,
double input_rows, List **pgset,
EstimationInfo *estinfo);
+extern double estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
+ double input_rows, List **pgset,
+ EstimationInfo *estinfo,
+ List **cache_varinfos, int prevNExprs);
+
extern void estimate_hash_bucket_stats(PlannerInfo *root,
Node *hashkey, double nbuckets,
Selectivity *mcv_freq,
diff --git a/src/test/regress/expected/aggregates.out b/src/test/regress/expected/aggregates.out
index d8271da4d1..205be28b3d 100644
--- a/src/test/regress/expected/aggregates.out
+++ b/src/test/regress/expected/aggregates.out
@@ -1234,7 +1234,8 @@ explain (costs off)
select distinct min(f1), max(f1) from minmaxtest;
QUERY PLAN
---------------------------------------------------------------------------------------------
- Unique
+ HashAggregate
+ Group Key: $0, $1
InitPlan 1 (returns $0)
-> Limit
-> Merge Append
@@ -1257,10 +1258,8 @@ explain (costs off)
-> Index Only Scan using minmaxtest2i on minmaxtest2 minmaxtest_8
Index Cond: (f1 IS NOT NULL)
-> Index Only Scan Backward using minmaxtest3i on minmaxtest3 minmaxtest_9
- -> Sort
- Sort Key: ($0), ($1)
- -> Result
-(26 rows)
+ -> Result
+(25 rows)
select distinct min(f1), max(f1) from minmaxtest;
min | max
@@ -2715,6 +2714,241 @@ SELECT balk(hundred) FROM tenk1;
(1 row)
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+SELECT
+ i AS id,
+ i/2 AS p,
+ format('%60s', i%2) AS v,
+ i/4 AS c,
+ i/8 AS d,
+ (random() * (10000/8))::int as e --the same as d but no correlation with p
+ INTO btg
+FROM
+ generate_series(1, 10000) i;
+VACUUM btg;
+ANALYZE btg;
+-- GROUP BY optimization by reorder columns by frequency
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Sort
+ Sort Key: p, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Sort
+ Sort Key: p, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, c, v
+ -> Sort
+ Sort Key: p, c, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: v, p, c
+ -> Sort
+ Sort Key: v, p, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: p, d, v, c
+ -> Sort
+ Sort Key: p, d, v, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: v, p, d, c
+ -> Sort
+ Sort Key: v, p, d, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: p, v, d, c
+ -> Sort
+ Sort Key: p, v, d, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, d, e
+ -> Sort
+ Sort Key: p, d, e
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, e, d
+ -> Sort
+ Sort Key: p, e, d
+ -> Seq Scan on btg
+(5 rows)
+
+CREATE STATISTICS btg_dep ON d, e, p FROM btg;
+ANALYZE btg;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, d, e
+ -> Sort
+ Sort Key: p, d, e
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, e, d
+ -> Sort
+ Sort Key: p, e, d
+ -> Seq Scan on btg
+(5 rows)
+
+-- GROUP BY optimization by reorder columns by index scan
+CREATE INDEX ON btg(p, v);
+SET enable_seqscan=off;
+SET enable_bitmapscan=off;
+VACUUM btg;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v ORDER BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p ORDER BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, c, v
+ -> Incremental Sort
+ Sort Key: p, c, v
+ Presorted Key: p
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY p, v;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, v, c
+ -> Incremental Sort
+ Sort Key: p, v, c
+ Presorted Key: p, v
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, c, v, d
+ -> Incremental Sort
+ Sort Key: p, c, v, d
+ Presorted Key: p
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d ORDER BY p, v;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, v, c, d
+ -> Incremental Sort
+ Sort Key: p, v, c, d
+ Presorted Key: p, v
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+DROP TABLE btg;
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+RESET enable_seqscan;
+RESET enable_bitmapscan;
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/expected/incremental_sort.out b/src/test/regress/expected/incremental_sort.out
index 7fdb685313..b8a0968b06 100644
--- a/src/test/regress/expected/incremental_sort.out
+++ b/src/test/regress/expected/incremental_sort.out
@@ -1426,7 +1426,7 @@ set parallel_setup_cost = 0;
set parallel_tuple_cost = 0;
set max_parallel_workers_per_gather = 2;
create table t (a int, b int, c int);
-insert into t select mod(i,10),mod(i,10),i from generate_series(1,10000) s(i);
+insert into t select mod(i,10),mod(i,10),i from generate_series(1,60000) s(i);
create index on t (a);
analyze t;
set enable_incremental_sort = off;
diff --git a/src/test/regress/expected/join.out b/src/test/regress/expected/join.out
index 9b8638f286..27684bfce2 100644
--- a/src/test/regress/expected/join.out
+++ b/src/test/regress/expected/join.out
@@ -1984,8 +1984,8 @@ USING (name);
------+----+----
bb | 12 | 13
cc | 22 | 23
- dd | | 33
ee | 42 |
+ dd | | 33
(4 rows)
-- Cases with non-nullable expressions in subquery results;
@@ -2019,8 +2019,8 @@ NATURAL FULL JOIN
------+------+------+------+------
bb | 12 | 2 | 13 | 3
cc | 22 | 2 | 23 | 3
- dd | | | 33 | 3
ee | 42 | 2 | |
+ dd | | | 33 | 3
(4 rows)
SELECT * FROM
@@ -5645,18 +5645,20 @@ select d.* from d left join (select * from b group by b.id, b.c_id) s
explain (costs off)
select d.* from d left join (select distinct * from b) s
on d.a = s.id;
- QUERY PLAN
---------------------------------------
- Merge Right Join
- Merge Cond: (b.id = d.a)
- -> Unique
- -> Sort
- Sort Key: b.id, b.c_id
- -> Seq Scan on b
+ QUERY PLAN
+---------------------------------------------
+ Merge Left Join
+ Merge Cond: (d.a = s.id)
-> Sort
Sort Key: d.a
-> Seq Scan on d
-(9 rows)
+ -> Sort
+ Sort Key: s.id
+ -> Subquery Scan on s
+ -> HashAggregate
+ Group Key: b.id, b.c_id
+ -> Seq Scan on b
+(11 rows)
-- join removal is not possible here
explain (costs off)
@@ -7597,44 +7599,39 @@ select * from j1 natural join j2;
explain (verbose, costs off)
select * from j1
inner join (select distinct id from j3) j3 on j1.id = j3.id;
- QUERY PLAN
------------------------------------------
+ QUERY PLAN
+-----------------------------------
Nested Loop
Output: j1.id, j3.id
Inner Unique: true
Join Filter: (j1.id = j3.id)
- -> Unique
+ -> HashAggregate
Output: j3.id
- -> Sort
+ Group Key: j3.id
+ -> Seq Scan on public.j3
Output: j3.id
- Sort Key: j3.id
- -> Seq Scan on public.j3
- Output: j3.id
-> Seq Scan on public.j1
Output: j1.id
-(13 rows)
+(11 rows)
-- ensure group by clause allows the inner to become unique
explain (verbose, costs off)
select * from j1
inner join (select id from j3 group by id) j3 on j1.id = j3.id;
- QUERY PLAN
------------------------------------------
+ QUERY PLAN
+-----------------------------------
Nested Loop
Output: j1.id, j3.id
Inner Unique: true
Join Filter: (j1.id = j3.id)
- -> Group
+ -> HashAggregate
Output: j3.id
Group Key: j3.id
- -> Sort
+ -> Seq Scan on public.j3
Output: j3.id
- Sort Key: j3.id
- -> Seq Scan on public.j3
- Output: j3.id
-> Seq Scan on public.j1
Output: j1.id
-(14 rows)
+(11 rows)
drop table j1;
drop table j2;
diff --git a/src/test/regress/expected/merge.out b/src/test/regress/expected/merge.out
index 28a6d0ba98..51a4596e9f 100644
--- a/src/test/regress/expected/merge.out
+++ b/src/test/regress/expected/merge.out
@@ -1459,18 +1459,15 @@ WHEN MATCHED AND t.a < 10 THEN
explain_merge
--------------------------------------------------------------------
Merge on ex_mtarget t (actual rows=0 loops=1)
- -> Merge Join (actual rows=0 loops=1)
- Merge Cond: (t.a = s.a)
- -> Sort (actual rows=0 loops=1)
- Sort Key: t.a
- Sort Method: quicksort Memory: xxx
+ -> Hash Join (actual rows=0 loops=1)
+ Hash Cond: (s.a = t.a)
+ -> Seq Scan on ex_msource s (actual rows=1 loops=1)
+ -> Hash (actual rows=0 loops=1)
+ Buckets: xxx Batches: xxx Memory Usage: xxx
-> Seq Scan on ex_mtarget t (actual rows=0 loops=1)
Filter: (a < '-1000'::integer)
Rows Removed by Filter: 54
- -> Sort (never executed)
- Sort Key: s.a
- -> Seq Scan on ex_msource s (never executed)
-(12 rows)
+(9 rows)
DROP TABLE ex_msource, ex_mtarget;
DROP FUNCTION explain_merge(text);
diff --git a/src/test/regress/expected/partition_aggregate.out b/src/test/regress/expected/partition_aggregate.out
index 1b900fddf8..800114dfde 100644
--- a/src/test/regress/expected/partition_aggregate.out
+++ b/src/test/regress/expected/partition_aggregate.out
@@ -948,12 +948,12 @@ SET parallel_setup_cost = 0;
-- is not partial agg safe.
EXPLAIN (COSTS OFF)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3 ORDER BY 1, 2, 3;
- QUERY PLAN
---------------------------------------------------------------------------------------
- Sort
- Sort Key: pagg_tab_ml.a, (sum(pagg_tab_ml.b)), (array_agg(DISTINCT pagg_tab_ml.c))
- -> Gather
- Workers Planned: 2
+ QUERY PLAN
+--------------------------------------------------------------------------------------------
+ Gather Merge
+ Workers Planned: 2
+ -> Sort
+ Sort Key: pagg_tab_ml.a, (sum(pagg_tab_ml.b)), (array_agg(DISTINCT pagg_tab_ml.c))
-> Parallel Append
-> GroupAggregate
Group Key: pagg_tab_ml.a
@@ -1380,28 +1380,26 @@ SELECT x, sum(y), avg(y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) <
-- When GROUP BY clause does not match; partial aggregation is performed for each partition.
EXPLAIN (COSTS OFF)
SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) < 12 ORDER BY 1, 2, 3;
- QUERY PLAN
--------------------------------------------------------------------------------------------
+ QUERY PLAN
+-------------------------------------------------------------------------------------
Sort
Sort Key: pagg_tab_para.y, (sum(pagg_tab_para.x)), (avg(pagg_tab_para.x))
- -> Finalize GroupAggregate
+ -> Finalize HashAggregate
Group Key: pagg_tab_para.y
Filter: (avg(pagg_tab_para.x) < '12'::numeric)
- -> Gather Merge
+ -> Gather
Workers Planned: 2
- -> Sort
- Sort Key: pagg_tab_para.y
- -> Parallel Append
- -> Partial HashAggregate
- Group Key: pagg_tab_para.y
- -> Parallel Seq Scan on pagg_tab_para_p1 pagg_tab_para
- -> Partial HashAggregate
- Group Key: pagg_tab_para_1.y
- -> Parallel Seq Scan on pagg_tab_para_p2 pagg_tab_para_1
- -> Partial HashAggregate
- Group Key: pagg_tab_para_2.y
- -> Parallel Seq Scan on pagg_tab_para_p3 pagg_tab_para_2
-(19 rows)
+ -> Parallel Append
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para.y
+ -> Parallel Seq Scan on pagg_tab_para_p1 pagg_tab_para
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para_1.y
+ -> Parallel Seq Scan on pagg_tab_para_p2 pagg_tab_para_1
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para_2.y
+ -> Parallel Seq Scan on pagg_tab_para_p3 pagg_tab_para_2
+(17 rows)
SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) < 12 ORDER BY 1, 2, 3;
y | sum | avg | count
diff --git a/src/test/regress/expected/partition_join.out b/src/test/regress/expected/partition_join.out
index 6560fe2416..680e82cfc5 100644
--- a/src/test/regress/expected/partition_join.out
+++ b/src/test/regress/expected/partition_join.out
@@ -512,52 +512,41 @@ EXPLAIN (COSTS OFF)
SELECT a, b FROM prt1 FULL JOIN prt2 p2(b,a,c) USING(a,b)
WHERE a BETWEEN 490 AND 510
GROUP BY 1, 2 ORDER BY 1, 2;
- QUERY PLAN
------------------------------------------------------------------------------------------------------------------
+ QUERY PLAN
+-----------------------------------------------------------------------------------------------------------
Group
Group Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Merge Append
+ -> Sort
Sort Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Group
- Group Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Sort
- Sort Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Merge Full Join
- Merge Cond: ((prt1.a = p2.a) AND (prt1.b = p2.b))
- Filter: ((COALESCE(prt1.a, p2.a) >= 490) AND (COALESCE(prt1.a, p2.a) <= 510))
- -> Sort
- Sort Key: prt1.a, prt1.b
- -> Seq Scan on prt1_p1 prt1
- -> Sort
- Sort Key: p2.a, p2.b
- -> Seq Scan on prt2_p1 p2
- -> Group
- Group Key: (COALESCE(prt1_1.a, p2_1.a)), (COALESCE(prt1_1.b, p2_1.b))
- -> Sort
- Sort Key: (COALESCE(prt1_1.a, p2_1.a)), (COALESCE(prt1_1.b, p2_1.b))
- -> Merge Full Join
- Merge Cond: ((prt1_1.a = p2_1.a) AND (prt1_1.b = p2_1.b))
- Filter: ((COALESCE(prt1_1.a, p2_1.a) >= 490) AND (COALESCE(prt1_1.a, p2_1.a) <= 510))
- -> Sort
- Sort Key: prt1_1.a, prt1_1.b
- -> Seq Scan on prt1_p2 prt1_1
- -> Sort
- Sort Key: p2_1.a, p2_1.b
- -> Seq Scan on prt2_p2 p2_1
- -> Group
- Group Key: (COALESCE(prt1_2.a, p2_2.a)), (COALESCE(prt1_2.b, p2_2.b))
- -> Sort
- Sort Key: (COALESCE(prt1_2.a, p2_2.a)), (COALESCE(prt1_2.b, p2_2.b))
- -> Merge Full Join
- Merge Cond: ((prt1_2.a = p2_2.a) AND (prt1_2.b = p2_2.b))
- Filter: ((COALESCE(prt1_2.a, p2_2.a) >= 490) AND (COALESCE(prt1_2.a, p2_2.a) <= 510))
- -> Sort
- Sort Key: prt1_2.a, prt1_2.b
- -> Seq Scan on prt1_p3 prt1_2
- -> Sort
- Sort Key: p2_2.a, p2_2.b
- -> Seq Scan on prt2_p3 p2_2
-(43 rows)
+ -> Append
+ -> Merge Full Join
+ Merge Cond: ((prt1_1.a = p2_1.a) AND (prt1_1.b = p2_1.b))
+ Filter: ((COALESCE(prt1_1.a, p2_1.a) >= 490) AND (COALESCE(prt1_1.a, p2_1.a) <= 510))
+ -> Sort
+ Sort Key: prt1_1.a, prt1_1.b
+ -> Seq Scan on prt1_p1 prt1_1
+ -> Sort
+ Sort Key: p2_1.a, p2_1.b
+ -> Seq Scan on prt2_p1 p2_1
+ -> Merge Full Join
+ Merge Cond: ((prt1_2.a = p2_2.a) AND (prt1_2.b = p2_2.b))
+ Filter: ((COALESCE(prt1_2.a, p2_2.a) >= 490) AND (COALESCE(prt1_2.a, p2_2.a) <= 510))
+ -> Sort
+ Sort Key: prt1_2.a, prt1_2.b
+ -> Seq Scan on prt1_p2 prt1_2
+ -> Sort
+ Sort Key: p2_2.a, p2_2.b
+ -> Seq Scan on prt2_p2 p2_2
+ -> Merge Full Join
+ Merge Cond: ((prt1_3.a = p2_3.a) AND (prt1_3.b = p2_3.b))
+ Filter: ((COALESCE(prt1_3.a, p2_3.a) >= 490) AND (COALESCE(prt1_3.a, p2_3.a) <= 510))
+ -> Sort
+ Sort Key: prt1_3.a, prt1_3.b
+ -> Seq Scan on prt1_p3 prt1_3
+ -> Sort
+ Sort Key: p2_3.a, p2_3.b
+ -> Seq Scan on prt2_p3 p2_3
+(32 rows)
SELECT a, b FROM prt1 FULL JOIN prt2 p2(b,a,c) USING(a,b)
WHERE a BETWEEN 490 AND 510
diff --git a/src/test/regress/expected/sysviews.out b/src/test/regress/expected/sysviews.out
index aae5d51e1c..487fe26906 100644
--- a/src/test/regress/expected/sysviews.out
+++ b/src/test/regress/expected/sysviews.out
@@ -114,6 +114,7 @@ select name, setting from pg_settings where name like 'enable%';
enable_async_append | on
enable_bitmapscan | on
enable_gathermerge | on
+ enable_group_by_reordering | on
enable_hashagg | on
enable_hashjoin | on
enable_incremental_sort | on
@@ -132,7 +133,7 @@ select name, setting from pg_settings where name like 'enable%';
enable_seqscan | on
enable_sort | on
enable_tidscan | on
-(21 rows)
+(22 rows)
-- There are always wait event descriptions for various types.
select type, count(*) > 0 as ok FROM pg_wait_events
diff --git a/src/test/regress/expected/union.out b/src/test/regress/expected/union.out
index e2613d6777..220f0c1629 100644
--- a/src/test/regress/expected/union.out
+++ b/src/test/regress/expected/union.out
@@ -1303,24 +1303,22 @@ select distinct q1 from
union all
select distinct * from int8_tbl i82) ss
where q2 = q2;
- QUERY PLAN
-----------------------------------------------------------
- Unique
- -> Merge Append
- Sort Key: "*SELECT* 1".q1
+ QUERY PLAN
+----------------------------------------------------
+ HashAggregate
+ Group Key: "*SELECT* 1".q1
+ -> Append
-> Subquery Scan on "*SELECT* 1"
- -> Unique
- -> Sort
- Sort Key: i81.q1, i81.q2
- -> Seq Scan on int8_tbl i81
- Filter: (q2 IS NOT NULL)
+ -> HashAggregate
+ Group Key: i81.q1, i81.q2
+ -> Seq Scan on int8_tbl i81
+ Filter: (q2 IS NOT NULL)
-> Subquery Scan on "*SELECT* 2"
- -> Unique
- -> Sort
- Sort Key: i82.q1, i82.q2
- -> Seq Scan on int8_tbl i82
- Filter: (q2 IS NOT NULL)
-(15 rows)
+ -> HashAggregate
+ Group Key: i82.q1, i82.q2
+ -> Seq Scan on int8_tbl i82
+ Filter: (q2 IS NOT NULL)
+(13 rows)
select distinct q1 from
(select distinct * from int8_tbl i81
@@ -1339,24 +1337,22 @@ select distinct q1 from
union all
select distinct * from int8_tbl i82) ss
where -q1 = q2;
- QUERY PLAN
---------------------------------------------------------
- Unique
- -> Merge Append
- Sort Key: "*SELECT* 1".q1
+ QUERY PLAN
+--------------------------------------------------
+ HashAggregate
+ Group Key: "*SELECT* 1".q1
+ -> Append
-> Subquery Scan on "*SELECT* 1"
- -> Unique
- -> Sort
- Sort Key: i81.q1, i81.q2
- -> Seq Scan on int8_tbl i81
- Filter: ((- q1) = q2)
+ -> HashAggregate
+ Group Key: i81.q1, i81.q2
+ -> Seq Scan on int8_tbl i81
+ Filter: ((- q1) = q2)
-> Subquery Scan on "*SELECT* 2"
- -> Unique
- -> Sort
- Sort Key: i82.q1, i82.q2
- -> Seq Scan on int8_tbl i82
- Filter: ((- q1) = q2)
-(15 rows)
+ -> HashAggregate
+ Group Key: i82.q1, i82.q2
+ -> Seq Scan on int8_tbl i82
+ Filter: ((- q1) = q2)
+(13 rows)
select distinct q1 from
(select distinct * from int8_tbl i81
diff --git a/src/test/regress/sql/aggregates.sql b/src/test/regress/sql/aggregates.sql
index 75c78be640..5c4e768825 100644
--- a/src/test/regress/sql/aggregates.sql
+++ b/src/test/regress/sql/aggregates.sql
@@ -1172,6 +1172,105 @@ SELECT balk(hundred) FROM tenk1;
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+
+SELECT
+ i AS id,
+ i/2 AS p,
+ format('%60s', i%2) AS v,
+ i/4 AS c,
+ i/8 AS d,
+ (random() * (10000/8))::int as e --the same as d but no correlation with p
+ INTO btg
+FROM
+ generate_series(1, 10000) i;
+
+VACUUM btg;
+ANALYZE btg;
+
+-- GROUP BY optimization by reorder columns by frequency
+
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+
+CREATE STATISTICS btg_dep ON d, e, p FROM btg;
+ANALYZE btg;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+
+
+-- GROUP BY optimization by reorder columns by index scan
+
+CREATE INDEX ON btg(p, v);
+SET enable_seqscan=off;
+SET enable_bitmapscan=off;
+VACUUM btg;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d ORDER BY p, v;
+
+DROP TABLE btg;
+
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+RESET enable_seqscan;
+RESET enable_bitmapscan;
+
+
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/sql/incremental_sort.sql b/src/test/regress/sql/incremental_sort.sql
index ab471bdfff..70f311a6f2 100644
--- a/src/test/regress/sql/incremental_sort.sql
+++ b/src/test/regress/sql/incremental_sort.sql
@@ -208,7 +208,7 @@ set parallel_tuple_cost = 0;
set max_parallel_workers_per_gather = 2;
create table t (a int, b int, c int);
-insert into t select mod(i,10),mod(i,10),i from generate_series(1,10000) s(i);
+insert into t select mod(i,10),mod(i,10),i from generate_series(1,60000) s(i);
create index on t (a);
analyze t;
--
2.42.0
Attachments:
[text/plain] 0001-Explore-alternative-orderings-of-group-by-pathkeys-d-20231001.patch (97.4K, ../../[email protected]/2-0001-Explore-alternative-orderings-of-group-by-pathkeys-d-20231001.patch)
download | inline diff:
From 2f5a42c8a53286f830e3376ff4d3f8b7d4215b4b Mon Sep 17 00:00:00 2001
From: "Andrey V. Lepikhov" <[email protected]>
Date: Wed, 13 Sep 2023 11:20:03 +0700
Subject: [PATCH] Explore alternative orderings of group-by pathkeys during
optimization.
When evaluating a query with a multi-column GROUP BY clause using sort,
the cost may depend heavily on the order in which the keys are compared when
building the groups. Grouping does not imply any ordering, so we can compare
the keys in arbitrary order, and a Hash Agg leverages this. But for Group Agg,
we simply compared keys in the order specified in the query. This commit
explores alternative ordering of the keys, trying to find a cheaper one.
In principle, we might generate grouping paths for all permutations of the keys
and leave the rest to the optimizer. But that might get very expensive, so we
try to pick only a couple interesting orderings based on both local and global
information.
When planning the grouping path, we explore statistics (number of distinct
values, cost of the comparison function) for the keys and reorder them
to minimize comparison costs. Intuitively, it may be better to perform more
expensive comparisons (for complex data types, etc.) last because maybe
the cheaper comparisons will be enough. Similarly, the higher the cardinality
of a key, the lower the probability we'll need to compare more keys. The patch
generates and costs various orderings, picking the cheapest ones.
The ordering of group keys may interact with other parts of the query, some of
which may not be known while planning the grouping. For example, there may be
an explicit ORDER BY clause or some other ordering-dependent operation higher up
in the query, and using the same ordering may allow using either incremental
sort or even eliminating the sort entirely.
The patch generates orderings and picks those, minimizing the comparison cost
(for various path keys), and then adds orderings that might be useful for
operations higher up in the plan (ORDER BY, etc.). Finally, it always keeps
the ordering specified in the query, assuming the user might have additional
insights.
This introduces a new GUC enable_group_by_reordering so that the optimization
may be disabled if needed.
---
.../postgres_fdw/expected/postgres_fdw.out | 36 +-
src/backend/optimizer/path/costsize.c | 363 ++++++++++-
src/backend/optimizer/path/equivclass.c | 13 +-
src/backend/optimizer/path/pathkeys.c | 600 ++++++++++++++++++
src/backend/optimizer/plan/planner.c | 465 ++++++++------
src/backend/optimizer/util/pathnode.c | 4 +-
src/backend/utils/adt/selfuncs.c | 38 +-
src/backend/utils/misc/guc_tables.c | 10 +
src/backend/utils/misc/postgresql.conf.sample | 1 +
src/include/nodes/pathnodes.h | 10 +
src/include/optimizer/cost.h | 4 +-
src/include/optimizer/paths.h | 7 +
src/include/utils/selfuncs.h | 5 +
src/test/regress/expected/aggregates.out | 244 ++++++-
.../regress/expected/incremental_sort.out | 2 +-
src/test/regress/expected/join.out | 51 +-
src/test/regress/expected/merge.out | 15 +-
.../regress/expected/partition_aggregate.out | 44 +-
src/test/regress/expected/partition_join.out | 75 +--
src/test/regress/expected/sysviews.out | 3 +-
src/test/regress/expected/union.out | 60 +-
src/test/regress/sql/aggregates.sql | 99 +++
src/test/regress/sql/incremental_sort.sql | 2 +-
23 files changed, 1769 insertions(+), 382 deletions(-)
diff --git a/contrib/postgres_fdw/expected/postgres_fdw.out b/contrib/postgres_fdw/expected/postgres_fdw.out
index 144c114d0f..63af7feabe 100644
--- a/contrib/postgres_fdw/expected/postgres_fdw.out
+++ b/contrib/postgres_fdw/expected/postgres_fdw.out
@@ -2319,18 +2319,21 @@ SELECT t1."C 1" FROM "S 1"."T 1" t1, LATERAL (SELECT DISTINCT t2.c1, t3.c1 FROM
-- join with pseudoconstant quals
EXPLAIN (VERBOSE, COSTS OFF)
SELECT t1.c1, t2.c1 FROM ft1 t1 JOIN ft2 t2 ON (t1.c1 = t2.c1 AND CURRENT_USER = SESSION_USER) ORDER BY t1.c3, t1.c1 OFFSET 100 LIMIT 10;
- QUERY PLAN
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+ QUERY PLAN
+-----------------------------------------------------------------------------------------------------------------------------------------------
Limit
Output: t1.c1, t2.c1, t1.c3
- -> Result
+ -> Sort
Output: t1.c1, t2.c1, t1.c3
- One-Time Filter: (CURRENT_USER = SESSION_USER)
- -> Foreign Scan
- Output: t1.c1, t1.c3, t2.c1
- Relations: (public.ft1 t1) INNER JOIN (public.ft2 t2)
- Remote SQL: SELECT r1."C 1", r2."C 1", r1.c3 FROM ("S 1"."T 1" r1 INNER JOIN "S 1"."T 1" r2 ON (((r2."C 1" = r1."C 1")))) ORDER BY r1.c3 ASC NULLS LAST, r1."C 1" ASC NULLS LAST
-(9 rows)
+ Sort Key: t1.c3, t1.c1
+ -> Result
+ Output: t1.c1, t2.c1, t1.c3
+ One-Time Filter: (CURRENT_USER = SESSION_USER)
+ -> Foreign Scan
+ Output: t1.c1, t1.c3, t2.c1
+ Relations: (public.ft1 t1) INNER JOIN (public.ft2 t2)
+ Remote SQL: SELECT r1."C 1", r2."C 1", r1.c3 FROM ("S 1"."T 1" r1 INNER JOIN "S 1"."T 1" r2 ON (((r2."C 1" = r1."C 1"))))
+(12 rows)
-- non-Var items in targetlist of the nullable rel of a join preventing
-- push-down in some cases
@@ -9609,13 +9612,16 @@ SELECT t1.a,t2.b,t3.c FROM fprt1 t1 INNER JOIN fprt2 t2 ON (t1.a = t2.b) INNER J
-- left outer join + nullable clause
EXPLAIN (VERBOSE, COSTS OFF)
SELECT t1.a,t2.b,t2.c FROM fprt1 t1 LEFT JOIN (SELECT * FROM fprt2 WHERE a < 10) t2 ON (t1.a = t2.b and t1.b = t2.a) WHERE t1.a < 10 ORDER BY 1,2,3;
- QUERY PLAN
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
- Foreign Scan
+ QUERY PLAN
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+ Sort
Output: t1.a, fprt2.b, fprt2.c
- Relations: (public.ftprt1_p1 t1) LEFT JOIN (public.ftprt2_p1 fprt2)
- Remote SQL: SELECT r5.a, r6.b, r6.c FROM (public.fprt1_p1 r5 LEFT JOIN public.fprt2_p1 r6 ON (((r5.a = r6.b)) AND ((r5.b = r6.a)) AND ((r6.a < 10)))) WHERE ((r5.a < 10)) ORDER BY r5.a ASC NULLS LAST, r6.b ASC NULLS LAST, r6.c ASC NULLS LAST
-(4 rows)
+ Sort Key: t1.a, fprt2.b, fprt2.c
+ -> Foreign Scan
+ Output: t1.a, fprt2.b, fprt2.c
+ Relations: (public.ftprt1_p1 t1) LEFT JOIN (public.ftprt2_p1 fprt2)
+ Remote SQL: SELECT r5.a, r6.b, r6.c FROM (public.fprt1_p1 r5 LEFT JOIN public.fprt2_p1 r6 ON (((r5.a = r6.b)) AND ((r5.b = r6.a)) AND ((r6.a < 10)))) WHERE ((r5.a < 10))
+(7 rows)
SELECT t1.a,t2.b,t2.c FROM fprt1 t1 LEFT JOIN (SELECT * FROM fprt2 WHERE a < 10) t2 ON (t1.a = t2.b and t1.b = t2.a) WHERE t1.a < 10 ORDER BY 1,2,3;
a | b | c
diff --git a/src/backend/optimizer/path/costsize.c b/src/backend/optimizer/path/costsize.c
index d6ceafd51c..6f11f679ba 100644
--- a/src/backend/optimizer/path/costsize.c
+++ b/src/backend/optimizer/path/costsize.c
@@ -1816,6 +1816,327 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
rterm->pathtarget->width);
}
+/*
+ * is_fake_var
+ * Workaround for generate_append_tlist() which generates fake Vars with
+ * varno == 0, that will cause a fail of estimate_num_group() call
+ *
+ * XXX Ummm, why would estimate_num_group fail with this?
+ */
+static bool
+is_fake_var(Expr *expr)
+{
+ if (IsA(expr, RelabelType))
+ expr = (Expr *) ((RelabelType *) expr)->arg;
+
+ return (IsA(expr, Var) && ((Var *) expr)->varno == 0);
+}
+
+/*
+ * get_width_cost_multiplier
+ * Returns relative complexity of comparing two values based on its width.
+ * The idea behind is that the comparison becomes more expensive the longer the
+ * value is. Return value is in cpu_operator_cost units.
+ */
+static double
+get_width_cost_multiplier(PlannerInfo *root, Expr *expr)
+{
+ double width = -1.0; /* fake value */
+
+ if (IsA(expr, RelabelType))
+ expr = (Expr *) ((RelabelType *) expr)->arg;
+
+ /* Try to find actual stat in corresponding relation */
+ if (IsA(expr, Var))
+ {
+ Var *var = (Var *) expr;
+
+ if (var->varno > 0 && var->varno < root->simple_rel_array_size)
+ {
+ RelOptInfo *rel = root->simple_rel_array[var->varno];
+
+ if (rel != NULL &&
+ var->varattno >= rel->min_attr &&
+ var->varattno <= rel->max_attr)
+ {
+ int ndx = var->varattno - rel->min_attr;
+
+ if (rel->attr_widths[ndx] > 0)
+ width = rel->attr_widths[ndx];
+ }
+ }
+ }
+
+ /* Didn't find any actual stats, try using type width instead. */
+ if (width < 0.0)
+ {
+ Node *node = (Node *) expr;
+
+ width = get_typavgwidth(exprType(node), exprTypmod(node));
+ }
+
+ /*
+ * Values are passed as Datum type, so comparisons can't be cheaper than
+ * comparing a Datum value.
+ *
+ * FIXME I find this reasoning questionable. We may pass int2, and
+ * comparing it is probably a bit cheaper than comparing a bigint.
+ */
+ if (width <= sizeof(Datum))
+ return 1.0;
+
+ /*
+ * We consider the cost of a comparison not to be directly proportional to
+ * width of the argument, because widths of the arguments could be
+ * slightly different (we only know the average width for the whole
+ * column). So we use log16(width) as an estimate.
+ */
+ return 1.0 + 0.125 * LOG2(width / sizeof(Datum));
+}
+
+/*
+ * compute_cpu_sort_cost
+ * compute CPU cost of sort (i.e. in-memory)
+ *
+ * The main thing we need to calculate to estimate sort CPU costs is the number
+ * of calls to the comparator functions. The difficulty is that for multi-column
+ * sorts there may be different data types involved (for some of which the calls
+ * may be much more expensive). Furthermore, columns may have a very different
+ * number of distinct values - the higher the number, the fewer comparisons will
+ * be needed for the following columns.
+ *
+ * The algorithm is incremental - we add pathkeys one by one, and at each step we
+ * estimate the number of necessary comparisons (based on the number of distinct
+ * groups in the current pathkey prefix and the new pathkey), and the comparison
+ * costs (which is data type specific).
+ *
+ * Estimation of the number of comparisons is based on ideas from:
+ *
+ * "Quicksort Is Optimal", Robert Sedgewick, Jon Bentley, 2002
+ * [https://www.cs.princeton.edu/~rs/talks/QuicksortIsOptimal.pdf]
+ *
+ * In term of that paper, let N - number of tuples, Xi - number of identical
+ * tuples with value Ki, then the estimate of number of comparisons is:
+ *
+ * log(N! / (X1! * X2! * ..)) ~ sum(Xi * log(N/Xi))
+ *
+ * We assume all Xi the same because now we don't have any estimation of
+ * group sizes, we have only know the estimate of number of groups (distinct
+ * values). In that case, formula becomes:
+ *
+ * N * log(NumberOfGroups)
+ *
+ * For multi-column sorts we need to estimate the number of comparisons for
+ * each individual column - for example with columns (c1, c2, ..., ck) we
+ * can estimate that number of comparisons on ck is roughly
+ *
+ * ncomparisons(c1, c2, ..., ck) / ncomparisons(c1, c2, ..., c(k-1))
+ *
+ * Let k be a column number, Gk - number of groups defined by k columns, and Fk
+ * the cost of the comparison is
+ *
+ * N * sum( Fk * log(Gk) )
+ *
+ * Note: We also consider column width, not just the comparator cost.
+ *
+ * NOTE: some callers currently pass NIL for pathkeys because they
+ * can't conveniently supply the sort keys. In this case, it will fallback to
+ * simple comparison cost estimate.
+ */
+static Cost
+compute_cpu_sort_cost(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
+ Cost comparison_cost, double tuples, double output_tuples,
+ bool heapSort)
+{
+ Cost per_tuple_cost = 0.0;
+ ListCell *lc;
+ List *pathkeyExprs = NIL;
+ double tuplesPerPrevGroup = tuples;
+ double totalFuncCost = 1.0;
+ bool has_fake_var = false;
+ int i = 0;
+ Oid prev_datatype = InvalidOid;
+ List *cache_varinfos = NIL;
+
+ /* fallback if pathkeys is unknown */
+ if (list_length(pathkeys) == 0)
+ {
+ /*
+ * If we'll use a bounded heap-sort keeping just K tuples in memory,
+ * for a total number of tuple comparisons of N log2 K; but the
+ * constant factor is a bit higher than for quicksort. Tweak it so
+ * that the cost curve is continuous at the crossover point.
+ */
+ output_tuples = (heapSort) ? 2.0 * output_tuples : tuples;
+ per_tuple_cost += 2.0 * cpu_operator_cost * LOG2(output_tuples);
+
+ /* add cost provided by caller */
+ per_tuple_cost += comparison_cost;
+
+ return per_tuple_cost * tuples;
+ }
+
+ /*
+ * Computing total cost of sorting takes into account the per-column
+ * comparison function cost. We try to compute the needed number of
+ * comparisons per column.
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(lc);
+ EquivalenceMember *em;
+ double nGroups,
+ correctedNGroups;
+ Cost funcCost = 1.0;
+
+ /*
+ * We believe that equivalence members aren't very different, so, to
+ * estimate cost we consider just the first member.
+ */
+ em = (EquivalenceMember *) linitial(pathkey->pk_eclass->ec_members);
+
+ if (em->em_datatype != InvalidOid)
+ {
+ /* do not lookup funcCost if the data type is the same */
+ if (prev_datatype != em->em_datatype)
+ {
+ Oid sortop;
+ QualCost cost;
+
+ sortop = get_opfamily_member(pathkey->pk_opfamily,
+ em->em_datatype, em->em_datatype,
+ pathkey->pk_strategy);
+
+ cost.startup = 0;
+ cost.per_tuple = 0;
+ add_function_cost(root, get_opcode(sortop), NULL, &cost);
+
+ /*
+ * add_function_cost returns the product of cpu_operator_cost
+ * and procost, but we need just procost, co undo that.
+ */
+ funcCost = cost.per_tuple / cpu_operator_cost;
+
+ prev_datatype = em->em_datatype;
+ }
+ }
+
+ /* factor in the width of the values in this column */
+ funcCost *= get_width_cost_multiplier(root, em->em_expr);
+
+ /* now we have per-key cost, so add to the running total */
+ totalFuncCost += funcCost;
+
+ /* remember if we have found a fake Var in pathkeys */
+ has_fake_var |= is_fake_var(em->em_expr);
+ pathkeyExprs = lappend(pathkeyExprs, em->em_expr);
+
+ /*
+ * We need to calculate the number of comparisons for this column,
+ * which requires knowing the group size. So we estimate the number of
+ * groups by calling estimate_num_groups_incremental(), which
+ * estimates the group size for "new" pathkeys.
+ *
+ * Note: estimate_num_groups_incremental does not handle fake Vars, so
+ * use a default estimate otherwise.
+ */
+ if (!has_fake_var)
+ nGroups = estimate_num_groups_incremental(root, pathkeyExprs,
+ tuplesPerPrevGroup, NULL, NULL,
+ &cache_varinfos,
+ list_length(pathkeyExprs) - 1);
+ else if (tuples > 4.0)
+
+ /*
+ * Use geometric mean as estimation if there are no stats.
+ *
+ * We don't use DEFAULT_NUM_DISTINCT here, because that's used for
+ * a single column, but here we're dealing with multiple columns.
+ */
+ nGroups = ceil(2.0 + sqrt(tuples) * (i + 1) / list_length(pathkeys));
+ else
+ nGroups = tuples;
+
+ /*
+ * Presorted keys are not considered in the cost above, but we still
+ * do have to compare them in the qsort comparator. So make sure to
+ * factor in the cost in that case.
+ */
+ if (i >= nPresortedKeys)
+ {
+ if (heapSort)
+ {
+ /*
+ * have to keep at least one group, and a multiple of group
+ * size
+ */
+ correctedNGroups = ceil(output_tuples / tuplesPerPrevGroup);
+ }
+ else
+ /* all groups in the input */
+ correctedNGroups = nGroups;
+
+ correctedNGroups = Max(1.0, ceil(correctedNGroups));
+
+ per_tuple_cost += totalFuncCost * LOG2(correctedNGroups);
+ }
+
+ i++;
+
+ /*
+ * Once we get single-row group, it means tuples in the group are
+ * unique and we can skip all remaining columns.
+ */
+ if (tuplesPerPrevGroup <= nGroups)
+ break;
+
+ /*
+ * Uniform distributions with all groups being of the same size are
+ * the best case, with nice smooth behavior. Real-world distributions
+ * tend not to be uniform, though, and we don't have any reliable
+ * easy-to-use information. As a basic defense against skewed
+ * distributions, we use a 1.5 factor to make the expected group a bit
+ * larger, but we need to be careful not to make the group larger than
+ * in the preceding step.
+ */
+ tuplesPerPrevGroup = Min(tuplesPerPrevGroup,
+ ceil(1.5 * tuplesPerPrevGroup / nGroups));
+ }
+
+ list_free(pathkeyExprs);
+
+ /* per_tuple_cost is in cpu_operator_cost units */
+ per_tuple_cost *= cpu_operator_cost;
+
+ /*
+ * Accordingly to "Introduction to algorithms", Thomas H. Cormen, Charles
+ * E. Leiserson, Ronald L. Rivest, ISBN 0-07-013143-0, quicksort
+ * estimation formula has additional term proportional to number of tuples
+ * (see Chapter 8.2 and Theorem 4.1). That affects cases with a low number
+ * of tuples, approximately less than 1e4. We could implement it as an
+ * additional multiplier under the logarithm, but we use a bit more
+ * complex formula which takes into account the number of unique tuples
+ * and it's not clear how to combine the multiplier with the number of
+ * groups. Estimate it as 10 cpu_operator_cost units.
+ */
+ per_tuple_cost += 10 * cpu_operator_cost;
+
+ per_tuple_cost += comparison_cost;
+
+ return tuples * per_tuple_cost;
+}
+
+/*
+ * simple wrapper just to estimate best sort path
+ */
+Cost
+cost_sort_estimate(PlannerInfo *root, List *pathkeys, int nPresortedKeys,
+ double tuples)
+{
+ return compute_cpu_sort_cost(root, pathkeys, nPresortedKeys,
+ 0, tuples, tuples, false);
+}
+
/*
* cost_tuplesort
* Determines and returns the cost of sorting a relation using tuplesort,
@@ -1832,7 +2153,7 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
* number of initial runs formed and M is the merge order used by tuplesort.c.
* Since the average initial run should be about sort_mem, we have
* disk traffic = 2 * relsize * ceil(logM(p / sort_mem))
- * cpu = comparison_cost * t * log2(t)
+ * and cpu cost (computed by compute_cpu_sort_cost()).
*
* If the sort is bounded (i.e., only the first k result tuples are needed)
* and k tuples can fit into sort_mem, we use a heap method that keeps only
@@ -1851,9 +2172,11 @@ cost_recursive_union(Path *runion, Path *nrterm, Path *rterm)
* 'comparison_cost' is the extra cost per comparison, if any
* 'sort_mem' is the number of kilobytes of work memory allowed for the sort
* 'limit_tuples' is the bound on the number of output tuples; -1 if no bound
+ * 'startup_cost' is expected to be 0 at input. If there is "input cost" it should
+ * be added by caller later
*/
static void
-cost_tuplesort(Cost *startup_cost, Cost *run_cost,
+cost_tuplesort(PlannerInfo *root, List *pathkeys, Cost *startup_cost, Cost *run_cost,
double tuples, int width,
Cost comparison_cost, int sort_mem,
double limit_tuples)
@@ -1870,9 +2193,6 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
if (tuples < 2.0)
tuples = 2.0;
- /* Include the default cost-per-comparison */
- comparison_cost += 2.0 * cpu_operator_cost;
-
/* Do we have a useful LIMIT? */
if (limit_tuples > 0 && limit_tuples < tuples)
{
@@ -1896,12 +2216,10 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
double log_runs;
double npageaccesses;
- /*
- * CPU costs
- *
- * Assume about N log2 N comparisons
- */
- *startup_cost = comparison_cost * tuples * LOG2(tuples);
+ /* CPU costs */
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ tuples, false);
/* Disk costs */
@@ -1917,18 +2235,17 @@ cost_tuplesort(Cost *startup_cost, Cost *run_cost,
}
else if (tuples > 2 * output_tuples || input_bytes > sort_mem_bytes)
{
- /*
- * We'll use a bounded heap-sort keeping just K tuples in memory, for
- * a total number of tuple comparisons of N log2 K; but the constant
- * factor is a bit higher than for quicksort. Tweak it so that the
- * cost curve is continuous at the crossover point.
- */
- *startup_cost = comparison_cost * tuples * LOG2(2.0 * output_tuples);
+ /* We'll use a bounded heap-sort keeping just K tuples in memory. */
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ output_tuples, true);
}
else
{
/* We'll use plain quicksort on all the input tuples */
- *startup_cost = comparison_cost * tuples * LOG2(tuples);
+ *startup_cost = compute_cpu_sort_cost(root, pathkeys, 0,
+ comparison_cost, tuples,
+ tuples, false);
}
/*
@@ -2042,7 +2359,7 @@ cost_incremental_sort(Path *path,
* Estimate the average cost of sorting of one group where presorted keys
* are equal.
*/
- cost_tuplesort(&group_startup_cost, &group_run_cost,
+ cost_tuplesort(root, pathkeys, &group_startup_cost, &group_run_cost,
group_tuples, width, comparison_cost, sort_mem,
limit_tuples);
@@ -2102,7 +2419,7 @@ cost_sort(Path *path, PlannerInfo *root,
Cost startup_cost;
Cost run_cost;
- cost_tuplesort(&startup_cost, &run_cost,
+ cost_tuplesort(root, pathkeys, &startup_cost, &run_cost,
tuples, width,
comparison_cost, sort_mem,
limit_tuples);
@@ -2200,7 +2517,7 @@ append_nonpartial_cost(List *subpaths, int numpaths, int parallel_workers)
* Determines and returns the cost of an Append node.
*/
void
-cost_append(AppendPath *apath)
+cost_append(AppendPath *apath, PlannerInfo *root)
{
ListCell *l;
@@ -2268,7 +2585,7 @@ cost_append(AppendPath *apath)
* any child.
*/
cost_sort(&sort_path,
- NULL, /* doesn't currently need root */
+ root,
pathkeys,
subpath->total_cost,
subpath->rows,
diff --git a/src/backend/optimizer/path/equivclass.c b/src/backend/optimizer/path/equivclass.c
index 7fa502d6e2..07edd4f38e 100644
--- a/src/backend/optimizer/path/equivclass.c
+++ b/src/backend/optimizer/path/equivclass.c
@@ -652,7 +652,18 @@ get_eclass_for_sort_expr(PlannerInfo *root,
if (opcintype == cur_em->em_datatype &&
equal(expr, cur_em->em_expr))
- return cur_ec; /* Match! */
+ {
+ /*
+ * Match!
+ *
+ * Copy the sortref if it wasn't set yet. That may happen if
+ * the ec was constructed from WHERE clause, i.e. it doesn't
+ * have a target reference at all.
+ */
+ if (cur_ec->ec_sortref == 0 && sortref > 0)
+ cur_ec->ec_sortref = sortref;
+ return cur_ec;
+ }
}
}
diff --git a/src/backend/optimizer/path/pathkeys.c b/src/backend/optimizer/path/pathkeys.c
index fdb60aaa8d..ac3e13ddd4 100644
--- a/src/backend/optimizer/path/pathkeys.c
+++ b/src/backend/optimizer/path/pathkeys.c
@@ -17,17 +17,24 @@
*/
#include "postgres.h"
+#include <float.h>
+
+#include "miscadmin.h"
#include "access/stratnum.h"
#include "catalog/pg_opfamily.h"
#include "nodes/makefuncs.h"
#include "nodes/nodeFuncs.h"
#include "nodes/plannodes.h"
+#include "optimizer/cost.h"
#include "optimizer/optimizer.h"
#include "optimizer/pathnode.h"
#include "optimizer/paths.h"
#include "partitioning/partbounds.h"
#include "utils/lsyscache.h"
+#include "utils/selfuncs.h"
+/* Consider reordering of GROUP BY keys? */
+bool enable_group_by_reordering = true;
static bool pathkey_is_redundant(PathKey *new_pathkey, List *pathkeys);
static bool matches_boolean_partition_clause(RestrictInfo *rinfo,
@@ -350,6 +357,546 @@ pathkeys_contained_in(List *keys1, List *keys2)
return false;
}
+/*
+ * group_keys_reorder_by_pathkeys
+ * Reorder GROUP BY keys to match pathkeys of input path.
+ *
+ * Function returns new lists (pathkeys and clauses), original GROUP BY lists
+ * stay untouched.
+ *
+ * Returns the number of GROUP BY keys with a matching pathkey.
+ */
+int
+group_keys_reorder_by_pathkeys(List *pathkeys, List **group_pathkeys,
+ List **group_clauses,
+ int num_groupby_pathkeys)
+{
+ List *new_group_pathkeys = NIL,
+ *new_group_clauses = NIL;
+ ListCell *lc;
+ int n;
+
+ if (pathkeys == NIL || *group_pathkeys == NIL)
+ return 0;
+
+ /*
+ * Walk the pathkeys (determining ordering of the input path) and see if
+ * there's a matching GROUP BY key. If we find one, we append it to the
+ * list, and do the same for the clauses.
+ *
+ * Once we find the first pathkey without a matching GROUP BY key, the
+ * rest of the pathkeys are useless and can't be used to evaluate the
+ * grouping, so we abort the loop and ignore the remaining pathkeys.
+ *
+ * XXX Pathkeys are built in a way to allow simply comparing pointers.
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(lc);
+ SortGroupClause *sgc;
+
+ if (foreach_current_index(lc) >= num_groupby_pathkeys ||
+ !list_member_ptr(*group_pathkeys, pathkey))
+ /* */
+ break;
+
+ /* abort on first mismatch */
+ sgc = get_sortgroupref_clause_noerr(pathkey->pk_eclass->ec_sortref,
+ *group_clauses);
+ if (!sgc)
+ /* The grouping clause is not cover this paathkey */
+ break;
+
+ new_group_pathkeys = lappend(new_group_pathkeys, pathkey);
+ new_group_clauses = lappend(new_group_clauses, sgc);
+ }
+
+ /* remember the number of pathkeys with a matching GROUP BY key */
+ n = list_length(new_group_pathkeys);
+
+ /* append the remaining group pathkeys (will be treated as not sorted) */
+ *group_pathkeys = list_concat_unique_ptr(new_group_pathkeys,
+ *group_pathkeys);
+ *group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+
+ return n;
+}
+
+/*
+ * Used to generate all permutations of a pathkey list.
+ */
+typedef struct PathkeyMutatorState
+{
+ List *elemsList;
+ ListCell **elemCells;
+ void **elems;
+ int *positions;
+ int mutatorNColumns;
+ int count;
+} PathkeyMutatorState;
+
+
+/*
+ * PathkeyMutatorInit
+ * Initialize state of the permutation generator.
+ *
+ * We want to generate permutations of elements in the "elems" list. We may want
+ * to skip some number of elements at the beginning (when treating as presorted)
+ * or at the end (we only permute a limited number of group keys).
+ *
+ * The list is decomposed into elements, and we also keep pointers to individual
+ * cells. This allows us to build the permuted list quickly and cheaply, without
+ * creating any copies.
+ */
+static void
+PathkeyMutatorInit(PathkeyMutatorState *state, List *elems, int start, int end)
+{
+ int i;
+ int n = end - start;
+ ListCell *lc;
+
+ memset(state, 0, sizeof(*state));
+
+ state->mutatorNColumns = n;
+
+ state->elemsList = list_copy(elems);
+
+ state->elems = palloc(sizeof(void *) * n);
+ state->elemCells = palloc(sizeof(ListCell *) * n);
+ state->positions = palloc(sizeof(int) * n);
+
+ i = 0;
+ for_each_cell(lc, state->elemsList, list_nth_cell(state->elemsList, start))
+ {
+ state->elemCells[i] = lc;
+ state->elems[i] = lfirst(lc);
+ state->positions[i] = i + 1;
+ i++;
+
+ if (i >= n)
+ break;
+ }
+}
+
+/* Swap two elements of an array. */
+static void
+PathkeyMutatorSwap(int *a, int i, int j)
+{
+ int s = a[i];
+
+ a[i] = a[j];
+ a[j] = s;
+}
+
+/*
+ * Generate the next permutation of elements.
+ */
+static bool
+PathkeyMutatorNextSet(int *a, int n)
+{
+ int j,
+ k,
+ l,
+ r;
+
+ j = n - 2;
+
+ while (j >= 0 && a[j] >= a[j + 1])
+ j--;
+
+ if (j < 0)
+ return false;
+
+ k = n - 1;
+
+ while (k > 0 && a[j] >= a[k])
+ k--;
+
+ PathkeyMutatorSwap(a, j, k);
+
+ l = j + 1;
+ r = n - 1;
+
+ while (l < r)
+ PathkeyMutatorSwap(a, l++, r--);
+
+ return true;
+}
+
+/*
+ * PathkeyMutatorNext
+ * Generate the next permutation of list of elements.
+ *
+ * Returns the next permutation (as a list of elements) or NIL if there are no
+ * more permutations.
+ */
+static List *
+PathkeyMutatorNext(PathkeyMutatorState *state)
+{
+ int i;
+
+ state->count++;
+
+ /* first permutation is original list */
+ if (state->count == 1)
+ return state->elemsList;
+
+ /* when there are no more permutations, return NIL */
+ if (!PathkeyMutatorNextSet(state->positions, state->mutatorNColumns))
+ {
+ pfree(state->elems);
+ pfree(state->elemCells);
+ pfree(state->positions);
+
+ list_free(state->elemsList);
+
+ return NIL;
+ }
+
+ /* update the list cells to point to the right elements */
+ for (i = 0; i < state->mutatorNColumns; i++)
+ lfirst(state->elemCells[i]) =
+ (void *) state->elems[state->positions[i] - 1];
+
+ return state->elemsList;
+}
+
+/*
+ * Cost of comparing pathkeys.
+ */
+typedef struct PathkeySortCost
+{
+ Cost cost;
+ PathKey *pathkey;
+} PathkeySortCost;
+
+static int
+pathkey_sort_cost_comparator(const void *_a, const void *_b)
+{
+ const PathkeySortCost *a = (PathkeySortCost *) _a;
+ const PathkeySortCost *b = (PathkeySortCost *) _b;
+
+ if (a->cost < b->cost)
+ return -1;
+ else if (a->cost == b->cost)
+ return 0;
+ return 1;
+}
+
+/*
+ * get_cheapest_group_keys_order
+ * Reorders the group pathkeys / clauses to minimize the comparison cost.
+ *
+ * Given the list of pathkeys in '*group_pathkeys', we try to arrange these
+ * in an order that minimizes the sort costs that will be incurred by the
+ * GROUP BY. The costs mainly depend on the cost of the sort comparator
+ * function(s) and the number of distinct values in each column of the GROUP
+ * BY clause (*group_clauses). Sorting on subsequent columns is only required
+ * for tiebreak situations where two values sort equally.
+ *
+ * In case the input is partially sorted, only the remaining pathkeys are
+ * considered. 'n_preordered' denotes how many of the leading *group_pathkeys
+ * the input is presorted by.
+ *
+ * Returns true and sets *group_pathkeys and *group_clauses to the newly
+ * ordered versions of the lists that were passed in via these parameters.
+ * If no reordering was deemed necessary then we return false, in which case
+ * the *group_pathkeys and *group_clauses lists are left untouched. The
+ * original *group_pathkeys and *group_clauses parameter values are never
+ * destructively modified in place.
+ */
+static bool
+get_cheapest_group_keys_order(PlannerInfo *root, double nrows,
+ List **group_pathkeys, List **group_clauses,
+ int n_preordered)
+{
+ List *new_group_pathkeys = NIL;
+ List *new_group_clauses = NIL;
+ List *var_group_pathkeys;
+ ListCell *cell;
+ PathkeyMutatorState mstate;
+ double cheapest_sort_cost = DBL_MAX;
+ int nFreeKeys;
+ int nToPermute;
+
+ /* If there are less than 2 unsorted pathkeys, we're done. */
+ if (root->num_groupby_pathkeys - n_preordered < 2)
+ return false;
+
+ /*
+ * We could exhaustively cost all possible orderings of the pathkeys, but
+ * for a large number of pathkeys it might be prohibitively expensive. So
+ * we try to apply simple cheap heuristics first - we sort the pathkeys by
+ * sort cost (as if the pathkey was sorted independently) and then check
+ * only the four cheapest pathkeys. The remaining pathkeys are kept
+ * ordered by cost.
+ *
+ * XXX This is a very simple heuristics, but likely to work fine for most
+ * cases (because the number of GROUP BY clauses tends to be lower than
+ * 4). But it ignores how the number of distinct values in each pathkey
+ * affects the following steps. It might be better to use "more expensive"
+ * pathkey first if it has many distinct values, because it then limits
+ * the number of comparisons for the remaining pathkeys. But evaluating
+ * that is likely quite the expensive.
+ */
+ nFreeKeys = root->num_groupby_pathkeys - n_preordered;
+ nToPermute = 4;
+ if (nFreeKeys > nToPermute)
+ {
+ PathkeySortCost *costs = palloc(sizeof(PathkeySortCost) * nFreeKeys);
+ PathkeySortCost *cost = costs;
+ int i;
+
+ /*
+ * Estimate cost for sorting individual pathkeys skipping the
+ * pre-ordered pathkeys.
+ */
+ for_each_from(cell, *group_pathkeys, n_preordered)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(cell);
+ List *to_cost;
+
+ if (foreach_current_index(cell) >= root->num_groupby_pathkeys)
+ break;
+
+ to_cost = list_make1(pathkey);
+ cost->pathkey = pathkey;
+ cost->cost = cost_sort_estimate(root, to_cost, 0, nrows);
+ cost++;
+
+ list_free(to_cost);
+ }
+
+ /* sort the pathkeys by sort cost in ascending order */
+ qsort(costs, nFreeKeys, sizeof(*costs), pathkey_sort_cost_comparator);
+
+ /*
+ * Rebuild the list of pathkeys - first the preordered ones, then the
+ * rest ordered by cost.
+ */
+ new_group_pathkeys = list_copy_head(*group_pathkeys, n_preordered);
+
+ for (i = 0; i < nFreeKeys; i++)
+ new_group_pathkeys = lappend(new_group_pathkeys, costs[i].pathkey);
+
+ new_group_pathkeys = list_copy_tail(*group_pathkeys,
+ root->num_groupby_pathkeys);
+ pfree(costs);
+ }
+ else
+ {
+ /* Copy the list, so that we can free the new list by list_free. */
+ new_group_pathkeys = list_copy(*group_pathkeys);
+ nToPermute = nFreeKeys;
+ }
+
+ Assert(list_length(new_group_pathkeys) == list_length(*group_pathkeys));
+
+ /*
+ * Generate pathkey lists with permutations of the first nToPermute
+ * pathkeys.
+ *
+ * XXX We simply calculate sort cost for each individual pathkey list, but
+ * there's room for two dynamic programming optimizations here. Firstly,
+ * we may pass the current "best" cost to cost_sort_estimate so that it
+ * can "abort" if the estimated pathkeys list exceeds it. Secondly, it
+ * could pass the return information about the position when it exceeded
+ * the cost, and we could skip all permutations with the same prefix.
+ *
+ * Imagine we've already found ordering with cost C1, and we're evaluating
+ * another ordering - cost_sort_estimate() calculates cost by adding the
+ * pathkeys one by one (more or less), and the cost only grows. If at any
+ * point it exceeds C1, it can't possibly be "better" so we can discard
+ * it. But we also know that we can discard all ordering with the same
+ * prefix, because if we're estimating (a,b,c,d) and we exceed C1 at (a,b)
+ * then the same thing will happen for any ordering with this prefix.
+ */
+ PathkeyMutatorInit(&mstate, new_group_pathkeys, n_preordered,
+ n_preordered + nToPermute);
+
+ while ((var_group_pathkeys = PathkeyMutatorNext(&mstate)) != NIL)
+ {
+ Cost cost;
+
+ cost = cost_sort_estimate(root, var_group_pathkeys, n_preordered, nrows);
+
+ if (cost < cheapest_sort_cost)
+ {
+ list_free(new_group_pathkeys);
+ new_group_pathkeys = list_copy(var_group_pathkeys);
+ cheapest_sort_cost = cost;
+ }
+ }
+
+ Assert(list_length(new_group_pathkeys) == list_length(root->group_pathkeys));
+
+ /* Reorder the group clauses according to the reordered pathkeys. */
+ foreach(cell, new_group_pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(cell);
+
+ if (foreach_current_index(cell) >= root->num_groupby_pathkeys)
+ break;
+
+ new_group_clauses =
+ lappend(new_group_clauses,
+ get_sortgroupref_clause(pathkey->pk_eclass->ec_sortref,
+ *group_clauses));
+ }
+
+ /* Just append the rest GROUP BY clauses */
+ new_group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+ *group_pathkeys = new_group_pathkeys;
+ *group_clauses = new_group_clauses;
+ Assert(list_length(new_group_clauses) == list_length(root->processed_groupClause));
+
+ return true;
+}
+
+/*
+ * get_useful_group_keys_orderings
+ * Determine which orderings of GROUP BY keys are potentially interesting.
+ *
+ * Returns list of PathKeyInfo items, each representing an interesting ordering
+ * of GROUP BY keys. Each item stores pathkeys and clauses in matching order.
+ *
+ * The function considers (and keeps) multiple group by orderings:
+ *
+ * - the original ordering, as specified by the GROUP BY clause
+ *
+ * - GROUP BY keys reordered to minimize the sort cost
+ *
+ * - GROUP BY keys reordered to match path ordering (as much as possible), with
+ * the tail reordered to minimize the sort cost
+ *
+ * - GROUP BY keys to match target ORDER BY clause (as much as possible), with
+ * the tail reordered to minimize the sort cost
+ *
+ * There are other potentially interesting orderings (e.g. it might be best to
+ * match the first ORDER BY key, order the remaining keys differently and then
+ * rely on the incremental sort to fix this), but we ignore those for now. To
+ * make this work we'd have to pretty much generate all possible permutations.
+ */
+List *
+get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys)
+{
+ Query *parse = root->parse;
+ List *infos = NIL;
+ PathKeyInfo *info;
+ int n_preordered = 0;
+
+ List *pathkeys = root->group_pathkeys;
+ List *clauses = root->processed_groupClause;
+
+ /* always return at least the original pathkeys/clauses */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+ infos = lappend(infos, info);
+
+ /*
+ * Should we try generating alternative orderings of the group keys? If
+ * not, we produce only the order specified in the query, i.e. the
+ * optimization is effectively disabled.
+ */
+ if (!enable_group_by_reordering)
+ return infos;
+
+ /* for grouping sets we can't do any reordering */
+ if (parse->groupingSets)
+ return infos;
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost, ignoring both the
+ * target ordering (ORDER BY) and ordering of the input path.
+ */
+ if (get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered))
+ {
+ Assert(list_length(pathkeys) == list_length(clauses));
+
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * If the path is sorted in some way, try reordering the group keys to
+ * match as much of the ordering as possible - we get this sort for free
+ * (mostly).
+ *
+ * We must not do this when there are no grouping sets, because those use
+ * more complex logic to decide the ordering.
+ *
+ * XXX Isn't this somewhat redundant with presorted_keys? Actually, it's
+ * more a complement, because it allows benefiting from incremental sort
+ * as much as possible.
+ *
+ * XXX This does nothing if (n_preordered == 0). We shouldn't create the
+ * info in this case.
+ */
+ if (path_pathkeys)
+ {
+ n_preordered = group_keys_reorder_by_pathkeys(path_pathkeys,
+ &pathkeys,
+ &clauses,
+ root->num_groupby_pathkeys);
+
+ /* reorder the tail to minimize sort cost */
+ get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, in
+ * which case get_cheapest_group_keys_order returns false. But we
+ * still want to keep the keys reordered to path_pathkeys.
+ */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost (this time consider
+ * the ORDER BY clause, but only if set debug_group_by_match_order_by).
+ */
+ if (root->sort_pathkeys)
+ {
+ n_preordered = group_keys_reorder_by_pathkeys(root->sort_pathkeys,
+ &pathkeys,
+ &clauses,
+ root->num_groupby_pathkeys);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, in
+ * which case get_cheapest_group_keys_order returns false. But we
+ * still want to keep the keys reordered to sort_pathkeys.
+ */
+ get_cheapest_group_keys_order(root, nrows, &pathkeys, &clauses,
+ n_preordered);
+
+ /* keep the group keys reordered to match ordering of input path */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ return infos;
+}
+
/*
* pathkeys_count_contained_in
* Same as pathkeys_contained_in, but also sets length of longest
@@ -1939,6 +2486,54 @@ pathkeys_useful_for_ordering(PlannerInfo *root, List *pathkeys)
return n_common_pathkeys;
}
+/*
+ * pathkeys_useful_for_grouping
+ * Count the number of pathkeys that are useful for grouping (instead of
+ * explicit sort)
+ *
+ * Group pathkeys could be reordered to benefit from the ordering. The
+ * ordering may not be "complete" and may require incremental sort, but that's
+ * fine. So we simply count prefix pathkeys with a matching group key, and
+ * stop once we find the first pathkey without a match.
+ *
+ * So e.g. with pathkeys (a,b,c) and group keys (a,b,e) this determines (a,b)
+ * pathkeys are useful for grouping, and we might do incremental sort to get
+ * path ordered by (a,b,e).
+ *
+ * This logic is necessary to retain paths with ordering not matching grouping
+ * keys directly, without the reordering.
+ *
+ * Returns the length of pathkey prefix with matching group keys.
+ */
+static int
+pathkeys_useful_for_grouping(PlannerInfo *root, List *pathkeys)
+{
+ ListCell *key;
+ int n = 0;
+
+ /* no special ordering requested for grouping */
+ if (root->group_pathkeys == NIL)
+ return 0;
+
+ /* unordered path */
+ if (pathkeys == NIL)
+ return 0;
+
+ /* walk the pathkeys and search for matching group key */
+ foreach(key, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(key);
+
+ /* no matching group key, we're done */
+ if (!list_member_ptr(root->group_pathkeys, pathkey))
+ break;
+
+ n++;
+ }
+
+ return n;
+}
+
/*
* truncate_useless_pathkeys
* Shorten the given pathkey list to just the useful pathkeys.
@@ -1953,6 +2548,9 @@ truncate_useless_pathkeys(PlannerInfo *root,
nuseful = pathkeys_useful_for_merging(root, rel, pathkeys);
nuseful2 = pathkeys_useful_for_ordering(root, pathkeys);
+ if (nuseful2 > nuseful)
+ nuseful = nuseful2;
+ nuseful2 = pathkeys_useful_for_grouping(root, pathkeys);
if (nuseful2 > nuseful)
nuseful = nuseful2;
@@ -1988,6 +2586,8 @@ has_useful_pathkeys(PlannerInfo *root, RelOptInfo *rel)
{
if (rel->joininfo != NIL || rel->has_eclass_joins)
return true; /* might be able to use pathkeys for merging */
+ if (root->group_pathkeys != NIL)
+ return true; /* might be able to use pathkeys for grouping */
if (root->query_pathkeys != NIL)
return true; /* might be able to use them for ordering */
return false; /* definitely useless */
diff --git a/src/backend/optimizer/plan/planner.c b/src/backend/optimizer/plan/planner.c
index aafef24cf1..4727a5a157 100644
--- a/src/backend/optimizer/plan/planner.c
+++ b/src/backend/optimizer/plan/planner.c
@@ -2811,8 +2811,9 @@ remove_useless_groupby_columns(PlannerInfo *root)
*
* In principle it might be interesting to consider other orderings of the
* GROUP BY elements, which could match the sort ordering of other
- * possible plans (eg an indexscan) and thereby reduce cost. We don't
- * bother with that, though. Hashed grouping will frequently win anyway.
+ * possible plans (eg an indexscan) and thereby reduce cost. However, we
+ * don't yet have sufficient information to do that here, so that's left until
+ * later in planning. See get_useful_group_keys_orderings().
*
* Note: we need no comparable processing of the distinctClause because
* the parser already enforced that that matches ORDER BY.
@@ -3469,6 +3470,7 @@ standard_qp_callback(PlannerInfo *root, void *extra)
tlist,
true,
&sortable);
+
if (!sortable)
{
/* Can't sort; no point in considering aggregate ordering either */
@@ -6843,90 +6845,108 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
+
+ Assert(list_length(pathkey_orderings) > 0);
- if (!is_sorted)
+ foreach(lc2, pathkey_orderings)
{
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially sorted
- * already (no need to deal with paths which have presorted
- * keys when incremental sort is disabled unless it's the
- * cheapest input path).
- */
- if (path != cheapest_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ bool is_sorted;
+ int presorted_keys;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- /*
- * We've no need to consider both a sort and incremental sort.
- * We'll just do a sort if there are no presorted keys and an
- * incremental sort when there are presorted keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
- /* Now decide what to stick atop it */
- if (parse->groupingSets)
- {
- consider_groupingsets_paths(root, grouped_rel,
- path, true, can_hash,
- gd, agg_costs, dNumGroups);
- }
- else if (parse->hasAggs)
- {
- /*
- * We have aggregation, possibly with plain GROUP BY. Make an
- * AggPath.
- */
- add_path(grouped_rel, (Path *)
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (!is_sorted)
+ {
+ /*
+ * Try at least sorting the cheapest path and also try
+ * incrementally sorting any path which is partially sorted
+ * already (no need to deal with paths which have presorted
+ * keys when incremental sort is disabled unless it's the
+ * cheapest input path).
+ */
+ if (path != cheapest_path &&
+ (presorted_keys == 0 || !enable_incremental_sort))
+ continue;
+
+ /*
+ * We've no need to consider both a sort and incremental sort.
+ * We'll just do a sort if there are no presorted keys and an
+ * incremental sort when there are presorted keys.
+ */
+ if (presorted_keys == 0 || !enable_incremental_sort)
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ else
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+ }
+
+ /* Now decide what to stick atop it */
+ if (parse->groupingSets)
+ {
+ consider_groupingsets_paths(root, grouped_rel,
+ path, true, can_hash,
+ gd, agg_costs, dNumGroups);
+ }
+ else if (parse->hasAggs)
+ {
+ /*
+ * We have aggregation, possibly with plain GROUP BY. Make an
+ * AggPath.
+ */
+ add_path(grouped_rel, (Path *)
create_agg_path(root,
grouped_rel,
path,
grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_SIMPLE,
- root->processed_groupClause,
+ info->clauses,
havingQual,
agg_costs,
dNumGroups));
- }
- else if (parse->groupClause)
- {
- /*
- * We have GROUP BY without aggregation or grouping sets. Make
- * a GroupPath.
- */
- add_path(grouped_rel, (Path *)
+ }
+ else if (parse->groupClause)
+ {
+ /*
+ * We have GROUP BY without aggregation or grouping sets. Make
+ * a GroupPath.
+ */
+ add_path(grouped_rel, (Path *)
create_group_path(root,
grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
havingQual,
dNumGroups));
- }
- else
- {
- /* Other cases should have been handled above */
- Assert(false);
+ }
+ else
+ {
+ /* Other cases should have been handled above */
+ Assert(false);
+ }
}
}
@@ -6938,69 +6958,84 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
{
foreach(lc, partially_grouped_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
+ List *pathkey_orderings = NIL;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
- if (!is_sorted)
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach(lc2, pathkey_orderings)
{
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially
- * sorted already (no need to deal with paths which have
- * presorted keys when incremental sort is disabled unless
- * it's the cheapest input path).
- */
- if (path != partially_grouped_rel->cheapest_total_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ bool is_sorted;
+ int presorted_keys;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- /*
- * We've no need to consider both a sort and incremental
- * sort. We'll just do a sort if there are no pre-sorted
- * keys and an incremental sort when there are presorted
- * keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
+ path->pathkeys,
+ &presorted_keys);
- if (parse->hasAggs)
- add_path(grouped_rel, (Path *)
+ if (!is_sorted)
+ {
+ /*
+ * Try at least sorting the cheapest path and also try
+ * incrementally sorting any path which is partially
+ * sorted already (no need to deal with paths which have
+ * presorted keys when incremental sort is disabled unless
+ * it's the cheapest input path).
+ */
+ if (path != partially_grouped_rel->cheapest_total_path &&
+ (presorted_keys == 0 || !enable_incremental_sort))
+ continue;
+
+ /*
+ * We've no need to consider both a sort and incremental
+ * sort. We'll just do a sort if there are no pre-sorted
+ * keys and an incremental sort when there are presorted
+ * keys.
+ */
+ if (presorted_keys == 0 || !enable_incremental_sort)
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ else
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_path(grouped_rel, (Path *)
create_agg_path(root,
grouped_rel,
path,
grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_FINAL_DESERIAL,
- root->processed_groupClause,
+ info->clauses,
havingQual,
agg_final_costs,
dNumGroups));
- else
- add_path(grouped_rel, (Path *)
+ else
+ add_path(grouped_rel, (Path *)
create_group_path(root,
grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
havingQual,
dNumGroups));
+ }
}
}
}
@@ -7203,66 +7238,85 @@ create_partial_grouping_paths(PlannerInfo *root,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
- if (!is_sorted)
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach(lc2, pathkey_orderings)
{
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially sorted
- * already (no need to deal with paths which have presorted
- * keys when incremental sort is disabled unless it's the
- * cheapest input path).
- */
- if (path != cheapest_total_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ bool is_sorted;
+ int presorted_keys;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- /*
- * We've no need to consider both a sort and incremental sort.
- * We'll just do a sort if there are no presorted keys and an
- * incremental sort when there are presorted keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
- if (parse->hasAggs)
- add_path(partially_grouped_rel, (Path *)
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+ if (!is_sorted)
+ {
+ /*
+ * Try at least sorting the cheapest path and also try
+ * incrementally sorting any path which is partially sorted
+ * already (no need to deal with paths which have presorted
+ * keys when incremental sort is disabled unless it's the
+ * cheapest input path).
+ */
+ if (path != cheapest_total_path &&
+ (presorted_keys == 0 || !enable_incremental_sort))
+ continue;
+
+ /*
+ * We've no need to consider both a sort and incremental sort.
+ * We'll just do a sort if there are no presorted keys and an
+ * incremental sort when there are presorted keys.
+ */
+ if (presorted_keys == 0 || !enable_incremental_sort)
+ path = (Path *) create_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ else
+ path = (Path *) create_incremental_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_path(partially_grouped_rel, (Path *)
create_agg_path(root,
partially_grouped_rel,
path,
partially_grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_INITIAL_SERIAL,
- root->processed_groupClause,
+ info->clauses,
NIL,
agg_partial_costs,
dNumPartialGroups));
- else
- add_path(partially_grouped_rel, (Path *)
- create_group_path(root,
- partially_grouped_rel,
- path,
- root->processed_groupClause,
- NIL,
- dNumPartialGroups));
+ else
+ add_path(partially_grouped_rel, (Path *)
+ create_group_path(root,
+ partially_grouped_rel,
+ path,
+ info->clauses,
+ NIL,
+ dNumPartialGroups));
+ }
}
}
@@ -7271,67 +7325,86 @@ create_partial_grouping_paths(PlannerInfo *root,
/* Similar to above logic, but for partial paths. */
foreach(lc, input_rel->partial_pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
+
+ Assert(list_length(pathkey_orderings) > 0);
- if (!is_sorted)
+ /* process all potentially interesting grouping reorderings */
+ foreach(lc2, pathkey_orderings)
{
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially sorted
- * already (no need to deal with paths which have presorted
- * keys when incremental sort is disabled unless it's the
- * cheapest input path).
- */
- if (path != cheapest_partial_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ bool is_sorted;
+ int presorted_keys;
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
- /*
- * We've no need to consider both a sort and incremental sort.
- * We'll just do a sort if there are no presorted keys and an
- * incremental sort when there are presorted keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
- if (parse->hasAggs)
- add_partial_path(partially_grouped_rel, (Path *)
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (!is_sorted)
+ {
+ /*
+ * Try at least sorting the cheapest path and also try
+ * incrementally sorting any path which is partially sorted
+ * already (no need to deal with paths which have presorted
+ * keys when incremental sort is disabled unless it's the
+ * cheapest input path).
+ */
+ if (path != cheapest_partial_path &&
+ (presorted_keys == 0 || !enable_incremental_sort))
+ continue;
+
+ /*
+ * We've no need to consider both a sort and incremental sort.
+ * We'll just do a sort if there are no presorted keys and an
+ * incremental sort when there are presorted keys.
+ */
+ if (presorted_keys == 0 || !enable_incremental_sort)
+ path = (Path *) create_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ -1.0);
+ else
+ path = (Path *) create_incremental_sort_path(root,
+ partially_grouped_rel,
+ path,
+ info->pathkeys,
+ presorted_keys,
+ -1.0);
+ }
+
+ if (parse->hasAggs)
+ add_partial_path(partially_grouped_rel, (Path *)
create_agg_path(root,
partially_grouped_rel,
path,
partially_grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_INITIAL_SERIAL,
- root->processed_groupClause,
+ info->clauses,
NIL,
agg_partial_costs,
dNumPartialPartialGroups));
- else
- add_partial_path(partially_grouped_rel, (Path *)
+ else
+ add_partial_path(partially_grouped_rel, (Path *)
create_group_path(root,
partially_grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
NIL,
dNumPartialPartialGroups));
+ }
}
}
@@ -7445,6 +7518,8 @@ gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
* We can also skip the entire loop when we only have a single-item
* group_pathkeys because then we can't possibly have a presorted prefix
* of the list without having the list be fully sorted.
+ *
+ * XXX Shouldn't this also consider the group-key-reordering?
*/
if (!enable_incremental_sort || list_length(root->group_pathkeys) == 1)
return;
diff --git a/src/backend/optimizer/util/pathnode.c b/src/backend/optimizer/util/pathnode.c
index 211ba65389..7b8d0cd294 100644
--- a/src/backend/optimizer/util/pathnode.c
+++ b/src/backend/optimizer/util/pathnode.c
@@ -1346,12 +1346,12 @@ create_append_path(PlannerInfo *root,
pathnode->path.total_cost = child->total_cost;
}
else
- cost_append(pathnode);
+ cost_append(pathnode, root);
/* Must do this last, else cost_append complains */
pathnode->path.pathkeys = child->pathkeys;
}
else
- cost_append(pathnode);
+ cost_append(pathnode, root);
/* If the caller provided a row estimate, override the computed value. */
if (rows >= 0)
diff --git a/src/backend/utils/adt/selfuncs.c b/src/backend/utils/adt/selfuncs.c
index c4fcd0076e..cb7e67099e 100644
--- a/src/backend/utils/adt/selfuncs.c
+++ b/src/backend/utils/adt/selfuncs.c
@@ -3386,11 +3386,28 @@ double
estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
List **pgset, EstimationInfo *estinfo)
{
- List *varinfos = NIL;
+ return estimate_num_groups_incremental(root, groupExprs,
+ input_rows, pgset, estinfo,
+ NULL, 0);
+}
+
+/*
+ * estimate_num_groups_incremental
+ * An estimate_num_groups variant, optimized for cases that are adding the
+ * expressions incrementally (e.g. one by one).
+ */
+double
+estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
+ double input_rows,
+ List **pgset, EstimationInfo *estinfo,
+ List **cache_varinfos, int prevNExprs)
+{
+ List *varinfos = (cache_varinfos) ? *cache_varinfos : NIL;
double srf_multiplier = 1.0;
double numdistinct;
ListCell *l;
- int i;
+ int i,
+ j;
/* Zero the estinfo output parameter, if non-NULL */
if (estinfo != NULL)
@@ -3421,7 +3438,7 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
*/
numdistinct = 1.0;
- i = 0;
+ i = j = 0;
foreach(l, groupExprs)
{
Node *groupexpr = (Node *) lfirst(l);
@@ -3430,6 +3447,14 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
List *varshere;
ListCell *l2;
+ /* was done on previous call */
+ if (cache_varinfos && j++ < prevNExprs)
+ {
+ if (pgset)
+ i++; /* to keep in sync with lines below */
+ continue;
+ }
+
/* is expression in this grouping set? */
if (pgset && !list_member_int(*pgset, i++))
continue;
@@ -3499,7 +3524,11 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
if (varshere == NIL)
{
if (contain_volatile_functions(groupexpr))
+ {
+ if (cache_varinfos)
+ *cache_varinfos = varinfos;
return input_rows;
+ }
continue;
}
@@ -3516,6 +3545,9 @@ estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
}
}
+ if (cache_varinfos)
+ *cache_varinfos = varinfos;
+
/*
* If now no Vars, we must have an all-constant or all-boolean GROUP BY
* list.
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index 16ec6c5ef0..9701d6c70f 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -1027,6 +1027,16 @@ struct config_bool ConfigureNamesBool[] =
true,
NULL, NULL, NULL
},
+ {
+ {"enable_group_by_reordering", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("Enables reordering of GROUP BY keys."),
+ NULL,
+ GUC_EXPLAIN
+ },
+ &enable_group_by_reordering,
+ true,
+ NULL, NULL, NULL
+ },
{
{"geqo", PGC_USERSET, QUERY_TUNING_GEQO,
gettext_noop("Enables genetic query optimization."),
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index d08d55c3fe..0fcd023c38 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -391,6 +391,7 @@
#enable_seqscan = on
#enable_sort = on
#enable_tidscan = on
+#enable_group_by_reordering = on
# - Planner Cost Constants -
diff --git a/src/include/nodes/pathnodes.h b/src/include/nodes/pathnodes.h
index 5702fbba60..15ba2f8218 100644
--- a/src/include/nodes/pathnodes.h
+++ b/src/include/nodes/pathnodes.h
@@ -1456,6 +1456,16 @@ typedef struct PathKey
bool pk_nulls_first; /* do NULLs come before normal values? */
} PathKey;
+/*
+ * Combines information about pathkeys and the associated clauses.
+ */
+typedef struct PathKeyInfo
+{
+ NodeTag type;
+ List *pathkeys;
+ List *clauses;
+} PathKeyInfo;
+
/*
* VolatileFunctionStatus -- allows nodes to cache their
* contain_volatile_functions properties. VOLATILITY_UNKNOWN means not yet
diff --git a/src/include/optimizer/cost.h b/src/include/optimizer/cost.h
index bee090ffc2..bbb1132473 100644
--- a/src/include/optimizer/cost.h
+++ b/src/include/optimizer/cost.h
@@ -116,7 +116,9 @@ extern void cost_incremental_sort(Path *path,
Cost input_startup_cost, Cost input_total_cost,
double input_tuples, int width, Cost comparison_cost, int sort_mem,
double limit_tuples);
-extern void cost_append(AppendPath *apath);
+extern Cost cost_sort_estimate(PlannerInfo *root, List *pathkeys,
+ int nPresortedKeys, double tuples);
+extern void cost_append(AppendPath *path, PlannerInfo *root);
extern void cost_merge_append(Path *path, PlannerInfo *root,
List *pathkeys, int n_streams,
Cost input_startup_cost, Cost input_total_cost,
diff --git a/src/include/optimizer/paths.h b/src/include/optimizer/paths.h
index 50bc3b503a..e4ff16d0c4 100644
--- a/src/include/optimizer/paths.h
+++ b/src/include/optimizer/paths.h
@@ -24,6 +24,7 @@ extern PGDLLIMPORT bool enable_geqo;
extern PGDLLIMPORT int geqo_threshold;
extern PGDLLIMPORT int min_parallel_table_scan_size;
extern PGDLLIMPORT int min_parallel_index_scan_size;
+extern PGDLLIMPORT bool enable_group_by_reordering;
/* Hook for plugins to get control in set_rel_pathlist() */
typedef void (*set_rel_pathlist_hook_type) (PlannerInfo *root,
@@ -205,6 +206,12 @@ typedef enum
extern PathKeysComparison compare_pathkeys(List *keys1, List *keys2);
extern bool pathkeys_contained_in(List *keys1, List *keys2);
extern bool pathkeys_count_contained_in(List *keys1, List *keys2, int *n_common);
+extern int group_keys_reorder_by_pathkeys(List *pathkeys,
+ List **group_pathkeys,
+ List **group_clauses,
+ int num_groupby_pathkeys);
+extern List *get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys);
extern Path *get_cheapest_path_for_pathkeys(List *paths, List *pathkeys,
Relids required_outer,
CostSelector cost_criterion,
diff --git a/src/include/utils/selfuncs.h b/src/include/utils/selfuncs.h
index 2f76c473db..2b21c06cc9 100644
--- a/src/include/utils/selfuncs.h
+++ b/src/include/utils/selfuncs.h
@@ -214,6 +214,11 @@ extern double estimate_num_groups(PlannerInfo *root, List *groupExprs,
double input_rows, List **pgset,
EstimationInfo *estinfo);
+extern double estimate_num_groups_incremental(PlannerInfo *root, List *groupExprs,
+ double input_rows, List **pgset,
+ EstimationInfo *estinfo,
+ List **cache_varinfos, int prevNExprs);
+
extern void estimate_hash_bucket_stats(PlannerInfo *root,
Node *hashkey, double nbuckets,
Selectivity *mcv_freq,
diff --git a/src/test/regress/expected/aggregates.out b/src/test/regress/expected/aggregates.out
index d8271da4d1..205be28b3d 100644
--- a/src/test/regress/expected/aggregates.out
+++ b/src/test/regress/expected/aggregates.out
@@ -1234,7 +1234,8 @@ explain (costs off)
select distinct min(f1), max(f1) from minmaxtest;
QUERY PLAN
---------------------------------------------------------------------------------------------
- Unique
+ HashAggregate
+ Group Key: $0, $1
InitPlan 1 (returns $0)
-> Limit
-> Merge Append
@@ -1257,10 +1258,8 @@ explain (costs off)
-> Index Only Scan using minmaxtest2i on minmaxtest2 minmaxtest_8
Index Cond: (f1 IS NOT NULL)
-> Index Only Scan Backward using minmaxtest3i on minmaxtest3 minmaxtest_9
- -> Sort
- Sort Key: ($0), ($1)
- -> Result
-(26 rows)
+ -> Result
+(25 rows)
select distinct min(f1), max(f1) from minmaxtest;
min | max
@@ -2715,6 +2714,241 @@ SELECT balk(hundred) FROM tenk1;
(1 row)
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+SELECT
+ i AS id,
+ i/2 AS p,
+ format('%60s', i%2) AS v,
+ i/4 AS c,
+ i/8 AS d,
+ (random() * (10000/8))::int as e --the same as d but no correlation with p
+ INTO btg
+FROM
+ generate_series(1, 10000) i;
+VACUUM btg;
+ANALYZE btg;
+-- GROUP BY optimization by reorder columns by frequency
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Sort
+ Sort Key: p, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Sort
+ Sort Key: p, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, c, v
+ -> Sort
+ Sort Key: p, c, v
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: v, p, c
+ -> Sort
+ Sort Key: v, p, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: p, d, v, c
+ -> Sort
+ Sort Key: p, d, v, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: v, p, d, c
+ -> Sort
+ Sort Key: v, p, d, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
+ QUERY PLAN
+------------------------------
+ GroupAggregate
+ Group Key: p, v, d, c
+ -> Sort
+ Sort Key: p, v, d, c
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, d, e
+ -> Sort
+ Sort Key: p, d, e
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, e, d
+ -> Sort
+ Sort Key: p, e, d
+ -> Seq Scan on btg
+(5 rows)
+
+CREATE STATISTICS btg_dep ON d, e, p FROM btg;
+ANALYZE btg;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, d, e
+ -> Sort
+ Sort Key: p, d, e
+ -> Seq Scan on btg
+(5 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+ QUERY PLAN
+-----------------------------
+ GroupAggregate
+ Group Key: p, e, d
+ -> Sort
+ Sort Key: p, e, d
+ -> Seq Scan on btg
+(5 rows)
+
+-- GROUP BY optimization by reorder columns by index scan
+CREATE INDEX ON btg(p, v);
+SET enable_seqscan=off;
+SET enable_bitmapscan=off;
+VACUUM btg;
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v ORDER BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p ORDER BY p, v;
+ QUERY PLAN
+------------------------------------------------
+ GroupAggregate
+ Group Key: p, v
+ -> Index Only Scan using btg_p_v_idx on btg
+(3 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, c, v
+ -> Incremental Sort
+ Sort Key: p, c, v
+ Presorted Key: p
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY p, v;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, v, c
+ -> Incremental Sort
+ Sort Key: p, v, c
+ Presorted Key: p, v
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, c, v, d
+ -> Incremental Sort
+ Sort Key: p, c, v, d
+ Presorted Key: p
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d ORDER BY p, v;
+ QUERY PLAN
+-------------------------------------------------
+ GroupAggregate
+ Group Key: p, v, c, d
+ -> Incremental Sort
+ Sort Key: p, v, c, d
+ Presorted Key: p, v
+ -> Index Scan using btg_p_v_idx on btg
+(6 rows)
+
+DROP TABLE btg;
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+RESET enable_seqscan;
+RESET enable_bitmapscan;
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/expected/incremental_sort.out b/src/test/regress/expected/incremental_sort.out
index 7fdb685313..b8a0968b06 100644
--- a/src/test/regress/expected/incremental_sort.out
+++ b/src/test/regress/expected/incremental_sort.out
@@ -1426,7 +1426,7 @@ set parallel_setup_cost = 0;
set parallel_tuple_cost = 0;
set max_parallel_workers_per_gather = 2;
create table t (a int, b int, c int);
-insert into t select mod(i,10),mod(i,10),i from generate_series(1,10000) s(i);
+insert into t select mod(i,10),mod(i,10),i from generate_series(1,60000) s(i);
create index on t (a);
analyze t;
set enable_incremental_sort = off;
diff --git a/src/test/regress/expected/join.out b/src/test/regress/expected/join.out
index 9b8638f286..27684bfce2 100644
--- a/src/test/regress/expected/join.out
+++ b/src/test/regress/expected/join.out
@@ -1984,8 +1984,8 @@ USING (name);
------+----+----
bb | 12 | 13
cc | 22 | 23
- dd | | 33
ee | 42 |
+ dd | | 33
(4 rows)
-- Cases with non-nullable expressions in subquery results;
@@ -2019,8 +2019,8 @@ NATURAL FULL JOIN
------+------+------+------+------
bb | 12 | 2 | 13 | 3
cc | 22 | 2 | 23 | 3
- dd | | | 33 | 3
ee | 42 | 2 | |
+ dd | | | 33 | 3
(4 rows)
SELECT * FROM
@@ -5645,18 +5645,20 @@ select d.* from d left join (select * from b group by b.id, b.c_id) s
explain (costs off)
select d.* from d left join (select distinct * from b) s
on d.a = s.id;
- QUERY PLAN
---------------------------------------
- Merge Right Join
- Merge Cond: (b.id = d.a)
- -> Unique
- -> Sort
- Sort Key: b.id, b.c_id
- -> Seq Scan on b
+ QUERY PLAN
+---------------------------------------------
+ Merge Left Join
+ Merge Cond: (d.a = s.id)
-> Sort
Sort Key: d.a
-> Seq Scan on d
-(9 rows)
+ -> Sort
+ Sort Key: s.id
+ -> Subquery Scan on s
+ -> HashAggregate
+ Group Key: b.id, b.c_id
+ -> Seq Scan on b
+(11 rows)
-- join removal is not possible here
explain (costs off)
@@ -7597,44 +7599,39 @@ select * from j1 natural join j2;
explain (verbose, costs off)
select * from j1
inner join (select distinct id from j3) j3 on j1.id = j3.id;
- QUERY PLAN
------------------------------------------
+ QUERY PLAN
+-----------------------------------
Nested Loop
Output: j1.id, j3.id
Inner Unique: true
Join Filter: (j1.id = j3.id)
- -> Unique
+ -> HashAggregate
Output: j3.id
- -> Sort
+ Group Key: j3.id
+ -> Seq Scan on public.j3
Output: j3.id
- Sort Key: j3.id
- -> Seq Scan on public.j3
- Output: j3.id
-> Seq Scan on public.j1
Output: j1.id
-(13 rows)
+(11 rows)
-- ensure group by clause allows the inner to become unique
explain (verbose, costs off)
select * from j1
inner join (select id from j3 group by id) j3 on j1.id = j3.id;
- QUERY PLAN
------------------------------------------
+ QUERY PLAN
+-----------------------------------
Nested Loop
Output: j1.id, j3.id
Inner Unique: true
Join Filter: (j1.id = j3.id)
- -> Group
+ -> HashAggregate
Output: j3.id
Group Key: j3.id
- -> Sort
+ -> Seq Scan on public.j3
Output: j3.id
- Sort Key: j3.id
- -> Seq Scan on public.j3
- Output: j3.id
-> Seq Scan on public.j1
Output: j1.id
-(14 rows)
+(11 rows)
drop table j1;
drop table j2;
diff --git a/src/test/regress/expected/merge.out b/src/test/regress/expected/merge.out
index 28a6d0ba98..51a4596e9f 100644
--- a/src/test/regress/expected/merge.out
+++ b/src/test/regress/expected/merge.out
@@ -1459,18 +1459,15 @@ WHEN MATCHED AND t.a < 10 THEN
explain_merge
--------------------------------------------------------------------
Merge on ex_mtarget t (actual rows=0 loops=1)
- -> Merge Join (actual rows=0 loops=1)
- Merge Cond: (t.a = s.a)
- -> Sort (actual rows=0 loops=1)
- Sort Key: t.a
- Sort Method: quicksort Memory: xxx
+ -> Hash Join (actual rows=0 loops=1)
+ Hash Cond: (s.a = t.a)
+ -> Seq Scan on ex_msource s (actual rows=1 loops=1)
+ -> Hash (actual rows=0 loops=1)
+ Buckets: xxx Batches: xxx Memory Usage: xxx
-> Seq Scan on ex_mtarget t (actual rows=0 loops=1)
Filter: (a < '-1000'::integer)
Rows Removed by Filter: 54
- -> Sort (never executed)
- Sort Key: s.a
- -> Seq Scan on ex_msource s (never executed)
-(12 rows)
+(9 rows)
DROP TABLE ex_msource, ex_mtarget;
DROP FUNCTION explain_merge(text);
diff --git a/src/test/regress/expected/partition_aggregate.out b/src/test/regress/expected/partition_aggregate.out
index 1b900fddf8..800114dfde 100644
--- a/src/test/regress/expected/partition_aggregate.out
+++ b/src/test/regress/expected/partition_aggregate.out
@@ -948,12 +948,12 @@ SET parallel_setup_cost = 0;
-- is not partial agg safe.
EXPLAIN (COSTS OFF)
SELECT a, sum(b), array_agg(distinct c), count(*) FROM pagg_tab_ml GROUP BY a HAVING avg(b) < 3 ORDER BY 1, 2, 3;
- QUERY PLAN
---------------------------------------------------------------------------------------
- Sort
- Sort Key: pagg_tab_ml.a, (sum(pagg_tab_ml.b)), (array_agg(DISTINCT pagg_tab_ml.c))
- -> Gather
- Workers Planned: 2
+ QUERY PLAN
+--------------------------------------------------------------------------------------------
+ Gather Merge
+ Workers Planned: 2
+ -> Sort
+ Sort Key: pagg_tab_ml.a, (sum(pagg_tab_ml.b)), (array_agg(DISTINCT pagg_tab_ml.c))
-> Parallel Append
-> GroupAggregate
Group Key: pagg_tab_ml.a
@@ -1380,28 +1380,26 @@ SELECT x, sum(y), avg(y), count(*) FROM pagg_tab_para GROUP BY x HAVING avg(y) <
-- When GROUP BY clause does not match; partial aggregation is performed for each partition.
EXPLAIN (COSTS OFF)
SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) < 12 ORDER BY 1, 2, 3;
- QUERY PLAN
--------------------------------------------------------------------------------------------
+ QUERY PLAN
+-------------------------------------------------------------------------------------
Sort
Sort Key: pagg_tab_para.y, (sum(pagg_tab_para.x)), (avg(pagg_tab_para.x))
- -> Finalize GroupAggregate
+ -> Finalize HashAggregate
Group Key: pagg_tab_para.y
Filter: (avg(pagg_tab_para.x) < '12'::numeric)
- -> Gather Merge
+ -> Gather
Workers Planned: 2
- -> Sort
- Sort Key: pagg_tab_para.y
- -> Parallel Append
- -> Partial HashAggregate
- Group Key: pagg_tab_para.y
- -> Parallel Seq Scan on pagg_tab_para_p1 pagg_tab_para
- -> Partial HashAggregate
- Group Key: pagg_tab_para_1.y
- -> Parallel Seq Scan on pagg_tab_para_p2 pagg_tab_para_1
- -> Partial HashAggregate
- Group Key: pagg_tab_para_2.y
- -> Parallel Seq Scan on pagg_tab_para_p3 pagg_tab_para_2
-(19 rows)
+ -> Parallel Append
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para.y
+ -> Parallel Seq Scan on pagg_tab_para_p1 pagg_tab_para
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para_1.y
+ -> Parallel Seq Scan on pagg_tab_para_p2 pagg_tab_para_1
+ -> Partial HashAggregate
+ Group Key: pagg_tab_para_2.y
+ -> Parallel Seq Scan on pagg_tab_para_p3 pagg_tab_para_2
+(17 rows)
SELECT y, sum(x), avg(x), count(*) FROM pagg_tab_para GROUP BY y HAVING avg(x) < 12 ORDER BY 1, 2, 3;
y | sum | avg | count
diff --git a/src/test/regress/expected/partition_join.out b/src/test/regress/expected/partition_join.out
index 6560fe2416..680e82cfc5 100644
--- a/src/test/regress/expected/partition_join.out
+++ b/src/test/regress/expected/partition_join.out
@@ -512,52 +512,41 @@ EXPLAIN (COSTS OFF)
SELECT a, b FROM prt1 FULL JOIN prt2 p2(b,a,c) USING(a,b)
WHERE a BETWEEN 490 AND 510
GROUP BY 1, 2 ORDER BY 1, 2;
- QUERY PLAN
------------------------------------------------------------------------------------------------------------------
+ QUERY PLAN
+-----------------------------------------------------------------------------------------------------------
Group
Group Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Merge Append
+ -> Sort
Sort Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Group
- Group Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Sort
- Sort Key: (COALESCE(prt1.a, p2.a)), (COALESCE(prt1.b, p2.b))
- -> Merge Full Join
- Merge Cond: ((prt1.a = p2.a) AND (prt1.b = p2.b))
- Filter: ((COALESCE(prt1.a, p2.a) >= 490) AND (COALESCE(prt1.a, p2.a) <= 510))
- -> Sort
- Sort Key: prt1.a, prt1.b
- -> Seq Scan on prt1_p1 prt1
- -> Sort
- Sort Key: p2.a, p2.b
- -> Seq Scan on prt2_p1 p2
- -> Group
- Group Key: (COALESCE(prt1_1.a, p2_1.a)), (COALESCE(prt1_1.b, p2_1.b))
- -> Sort
- Sort Key: (COALESCE(prt1_1.a, p2_1.a)), (COALESCE(prt1_1.b, p2_1.b))
- -> Merge Full Join
- Merge Cond: ((prt1_1.a = p2_1.a) AND (prt1_1.b = p2_1.b))
- Filter: ((COALESCE(prt1_1.a, p2_1.a) >= 490) AND (COALESCE(prt1_1.a, p2_1.a) <= 510))
- -> Sort
- Sort Key: prt1_1.a, prt1_1.b
- -> Seq Scan on prt1_p2 prt1_1
- -> Sort
- Sort Key: p2_1.a, p2_1.b
- -> Seq Scan on prt2_p2 p2_1
- -> Group
- Group Key: (COALESCE(prt1_2.a, p2_2.a)), (COALESCE(prt1_2.b, p2_2.b))
- -> Sort
- Sort Key: (COALESCE(prt1_2.a, p2_2.a)), (COALESCE(prt1_2.b, p2_2.b))
- -> Merge Full Join
- Merge Cond: ((prt1_2.a = p2_2.a) AND (prt1_2.b = p2_2.b))
- Filter: ((COALESCE(prt1_2.a, p2_2.a) >= 490) AND (COALESCE(prt1_2.a, p2_2.a) <= 510))
- -> Sort
- Sort Key: prt1_2.a, prt1_2.b
- -> Seq Scan on prt1_p3 prt1_2
- -> Sort
- Sort Key: p2_2.a, p2_2.b
- -> Seq Scan on prt2_p3 p2_2
-(43 rows)
+ -> Append
+ -> Merge Full Join
+ Merge Cond: ((prt1_1.a = p2_1.a) AND (prt1_1.b = p2_1.b))
+ Filter: ((COALESCE(prt1_1.a, p2_1.a) >= 490) AND (COALESCE(prt1_1.a, p2_1.a) <= 510))
+ -> Sort
+ Sort Key: prt1_1.a, prt1_1.b
+ -> Seq Scan on prt1_p1 prt1_1
+ -> Sort
+ Sort Key: p2_1.a, p2_1.b
+ -> Seq Scan on prt2_p1 p2_1
+ -> Merge Full Join
+ Merge Cond: ((prt1_2.a = p2_2.a) AND (prt1_2.b = p2_2.b))
+ Filter: ((COALESCE(prt1_2.a, p2_2.a) >= 490) AND (COALESCE(prt1_2.a, p2_2.a) <= 510))
+ -> Sort
+ Sort Key: prt1_2.a, prt1_2.b
+ -> Seq Scan on prt1_p2 prt1_2
+ -> Sort
+ Sort Key: p2_2.a, p2_2.b
+ -> Seq Scan on prt2_p2 p2_2
+ -> Merge Full Join
+ Merge Cond: ((prt1_3.a = p2_3.a) AND (prt1_3.b = p2_3.b))
+ Filter: ((COALESCE(prt1_3.a, p2_3.a) >= 490) AND (COALESCE(prt1_3.a, p2_3.a) <= 510))
+ -> Sort
+ Sort Key: prt1_3.a, prt1_3.b
+ -> Seq Scan on prt1_p3 prt1_3
+ -> Sort
+ Sort Key: p2_3.a, p2_3.b
+ -> Seq Scan on prt2_p3 p2_3
+(32 rows)
SELECT a, b FROM prt1 FULL JOIN prt2 p2(b,a,c) USING(a,b)
WHERE a BETWEEN 490 AND 510
diff --git a/src/test/regress/expected/sysviews.out b/src/test/regress/expected/sysviews.out
index aae5d51e1c..487fe26906 100644
--- a/src/test/regress/expected/sysviews.out
+++ b/src/test/regress/expected/sysviews.out
@@ -114,6 +114,7 @@ select name, setting from pg_settings where name like 'enable%';
enable_async_append | on
enable_bitmapscan | on
enable_gathermerge | on
+ enable_group_by_reordering | on
enable_hashagg | on
enable_hashjoin | on
enable_incremental_sort | on
@@ -132,7 +133,7 @@ select name, setting from pg_settings where name like 'enable%';
enable_seqscan | on
enable_sort | on
enable_tidscan | on
-(21 rows)
+(22 rows)
-- There are always wait event descriptions for various types.
select type, count(*) > 0 as ok FROM pg_wait_events
diff --git a/src/test/regress/expected/union.out b/src/test/regress/expected/union.out
index e2613d6777..220f0c1629 100644
--- a/src/test/regress/expected/union.out
+++ b/src/test/regress/expected/union.out
@@ -1303,24 +1303,22 @@ select distinct q1 from
union all
select distinct * from int8_tbl i82) ss
where q2 = q2;
- QUERY PLAN
-----------------------------------------------------------
- Unique
- -> Merge Append
- Sort Key: "*SELECT* 1".q1
+ QUERY PLAN
+----------------------------------------------------
+ HashAggregate
+ Group Key: "*SELECT* 1".q1
+ -> Append
-> Subquery Scan on "*SELECT* 1"
- -> Unique
- -> Sort
- Sort Key: i81.q1, i81.q2
- -> Seq Scan on int8_tbl i81
- Filter: (q2 IS NOT NULL)
+ -> HashAggregate
+ Group Key: i81.q1, i81.q2
+ -> Seq Scan on int8_tbl i81
+ Filter: (q2 IS NOT NULL)
-> Subquery Scan on "*SELECT* 2"
- -> Unique
- -> Sort
- Sort Key: i82.q1, i82.q2
- -> Seq Scan on int8_tbl i82
- Filter: (q2 IS NOT NULL)
-(15 rows)
+ -> HashAggregate
+ Group Key: i82.q1, i82.q2
+ -> Seq Scan on int8_tbl i82
+ Filter: (q2 IS NOT NULL)
+(13 rows)
select distinct q1 from
(select distinct * from int8_tbl i81
@@ -1339,24 +1337,22 @@ select distinct q1 from
union all
select distinct * from int8_tbl i82) ss
where -q1 = q2;
- QUERY PLAN
---------------------------------------------------------
- Unique
- -> Merge Append
- Sort Key: "*SELECT* 1".q1
+ QUERY PLAN
+--------------------------------------------------
+ HashAggregate
+ Group Key: "*SELECT* 1".q1
+ -> Append
-> Subquery Scan on "*SELECT* 1"
- -> Unique
- -> Sort
- Sort Key: i81.q1, i81.q2
- -> Seq Scan on int8_tbl i81
- Filter: ((- q1) = q2)
+ -> HashAggregate
+ Group Key: i81.q1, i81.q2
+ -> Seq Scan on int8_tbl i81
+ Filter: ((- q1) = q2)
-> Subquery Scan on "*SELECT* 2"
- -> Unique
- -> Sort
- Sort Key: i82.q1, i82.q2
- -> Seq Scan on int8_tbl i82
- Filter: ((- q1) = q2)
-(15 rows)
+ -> HashAggregate
+ Group Key: i82.q1, i82.q2
+ -> Seq Scan on int8_tbl i82
+ Filter: ((- q1) = q2)
+(13 rows)
select distinct q1 from
(select distinct * from int8_tbl i81
diff --git a/src/test/regress/sql/aggregates.sql b/src/test/regress/sql/aggregates.sql
index 75c78be640..5c4e768825 100644
--- a/src/test/regress/sql/aggregates.sql
+++ b/src/test/regress/sql/aggregates.sql
@@ -1172,6 +1172,105 @@ SELECT balk(hundred) FROM tenk1;
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+
+SELECT
+ i AS id,
+ i/2 AS p,
+ format('%60s', i%2) AS v,
+ i/4 AS c,
+ i/8 AS d,
+ (random() * (10000/8))::int as e --the same as d but no correlation with p
+ INTO btg
+FROM
+ generate_series(1, 10000) i;
+
+VACUUM btg;
+ANALYZE btg;
+
+-- GROUP BY optimization by reorder columns by frequency
+
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY v, p, d ,c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, d, c ORDER BY p, v, d ,c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+
+CREATE STATISTICS btg_dep ON d, e, p FROM btg;
+ANALYZE btg;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, d, e;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, e, d;
+
+
+-- GROUP BY optimization by reorder columns by index scan
+
+CREATE INDEX ON btg(p, v);
+SET enable_seqscan=off;
+SET enable_bitmapscan=off;
+VACUUM btg;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY p, v ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, p, c ORDER BY p, v;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d;
+
+EXPLAIN (COSTS off)
+SELECT count(*) FROM btg GROUP BY v, c, p, d ORDER BY p, v;
+
+DROP TABLE btg;
+
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+RESET enable_seqscan;
+RESET enable_bitmapscan;
+
+
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/sql/incremental_sort.sql b/src/test/regress/sql/incremental_sort.sql
index ab471bdfff..70f311a6f2 100644
--- a/src/test/regress/sql/incremental_sort.sql
+++ b/src/test/regress/sql/incremental_sort.sql
@@ -208,7 +208,7 @@ set parallel_tuple_cost = 0;
set max_parallel_workers_per_gather = 2;
create table t (a int, b int, c int);
-insert into t select mod(i,10),mod(i,10),i from generate_series(1,10000) s(i);
+insert into t select mod(i,10),mod(i,10),i from generate_series(1,60000) s(i);
create index on t (a);
analyze t;
--
2.42.0
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2023-12-21 10:53 Alexander Korotkov <[email protected]>
parent: Andrei Lepikhov <[email protected]>
0 siblings, 2 replies; 38+ messages in thread
From: Alexander Korotkov @ 2023-12-21 10:53 UTC (permalink / raw)
To: Andrei Lepikhov <[email protected]>; +Cc: PostgreSQL Developers <[email protected]>; Tomas Vondra <[email protected]>; Teodor Sigaev <[email protected]>; David Rowley <[email protected]>; a.rybakina <[email protected]>; Белялов Дамир Наилевич <[email protected]>; Tom Lane <[email protected]>
Hi!
On Sun, Oct 1, 2023 at 11:45 AM Andrei Lepikhov
<[email protected]> wrote:
>
> New version of the patch. Fixed minor inconsistencies and rebased onto
> current master.
Thank you (and other authors) for working on this subject. Indeed to
GROUP BY clauses are order-agnostic. Reordering them in the most
suitable order could give up significant query planning benefits. I
went through the thread: I see significant work has been already made
on this patch, the code is quite polished.
I'd like to make some notes.
1) As already mentioned, there is clearly a repetitive pattern for the
code following after get_useful_group_keys_orderings() calls. I think
it would be good to extract it into a separate function. Please, do
this as a separate patch coming before the group-by patch. That would
simplify the review.
2) I wonder what planning overhead this patch could introduce? Could
you try to measure the worst case? What if we have a table with a lot
of indexes and a long list of group-by clauses partially patching
every index. This should give us an understanding on whether we need
a separate GUC to control this feature.
3) I see that get_useful_group_keys_orderings() makes 3 calls to
get_cheapest_group_keys_order() function. Each time
get_cheapest_group_keys_order() performs the cost estimate and
reorders the free keys. However, cost estimation implies the system
catalog lookups (that is quite expensive). I wonder if we could
change the algorithm. Could we just sort the group-by keys by cost
once, save this ordering and then just re-use it. So, every time we
need to reorder a group by, we can just pull the required keys to the
top and use saved ordering for the rest. I also wonder if we could do
this once for add_paths_to_grouping_rel() and
create_partial_grouping_paths() calls. So, it probably should be
somewhere in create_ordinary_grouping_paths().
4) I think we can do some optimizations when enable_incremental_sort
== off. Then in get_useful_group_keys_orderings() we should only deal
with input_path fully matching the group-by clause, and try only full
match of group-by output to the required order.
------
Regards,
Alexander Korotkov
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2023-12-26 11:37 Andrei Lepikhov <[email protected]>
parent: Alexander Korotkov <[email protected]>
1 sibling, 0 replies; 38+ messages in thread
From: Andrei Lepikhov @ 2023-12-26 11:37 UTC (permalink / raw)
To: Alexander Korotkov <[email protected]>; +Cc: PostgreSQL Developers <[email protected]>; Tomas Vondra <[email protected]>; Teodor Sigaev <[email protected]>; David Rowley <[email protected]>; a.rybakina <[email protected]>; Белялов Дамир Наилевич <[email protected]>; Tom Lane <[email protected]>
On 21/12/2023 17:53, Alexander Korotkov wrote:
> On Sun, Oct 1, 2023 at 11:45 AM Andrei Lepikhov
> <[email protected]> wrote:
>> New version of the patch. Fixed minor inconsistencies and rebased onto
>> current master.
> Thank you (and other authors) for working on this subject. Indeed to
> GROUP BY clauses are order-agnostic. Reordering them in the most
> suitable order could give up significant query planning benefits. I
> went through the thread: I see significant work has been already made
> on this patch, the code is quite polished.
Maybe, but issues, mentioned in [1], still not resolved. It is the only
reason, why this thread hasn't been active.
> I'd like to make some notes.
> 1) As already mentioned, there is clearly a repetitive pattern for the
> code following after get_useful_group_keys_orderings() calls. I think
> it would be good to extract it into a separate function. Please, do
> this as a separate patch coming before the group-by patch. That would
> simplify the review.
Yeah, these parts of code a bit different. I will try to make common
routine.
> 2) I wonder what planning overhead this patch could introduce? Could
> you try to measure the worst case? What if we have a table with a lot
> of indexes and a long list of group-by clauses partially patching
> every index. This should give us an understanding on whether we need
> a separate GUC to control this feature.
Ok> 3) I see that get_useful_group_keys_orderings() makes 3 calls to
> get_cheapest_group_keys_order() function. Each time
> get_cheapest_group_keys_order() performs the cost estimate and
> reorders the free keys. However, cost estimation implies the system
> catalog lookups (that is quite expensive). I wonder if we could
> change the algorithm. Could we just sort the group-by keys by cost
> once, save this ordering and then just re-use it. So, every time we
> need to reorder a group by, we can just pull the required keys to the
> top and use saved ordering for the rest. I also wonder if we could do
> this once for add_paths_to_grouping_rel() and
> create_partial_grouping_paths() calls. So, it probably should be
> somewhere in create_ordinary_grouping_paths().
Thanks for the idea!> 4) I think we can do some optimizations when
enable_incremental_sort
> == off. Then in get_useful_group_keys_orderings() we should only deal
> with input_path fully matching the group-by clause, and try only full
> match of group-by output to the required order.
Oh, we had designed before the incremental sort was invented. Will see
what we can do here.
[1]
https://www.postgresql.org/message-id/60610df1-c32f-ebdf-e58c-7a664431f452%40enterprisedb.com
--
regards,
Andrei Lepikhov
Postgres Professional
^ permalink raw reply [nested|flat] 38+ messages in thread
* Re: POC: GROUP BY optimization
@ 2024-01-09 09:01 Andrei Lepikhov <[email protected]>
parent: Alexander Korotkov <[email protected]>
1 sibling, 0 replies; 38+ messages in thread
From: Andrei Lepikhov @ 2024-01-09 09:01 UTC (permalink / raw)
To: Alexander Korotkov <[email protected]>; +Cc: PostgreSQL Developers <[email protected]>; Tomas Vondra <[email protected]>; Teodor Sigaev <[email protected]>; David Rowley <[email protected]>; a.rybakina <[email protected]>; Tom Lane <[email protected]>
Here is a new version of GROUP-BY optimization without sort model.
On 21/12/2023 17:53, Alexander Korotkov wrote:
> I'd like to make some notes.
>
> 1) As already mentioned, there is clearly a repetitive pattern for the
> code following after get_useful_group_keys_orderings() calls. I think
> it would be good to extract it into a separate function. Please, do
> this as a separate patch coming before the group-by patch. That would
> simplify the review.
Done. See patch 0001-*. Unfortunately, extraction of whole cycle isn't
practical, because it blows out the interface of the routine.
> 2) I wonder what planning overhead this patch could introduce? Could
> you try to measure the worst case? What if we have a table with a lot
> of indexes and a long list of group-by clauses partially patching
> every index. This should give us an understanding on whether we need
> a separate GUC to control this feature.
In current implementation I don't anticipate any significant overhead.
GUC is needed here to allow users adhere their own ordering and to
disable feature in the case of problems.
> 4) I think we can do some optimizations when enable_incremental_sort
> == off. Then in get_useful_group_keys_orderings() we should only deal
> with input_path fully matching the group-by clause, and try only full
> match of group-by output to the required order.
Hm, is it really make sense in current implementation?
--
regards,
Andrei Lepikhov
Postgres Professional
From ea068e221a754a463142575f6e0360d3118475f8 Mon Sep 17 00:00:00 2001
From: "Andrey V. Lepikhov" <[email protected]>
Date: Tue, 9 Jan 2024 12:00:27 +0700
Subject: [PATCH 1/2] Generalize common code of adding sort before generation
of grouping paths.
---
src/backend/optimizer/plan/planner.c | 209 ++++++++-------------------
1 file changed, 61 insertions(+), 148 deletions(-)
diff --git a/src/backend/optimizer/plan/planner.c b/src/backend/optimizer/plan/planner.c
index 667723b675..fcf647940e 100644
--- a/src/backend/optimizer/plan/planner.c
+++ b/src/backend/optimizer/plan/planner.c
@@ -6809,6 +6809,52 @@ done:
return parallel_workers;
}
+static Path *
+try_presorting(PlannerInfo *root, RelOptInfo *grouped_rel,
+ Path *path, Path *cheapest_path)
+{
+ bool is_sorted;
+ int presorted_keys;
+
+ is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (!is_sorted)
+ {
+ /*
+ * Try at least sorting the cheapest path and also try
+ * incrementally sorting any path which is partially sorted
+ * already (no need to deal with paths which have presorted
+ * keys when incremental sort is disabled unless it's the
+ * cheapest input path).
+ */
+ if (path != cheapest_path &&
+ (presorted_keys == 0 || !enable_incremental_sort))
+ return NULL;
+
+ /*
+ * We've no need to consider both a sort and incremental sort.
+ * We'll just do a sort if there are no presorted keys and an
+ * incremental sort when there are presorted keys.
+ */
+ if (presorted_keys == 0 || !enable_incremental_sort)
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ root->group_pathkeys,
+ -1.0);
+ else
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ root->group_pathkeys,
+ presorted_keys,
+ -1.0);
+ }
+ return path;
+}
+
/*
* add_paths_to_grouping_rel
*
@@ -6840,45 +6886,11 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
foreach(lc, input_rel->pathlist)
{
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
-
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
- if (!is_sorted)
- {
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially sorted
- * already (no need to deal with paths which have presorted
- * keys when incremental sort is disabled unless it's the
- * cheapest input path).
- */
- if (path != cheapest_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ path = try_presorting(root, grouped_rel, path, cheapest_path);
- /*
- * We've no need to consider both a sort and incremental sort.
- * We'll just do a sort if there are no presorted keys and an
- * incremental sort when there are presorted keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ if (path == NULL)
+ continue;
/* Now decide what to stick atop it */
if (parse->groupingSets)
@@ -6935,46 +6947,12 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
foreach(lc, partially_grouped_rel->pathlist)
{
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
-
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
- if (!is_sorted)
- {
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially
- * sorted already (no need to deal with paths which have
- * presorted keys when incremental sort is disabled unless
- * it's the cheapest input path).
- */
- if (path != partially_grouped_rel->cheapest_total_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ path = try_presorting(root, grouped_rel, path,
+ partially_grouped_rel->cheapest_total_path);
- /*
- * We've no need to consider both a sort and incremental
- * sort. We'll just do a sort if there are no pre-sorted
- * keys and an incremental sort when there are presorted
- * keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ if (path == NULL)
+ continue;
if (parse->hasAggs)
add_path(grouped_rel, (Path *)
@@ -7200,44 +7178,12 @@ create_partial_grouping_paths(PlannerInfo *root,
foreach(lc, input_rel->pathlist)
{
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
- if (!is_sorted)
- {
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially sorted
- * already (no need to deal with paths which have presorted
- * keys when incremental sort is disabled unless it's the
- * cheapest input path).
- */
- if (path != cheapest_total_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ path = try_presorting(root, partially_grouped_rel, path,
+ cheapest_total_path);
- /*
- * We've no need to consider both a sort and incremental sort.
- * We'll just do a sort if there are no presorted keys and an
- * incremental sort when there are presorted keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ if (path == NULL)
+ continue;
if (parse->hasAggs)
add_path(partially_grouped_rel, (Path *)
@@ -7268,45 +7214,12 @@ create_partial_grouping_paths(PlannerInfo *root,
foreach(lc, input_rel->partial_pathlist)
{
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
-
- if (!is_sorted)
- {
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially sorted
- * already (no need to deal with paths which have presorted
- * keys when incremental sort is disabled unless it's the
- * cheapest input path).
- */
- if (path != cheapest_partial_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ path = try_presorting(root, partially_grouped_rel, path,
+ cheapest_partial_path);
- /*
- * We've no need to consider both a sort and incremental sort.
- * We'll just do a sort if there are no presorted keys and an
- * incremental sort when there are presorted keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ if (path == NULL)
+ continue;
if (parse->hasAggs)
add_partial_path(partially_grouped_rel, (Path *)
--
2.43.0
From 886227bd33a368df1b564121506842be39d31f3d Mon Sep 17 00:00:00 2001
From: "Andrey V. Lepikhov" <[email protected]>
Date: Tue, 9 Jan 2024 12:32:15 +0700
Subject: [PATCH 2/2] Explore alternative orderings of group-by pathkeys during
optimization.
When evaluating a query with a multi-column GROUP BY clause, we can minimize
sort operations or avoid them if we synchronize the order of GROUP BY clauses
with the ORDER BY sort clause or sort order, which comes from the underlying
query tree. Grouping does not imply any ordering, so we can compare
the keys in arbitrary order, and a Hash Agg leverages this. But for Group Agg,
we simply compared keys in the order specified in the query. This commit
explores alternative ordering of the keys, trying to find a cheaper one.
The ordering of group keys may interact with other parts of the query, some of
which may not be known while planning the grouping. For example, there may be
an explicit ORDER BY clause or some other ordering-dependent operation higher up
in the query, and using the same ordering may allow using either incremental
sort or even eliminating the sort entirely.
The patch always keeps the ordering specified in the query, assuming the user
might have additional insights.
This introduces a new GUC enable_group_by_reordering so that the optimization
may be disabled if needed.
---
src/backend/optimizer/path/equivclass.c | 13 +-
src/backend/optimizer/path/pathkeys.c | 222 ++++++++++++++++++
src/backend/optimizer/plan/planner.c | 214 +++++++++++------
src/backend/utils/misc/guc_tables.c | 10 +
src/backend/utils/misc/postgresql.conf.sample | 1 +
src/include/nodes/pathnodes.h | 10 +
src/include/optimizer/paths.h | 3 +
src/test/regress/expected/aggregates.out | 132 +++++++++++
src/test/regress/expected/sysviews.out | 3 +-
src/test/regress/sql/aggregates.sql | 47 ++++
10 files changed, 586 insertions(+), 69 deletions(-)
diff --git a/src/backend/optimizer/path/equivclass.c b/src/backend/optimizer/path/equivclass.c
index e86dfeaecd..7dd14d0a43 100644
--- a/src/backend/optimizer/path/equivclass.c
+++ b/src/backend/optimizer/path/equivclass.c
@@ -652,7 +652,18 @@ get_eclass_for_sort_expr(PlannerInfo *root,
if (opcintype == cur_em->em_datatype &&
equal(expr, cur_em->em_expr))
- return cur_ec; /* Match! */
+ {
+ /*
+ * Match!
+ *
+ * Copy the sortref if it wasn't set yet. That may happen if
+ * the ec was constructed from WHERE clause, i.e. it doesn't
+ * have a target reference at all.
+ */
+ if (cur_ec->ec_sortref == 0 && sortref > 0)
+ cur_ec->ec_sortref = sortref;
+ return cur_ec;
+ }
}
}
diff --git a/src/backend/optimizer/path/pathkeys.c b/src/backend/optimizer/path/pathkeys.c
index ca94a31f71..8bc4eb9f03 100644
--- a/src/backend/optimizer/path/pathkeys.c
+++ b/src/backend/optimizer/path/pathkeys.c
@@ -28,6 +28,8 @@
#include "partitioning/partbounds.h"
#include "utils/lsyscache.h"
+/* Consider reordering of GROUP BY keys? */
+bool enable_group_by_reordering = true;
static bool pathkey_is_redundant(PathKey *new_pathkey, List *pathkeys);
static bool matches_boolean_partition_clause(RestrictInfo *rinfo,
@@ -350,6 +352,173 @@ pathkeys_contained_in(List *keys1, List *keys2)
return false;
}
+/*
+ * group_keys_reorder_by_pathkeys
+ * Reorder GROUP BY keys to match pathkeys of input path.
+ *
+ * Function returns new lists (pathkeys and clauses), original GROUP BY lists
+ * stay untouched.
+ *
+ * Returns the number of GROUP BY keys with a matching pathkey.
+ */
+static int
+group_keys_reorder_by_pathkeys(List *pathkeys, List **group_pathkeys,
+ List **group_clauses,
+ int num_groupby_pathkeys)
+{
+ List *new_group_pathkeys = NIL,
+ *new_group_clauses = NIL;
+ ListCell *lc;
+ int n;
+
+ if (pathkeys == NIL || *group_pathkeys == NIL)
+ return 0;
+
+ /*
+ * Walk the pathkeys (determining ordering of the input path) and see if
+ * there's a matching GROUP BY key. If we find one, we append it to the
+ * list, and do the same for the clauses.
+ *
+ * Once we find the first pathkey without a matching GROUP BY key, the
+ * rest of the pathkeys are useless and can't be used to evaluate the
+ * grouping, so we abort the loop and ignore the remaining pathkeys.
+ *
+ * XXX Pathkeys are built in a way to allow simply comparing pointers.
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(lc);
+ SortGroupClause *sgc;
+
+ if (foreach_current_index(lc) >= num_groupby_pathkeys ||
+ !list_member_ptr(*group_pathkeys, pathkey))
+ /* */
+ break;
+
+ /* abort on first mismatch */
+ sgc = get_sortgroupref_clause_noerr(pathkey->pk_eclass->ec_sortref,
+ *group_clauses);
+ if (!sgc)
+ /* The grouping clause is not cover this pathkey */
+ break;
+
+ new_group_pathkeys = lappend(new_group_pathkeys, pathkey);
+ new_group_clauses = lappend(new_group_clauses, sgc);
+ }
+
+ /* remember the number of pathkeys with a matching GROUP BY key */
+ n = list_length(new_group_pathkeys);
+
+ /* append the remaining group pathkeys (will be treated as not sorted) */
+ *group_pathkeys = list_concat_unique_ptr(new_group_pathkeys,
+ *group_pathkeys);
+ *group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+
+ return n;
+}
+
+/*
+ * get_useful_group_keys_orderings
+ * Determine which orderings of GROUP BY keys are potentially interesting.
+ *
+ * Returns list of PathKeyInfo items, each representing an interesting ordering
+ * of GROUP BY keys. Each item stores pathkeys and clauses in matching order.
+ *
+ * The function considers (and keeps) multiple group by orderings:
+ *
+ * - the original ordering, as specified by the GROUP BY clause
+ *
+ * - GROUP BY keys reordered to match path ordering (as much as possible), with
+ * the tail reordered to minimize the sort cost
+ *
+ * - GROUP BY keys to match target ORDER BY clause (as much as possible)
+ *
+ * There are other potentially interesting orderings. Just to memorize:
+ * 1. It might be best to match the first ORDER BY key, order the remaining keys
+ * differently and then rely on the incremental sort to fix this).
+ * 2. Place integer fixed-length columns in the front of text/bytea columns
+ * 3. Rearrange columns according to ndistinct estimations.
+ */
+List *
+get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys)
+{
+ Query *parse = root->parse;
+ List *infos = NIL;
+ PathKeyInfo *info;
+
+ List *pathkeys = root->group_pathkeys;
+ List *clauses = root->processed_groupClause;
+
+ /* always return at least the original pathkeys/clauses */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+ infos = lappend(infos, info);
+
+ /*
+ * Should we try generating alternative orderings of the group keys? If
+ * not, we produce only the order specified in the query, i.e. the
+ * optimization is effectively disabled.
+ */
+ if (!enable_group_by_reordering)
+ return infos;
+
+ /* for grouping sets we can't do any reordering */
+ if (parse->groupingSets)
+ return infos;
+
+ /*
+ * If the path is sorted in some way, try reordering the group keys to
+ * match as much of the ordering as possible - we get this sort for free
+ * (mostly).
+ *
+ * We must not do this when there are no grouping sets, because those use
+ * more complex logic to decide the ordering.
+ *
+ * XXX Isn't this somewhat redundant with presorted_keys? Actually, it's
+ * more a complement, because it allows benefiting from incremental sort
+ * as much as possible.
+ */
+ if (path_pathkeys)
+ {
+ group_keys_reorder_by_pathkeys(path_pathkeys, &pathkeys, &clauses,
+ root->num_groupby_pathkeys);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, But we
+ * still want to keep the keys reordered to path_pathkeys.
+ */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost (this time consider
+ * the ORDER BY clause).
+ */
+ if (root->sort_pathkeys)
+ {
+ group_keys_reorder_by_pathkeys(root->sort_pathkeys, &pathkeys, &clauses,
+ root->num_groupby_pathkeys);
+
+ /* keep the group keys reordered to match ordering of input path */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ return infos;
+}
+
/*
* pathkeys_count_contained_in
* Same as pathkeys_contained_in, but also sets length of longest
@@ -1939,6 +2108,54 @@ pathkeys_useful_for_ordering(PlannerInfo *root, List *pathkeys)
return n_common_pathkeys;
}
+/*
+ * pathkeys_useful_for_grouping
+ * Count the number of pathkeys that are useful for grouping (instead of
+ * explicit sort)
+ *
+ * Group pathkeys could be reordered to benefit from the ordering. The
+ * ordering may not be "complete" and may require incremental sort, but that's
+ * fine. So we simply count prefix pathkeys with a matching group key, and
+ * stop once we find the first pathkey without a match.
+ *
+ * So e.g. with pathkeys (a,b,c) and group keys (a,b,e) this determines (a,b)
+ * pathkeys are useful for grouping, and we might do incremental sort to get
+ * path ordered by (a,b,e).
+ *
+ * This logic is necessary to retain paths with ordering not matching grouping
+ * keys directly, without the reordering.
+ *
+ * Returns the length of pathkey prefix with matching group keys.
+ */
+static int
+pathkeys_useful_for_grouping(PlannerInfo *root, List *pathkeys)
+{
+ ListCell *key;
+ int n = 0;
+
+ /* no special ordering requested for grouping */
+ if (root->group_pathkeys == NIL)
+ return 0;
+
+ /* unordered path */
+ if (pathkeys == NIL)
+ return 0;
+
+ /* walk the pathkeys and search for matching group key */
+ foreach(key, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(key);
+
+ /* no matching group key, we're done */
+ if (!list_member_ptr(root->group_pathkeys, pathkey))
+ break;
+
+ n++;
+ }
+
+ return n;
+}
+
/*
* truncate_useless_pathkeys
* Shorten the given pathkey list to just the useful pathkeys.
@@ -1953,6 +2170,9 @@ truncate_useless_pathkeys(PlannerInfo *root,
nuseful = pathkeys_useful_for_merging(root, rel, pathkeys);
nuseful2 = pathkeys_useful_for_ordering(root, pathkeys);
+ if (nuseful2 > nuseful)
+ nuseful = nuseful2;
+ nuseful2 = pathkeys_useful_for_grouping(root, pathkeys);
if (nuseful2 > nuseful)
nuseful = nuseful2;
@@ -1988,6 +2208,8 @@ has_useful_pathkeys(PlannerInfo *root, RelOptInfo *rel)
{
if (rel->joininfo != NIL || rel->has_eclass_joins)
return true; /* might be able to use pathkeys for merging */
+ if (root->group_pathkeys != NIL)
+ return true; /* might be able to use pathkeys for grouping */
if (root->query_pathkeys != NIL)
return true; /* might be able to use them for ordering */
return false; /* definitely useless */
diff --git a/src/backend/optimizer/plan/planner.c b/src/backend/optimizer/plan/planner.c
index fcf647940e..90b0f5b6c5 100644
--- a/src/backend/optimizer/plan/planner.c
+++ b/src/backend/optimizer/plan/planner.c
@@ -2805,8 +2805,9 @@ remove_useless_groupby_columns(PlannerInfo *root)
*
* In principle it might be interesting to consider other orderings of the
* GROUP BY elements, which could match the sort ordering of other
- * possible plans (eg an indexscan) and thereby reduce cost. We don't
- * bother with that, though. Hashed grouping will frequently win anyway.
+ * possible plans (eg an indexscan) and thereby reduce cost. However, we
+ * don't yet have sufficient information to do that here, so that's left until
+ * later in planning. See get_useful_group_keys_orderings().
*
* Note: we need no comparable processing of the distinctClause because
* the parser already enforced that that matches ORDER BY.
@@ -6811,12 +6812,12 @@ done:
static Path *
try_presorting(PlannerInfo *root, RelOptInfo *grouped_rel,
- Path *path, Path *cheapest_path)
+ Path *path, Path *cheapest_path, PathKeyInfo *info)
{
bool is_sorted;
int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
path->pathkeys,
&presorted_keys);
@@ -6842,13 +6843,13 @@ try_presorting(PlannerInfo *root, RelOptInfo *grouped_rel,
path = (Path *) create_sort_path(root,
grouped_rel,
path,
- root->group_pathkeys,
+ info->pathkeys,
-1.0);
else
path = (Path *) create_incremental_sort_path(root,
grouped_rel,
path,
- root->group_pathkeys,
+ info->pathkeys,
presorted_keys,
-1.0);
}
@@ -6885,56 +6886,74 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
+
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ foreach(lc2, pathkey_orderings)
+ {
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
- path = try_presorting(root, grouped_rel, path, cheapest_path);
+ path = try_presorting(root, grouped_rel, path, cheapest_path, info);
- if (path == NULL)
+ if (path == NULL)
continue;
- /* Now decide what to stick atop it */
- if (parse->groupingSets)
- {
- consider_groupingsets_paths(root, grouped_rel,
- path, true, can_hash,
- gd, agg_costs, dNumGroups);
- }
- else if (parse->hasAggs)
- {
- /*
- * We have aggregation, possibly with plain GROUP BY. Make an
- * AggPath.
- */
- add_path(grouped_rel, (Path *)
+ /* Now decide what to stick atop it */
+ if (parse->groupingSets)
+ {
+ consider_groupingsets_paths(root, grouped_rel,
+ path, true, can_hash,
+ gd, agg_costs, dNumGroups);
+ }
+ else if (parse->hasAggs)
+ {
+ /*
+ * We have aggregation, possibly with plain GROUP BY. Make an
+ * AggPath.
+ */
+ add_path(grouped_rel, (Path *)
create_agg_path(root,
grouped_rel,
path,
grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_SIMPLE,
- root->processed_groupClause,
+ info->clauses,
havingQual,
agg_costs,
dNumGroups));
- }
- else if (parse->groupClause)
- {
- /*
- * We have GROUP BY without aggregation or grouping sets. Make
- * a GroupPath.
- */
- add_path(grouped_rel, (Path *)
+ }
+ else if (parse->groupClause)
+ {
+ /*
+ * We have GROUP BY without aggregation or grouping sets. Make
+ * a GroupPath.
+ */
+ add_path(grouped_rel, (Path *)
create_group_path(root,
grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
havingQual,
dNumGroups));
- }
- else
- {
- /* Other cases should have been handled above */
- Assert(false);
+ }
+ else
+ {
+ /* Other cases should have been handled above */
+ Assert(false);
+ }
}
}
@@ -6946,35 +6965,55 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
{
foreach(lc, partially_grouped_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
- path = try_presorting(root, grouped_rel, path,
- partially_grouped_rel->cheapest_total_path);
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
- if (path == NULL)
- continue;
+ Assert(list_length(pathkey_orderings) > 0);
- if (parse->hasAggs)
- add_path(grouped_rel, (Path *)
+ /* process all potentially interesting grouping reorderings */
+ foreach(lc2, pathkey_orderings)
+ {
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
+
+ path = try_presorting(root, grouped_rel, path,
+ partially_grouped_rel->cheapest_total_path,
+ info);
+
+ if (path == NULL)
+ continue;
+
+ if (parse->hasAggs)
+ add_path(grouped_rel, (Path *)
create_agg_path(root,
grouped_rel,
path,
grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_FINAL_DESERIAL,
- root->processed_groupClause,
+ info->clauses,
havingQual,
agg_final_costs,
dNumGroups));
- else
- add_path(grouped_rel, (Path *)
+ else
+ add_path(grouped_rel, (Path *)
create_group_path(root,
grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
havingQual,
dNumGroups));
+ }
}
}
}
@@ -7177,34 +7216,53 @@ create_partial_grouping_paths(PlannerInfo *root,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
- path = try_presorting(root, partially_grouped_rel, path,
- cheapest_total_path);
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
- if (path == NULL)
- continue;
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach(lc2, pathkey_orderings)
+ {
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
+
+ path = try_presorting(root, partially_grouped_rel, path,
+ cheapest_total_path, info);
- if (parse->hasAggs)
- add_path(partially_grouped_rel, (Path *)
+ if (path == NULL)
+ continue;
+
+ if (parse->hasAggs)
+ add_path(partially_grouped_rel, (Path *)
create_agg_path(root,
partially_grouped_rel,
path,
partially_grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_INITIAL_SERIAL,
- root->processed_groupClause,
+ info->clauses,
NIL,
agg_partial_costs,
dNumPartialGroups));
- else
- add_path(partially_grouped_rel, (Path *)
+ else
+ add_path(partially_grouped_rel, (Path *)
create_group_path(root,
partially_grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
NIL,
dNumPartialGroups));
+ }
}
}
@@ -7213,34 +7271,54 @@ create_partial_grouping_paths(PlannerInfo *root,
/* Similar to above logic, but for partial paths. */
foreach(lc, input_rel->partial_pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
- path = try_presorting(root, partially_grouped_rel, path,
- cheapest_partial_path);
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
- if (path == NULL)
- continue;
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach(lc2, pathkey_orderings)
+ {
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
- if (parse->hasAggs)
- add_partial_path(partially_grouped_rel, (Path *)
+ path = try_presorting(root, partially_grouped_rel, path,
+ cheapest_partial_path, info);
+
+ if (path == NULL)
+ continue;
+
+ if (parse->hasAggs)
+ add_partial_path(partially_grouped_rel, (Path *)
create_agg_path(root,
partially_grouped_rel,
path,
partially_grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_INITIAL_SERIAL,
- root->processed_groupClause,
+ info->clauses,
NIL,
agg_partial_costs,
dNumPartialPartialGroups));
- else
- add_partial_path(partially_grouped_rel, (Path *)
+ else
+ add_partial_path(partially_grouped_rel, (Path *)
create_group_path(root,
partially_grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
NIL,
dNumPartialPartialGroups));
+ }
}
}
@@ -7354,6 +7432,8 @@ gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
* We can also skip the entire loop when we only have a single-item
* group_pathkeys because then we can't possibly have a presorted prefix
* of the list without having the list be fully sorted.
+ *
+ * XXX Shouldn't this also consider the group-key-reordering?
*/
if (!enable_incremental_sort || list_length(root->group_pathkeys) == 1)
return;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index e53ebc6dc2..7fe58518d7 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -1050,6 +1050,16 @@ struct config_bool ConfigureNamesBool[] =
true,
NULL, NULL, NULL
},
+ {
+ {"enable_group_by_reordering", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("Enables reordering of GROUP BY keys."),
+ NULL,
+ GUC_EXPLAIN
+ },
+ &enable_group_by_reordering,
+ true,
+ NULL, NULL, NULL
+ },
{
{"geqo", PGC_USERSET, QUERY_TUNING_GEQO,
gettext_noop("Enables genetic query optimization."),
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index b2809c711a..dc82fceb05 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -399,6 +399,7 @@
#enable_seqscan = on
#enable_sort = on
#enable_tidscan = on
+#enable_group_by_reordering = on
# - Planner Cost Constants -
diff --git a/src/include/nodes/pathnodes.h b/src/include/nodes/pathnodes.h
index b9713ec9aa..0a251984c1 100644
--- a/src/include/nodes/pathnodes.h
+++ b/src/include/nodes/pathnodes.h
@@ -1456,6 +1456,16 @@ typedef struct PathKey
bool pk_nulls_first; /* do NULLs come before normal values? */
} PathKey;
+/*
+ * Combines information about pathkeys and the associated clauses.
+ */
+typedef struct PathKeyInfo
+{
+ NodeTag type;
+ List *pathkeys;
+ List *clauses;
+} PathKeyInfo;
+
/*
* VolatileFunctionStatus -- allows nodes to cache their
* contain_volatile_functions properties. VOLATILITY_UNKNOWN means not yet
diff --git a/src/include/optimizer/paths.h b/src/include/optimizer/paths.h
index efd4abc28f..e80156c49f 100644
--- a/src/include/optimizer/paths.h
+++ b/src/include/optimizer/paths.h
@@ -24,6 +24,7 @@ extern PGDLLIMPORT bool enable_geqo;
extern PGDLLIMPORT int geqo_threshold;
extern PGDLLIMPORT int min_parallel_table_scan_size;
extern PGDLLIMPORT int min_parallel_index_scan_size;
+extern PGDLLIMPORT bool enable_group_by_reordering;
/* Hook for plugins to get control in set_rel_pathlist() */
typedef void (*set_rel_pathlist_hook_type) (PlannerInfo *root,
@@ -204,6 +205,8 @@ typedef enum
extern PathKeysComparison compare_pathkeys(List *keys1, List *keys2);
extern bool pathkeys_contained_in(List *keys1, List *keys2);
extern bool pathkeys_count_contained_in(List *keys1, List *keys2, int *n_common);
+extern List *get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys);
extern Path *get_cheapest_path_for_pathkeys(List *paths, List *pathkeys,
Relids required_outer,
CostSelector cost_criterion,
diff --git a/src/test/regress/expected/aggregates.out b/src/test/regress/expected/aggregates.out
index f635c5a1af..423c8ec3b6 100644
--- a/src/test/regress/expected/aggregates.out
+++ b/src/test/regress/expected/aggregates.out
@@ -2728,6 +2728,138 @@ SELECT balk(hundred) FROM tenk1;
(1 row)
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+CREATE TABLE btg AS SELECT
+ i % 100 AS x,
+ i % 100 AS y,
+ 'abc' || i % 10 AS z,
+ i AS w
+FROM generate_series(1,10000) AS i;
+CREATE INDEX abc ON btg(x,y);
+ANALYZE btg;
+-- GROUP BY optimization by reorder columns by frequency
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+-- Utilize index scan ordering to avoid a Sort operation
+EXPLAIN (COSTS OFF) SELECT count(*) FROM btg GROUP BY x,y;
+ QUERY PLAN
+----------------------------------------
+ GroupAggregate
+ Group Key: x, y
+ -> Index Only Scan using abc on btg
+(3 rows)
+
+EXPLAIN (COSTS OFF) SELECT count(*) FROM btg GROUP BY y,x;
+ QUERY PLAN
+----------------------------------------
+ GroupAggregate
+ Group Key: x, y
+ -> Index Only Scan using abc on btg
+(3 rows)
+
+-- Engage incremental sort
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY x,y,z,w;
+ QUERY PLAN
+-----------------------------------------
+ Group
+ Group Key: x, y, z, w
+ -> Incremental Sort
+ Sort Key: x, y, z, w
+ Presorted Key: x, y
+ -> Index Scan using abc on btg
+(6 rows)
+
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY z,y,w,x;
+ QUERY PLAN
+-----------------------------------------
+ Group
+ Group Key: x, y, z, w
+ -> Incremental Sort
+ Sort Key: x, y, z, w
+ Presorted Key: x, y
+ -> Index Scan using abc on btg
+(6 rows)
+
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY w,z,x,y;
+ QUERY PLAN
+-----------------------------------------
+ Group
+ Group Key: x, y, w, z
+ -> Incremental Sort
+ Sort Key: x, y, w, z
+ Presorted Key: x, y
+ -> Index Scan using abc on btg
+(6 rows)
+
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY w,x,z,y;
+ QUERY PLAN
+-----------------------------------------
+ Group
+ Group Key: x, y, w, z
+ -> Incremental Sort
+ Sort Key: x, y, w, z
+ Presorted Key: x, y
+ -> Index Scan using abc on btg
+(6 rows)
+
+-- Subqueries
+explain (COSTS OFF) SELECT x,y
+FROM (SELECT * FROM btg ORDER BY x,y,w,z) AS q1
+GROUP BY (w,x,z,y);
+ QUERY PLAN
+----------------------------------------------
+ Group
+ Group Key: btg.x, btg.y, btg.w, btg.z
+ -> Incremental Sort
+ Sort Key: btg.x, btg.y, btg.w, btg.z
+ Presorted Key: btg.x, btg.y
+ -> Index Scan using abc on btg
+(6 rows)
+
+explain (COSTS OFF) SELECT x,y
+FROM (SELECT * FROM btg ORDER BY x,y,w,z LIMIT 100) AS q1
+GROUP BY (w,x,z,y);
+ QUERY PLAN
+----------------------------------------------------
+ Group
+ Group Key: btg.x, btg.y, btg.w, btg.z
+ -> Limit
+ -> Incremental Sort
+ Sort Key: btg.x, btg.y, btg.w, btg.z
+ Presorted Key: btg.x, btg.y
+ -> Index Scan using abc on btg
+(7 rows)
+
+-- Should work with and without GROUP-BY optimization
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY w,x,z,y ORDER BY y,x,z,w;
+ QUERY PLAN
+------------------------------
+ Group
+ Group Key: y, x, z, w
+ -> Sort
+ Sort Key: y, x, z, w
+ -> Seq Scan on btg
+(5 rows)
+
+-- Utilize incremental sort to make the ORDER BY rule a bit cheaper
+explain (COSTS OFF) SELECT x,w FROM btg GROUP BY w,x,y,z ORDER BY x*x,z;
+ QUERY PLAN
+-----------------------------------------------
+ Sort
+ Sort Key: ((x * x)), z
+ -> Group
+ Group Key: x, y, w, z
+ -> Incremental Sort
+ Sort Key: x, y, w, z
+ Presorted Key: x, y
+ -> Index Scan using abc on btg
+(8 rows)
+
+DROP TABLE btg;
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/expected/sysviews.out b/src/test/regress/expected/sysviews.out
index 271313ebf8..9be7aca2b8 100644
--- a/src/test/regress/expected/sysviews.out
+++ b/src/test/regress/expected/sysviews.out
@@ -114,6 +114,7 @@ select name, setting from pg_settings where name like 'enable%';
enable_async_append | on
enable_bitmapscan | on
enable_gathermerge | on
+ enable_group_by_reordering | on
enable_hashagg | on
enable_hashjoin | on
enable_incremental_sort | on
@@ -133,7 +134,7 @@ select name, setting from pg_settings where name like 'enable%';
enable_seqscan | on
enable_sort | on
enable_tidscan | on
-(22 rows)
+(23 rows)
-- There are always wait event descriptions for various types.
select type, count(*) > 0 as ok FROM pg_wait_events
diff --git a/src/test/regress/sql/aggregates.sql b/src/test/regress/sql/aggregates.sql
index cc8f0efad5..b9fcceedd7 100644
--- a/src/test/regress/sql/aggregates.sql
+++ b/src/test/regress/sql/aggregates.sql
@@ -1181,6 +1181,53 @@ SELECT balk(hundred) FROM tenk1;
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+
+CREATE TABLE btg AS SELECT
+ i % 100 AS x,
+ i % 100 AS y,
+ 'abc' || i % 10 AS z,
+ i AS w
+FROM generate_series(1,10000) AS i;
+CREATE INDEX abc ON btg(x,y);
+ANALYZE btg;
+
+-- GROUP BY optimization by reorder columns by frequency
+
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+
+-- Utilize index scan ordering to avoid a Sort operation
+EXPLAIN (COSTS OFF) SELECT count(*) FROM btg GROUP BY x,y;
+EXPLAIN (COSTS OFF) SELECT count(*) FROM btg GROUP BY y,x;
+
+-- Engage incremental sort
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY x,y,z,w;
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY z,y,w,x;
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY w,z,x,y;
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY w,x,z,y;
+
+-- Subqueries
+explain (COSTS OFF) SELECT x,y
+FROM (SELECT * FROM btg ORDER BY x,y,w,z) AS q1
+GROUP BY (w,x,z,y);
+explain (COSTS OFF) SELECT x,y
+FROM (SELECT * FROM btg ORDER BY x,y,w,z LIMIT 100) AS q1
+GROUP BY (w,x,z,y);
+
+-- Should work with and without GROUP-BY optimization
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY w,x,z,y ORDER BY y,x,z,w;
+
+-- Utilize incremental sort to make the ORDER BY rule a bit cheaper
+explain (COSTS OFF) SELECT x,w FROM btg GROUP BY w,x,y,z ORDER BY x*x,z;
+
+DROP TABLE btg;
+
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
--
2.43.0
Attachments:
[text/plain] 0001-Generalize-common-code-of-adding-sort-before-generat.patch (8.5K, ../../[email protected]/2-0001-Generalize-common-code-of-adding-sort-before-generat.patch)
download | inline diff:
From ea068e221a754a463142575f6e0360d3118475f8 Mon Sep 17 00:00:00 2001
From: "Andrey V. Lepikhov" <[email protected]>
Date: Tue, 9 Jan 2024 12:00:27 +0700
Subject: [PATCH 1/2] Generalize common code of adding sort before generation
of grouping paths.
---
src/backend/optimizer/plan/planner.c | 209 ++++++++-------------------
1 file changed, 61 insertions(+), 148 deletions(-)
diff --git a/src/backend/optimizer/plan/planner.c b/src/backend/optimizer/plan/planner.c
index 667723b675..fcf647940e 100644
--- a/src/backend/optimizer/plan/planner.c
+++ b/src/backend/optimizer/plan/planner.c
@@ -6809,6 +6809,52 @@ done:
return parallel_workers;
}
+static Path *
+try_presorting(PlannerInfo *root, RelOptInfo *grouped_rel,
+ Path *path, Path *cheapest_path)
+{
+ bool is_sorted;
+ int presorted_keys;
+
+ is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
+ path->pathkeys,
+ &presorted_keys);
+
+ if (!is_sorted)
+ {
+ /*
+ * Try at least sorting the cheapest path and also try
+ * incrementally sorting any path which is partially sorted
+ * already (no need to deal with paths which have presorted
+ * keys when incremental sort is disabled unless it's the
+ * cheapest input path).
+ */
+ if (path != cheapest_path &&
+ (presorted_keys == 0 || !enable_incremental_sort))
+ return NULL;
+
+ /*
+ * We've no need to consider both a sort and incremental sort.
+ * We'll just do a sort if there are no presorted keys and an
+ * incremental sort when there are presorted keys.
+ */
+ if (presorted_keys == 0 || !enable_incremental_sort)
+ path = (Path *) create_sort_path(root,
+ grouped_rel,
+ path,
+ root->group_pathkeys,
+ -1.0);
+ else
+ path = (Path *) create_incremental_sort_path(root,
+ grouped_rel,
+ path,
+ root->group_pathkeys,
+ presorted_keys,
+ -1.0);
+ }
+ return path;
+}
+
/*
* add_paths_to_grouping_rel
*
@@ -6840,45 +6886,11 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
foreach(lc, input_rel->pathlist)
{
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
-
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
- if (!is_sorted)
- {
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially sorted
- * already (no need to deal with paths which have presorted
- * keys when incremental sort is disabled unless it's the
- * cheapest input path).
- */
- if (path != cheapest_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ path = try_presorting(root, grouped_rel, path, cheapest_path);
- /*
- * We've no need to consider both a sort and incremental sort.
- * We'll just do a sort if there are no presorted keys and an
- * incremental sort when there are presorted keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ if (path == NULL)
+ continue;
/* Now decide what to stick atop it */
if (parse->groupingSets)
@@ -6935,46 +6947,12 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
foreach(lc, partially_grouped_rel->pathlist)
{
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
-
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
- if (!is_sorted)
- {
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially
- * sorted already (no need to deal with paths which have
- * presorted keys when incremental sort is disabled unless
- * it's the cheapest input path).
- */
- if (path != partially_grouped_rel->cheapest_total_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ path = try_presorting(root, grouped_rel, path,
+ partially_grouped_rel->cheapest_total_path);
- /*
- * We've no need to consider both a sort and incremental
- * sort. We'll just do a sort if there are no pre-sorted
- * keys and an incremental sort when there are presorted
- * keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ if (path == NULL)
+ continue;
if (parse->hasAggs)
add_path(grouped_rel, (Path *)
@@ -7200,44 +7178,12 @@ create_partial_grouping_paths(PlannerInfo *root,
foreach(lc, input_rel->pathlist)
{
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
- if (!is_sorted)
- {
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially sorted
- * already (no need to deal with paths which have presorted
- * keys when incremental sort is disabled unless it's the
- * cheapest input path).
- */
- if (path != cheapest_total_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ path = try_presorting(root, partially_grouped_rel, path,
+ cheapest_total_path);
- /*
- * We've no need to consider both a sort and incremental sort.
- * We'll just do a sort if there are no presorted keys and an
- * incremental sort when there are presorted keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ if (path == NULL)
+ continue;
if (parse->hasAggs)
add_path(partially_grouped_rel, (Path *)
@@ -7268,45 +7214,12 @@ create_partial_grouping_paths(PlannerInfo *root,
foreach(lc, input_rel->partial_pathlist)
{
Path *path = (Path *) lfirst(lc);
- bool is_sorted;
- int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
- path->pathkeys,
- &presorted_keys);
-
- if (!is_sorted)
- {
- /*
- * Try at least sorting the cheapest path and also try
- * incrementally sorting any path which is partially sorted
- * already (no need to deal with paths which have presorted
- * keys when incremental sort is disabled unless it's the
- * cheapest input path).
- */
- if (path != cheapest_partial_path &&
- (presorted_keys == 0 || !enable_incremental_sort))
- continue;
+ path = try_presorting(root, partially_grouped_rel, path,
+ cheapest_partial_path);
- /*
- * We've no need to consider both a sort and incremental sort.
- * We'll just do a sort if there are no presorted keys and an
- * incremental sort when there are presorted keys.
- */
- if (presorted_keys == 0 || !enable_incremental_sort)
- path = (Path *) create_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- -1.0);
- else
- path = (Path *) create_incremental_sort_path(root,
- partially_grouped_rel,
- path,
- root->group_pathkeys,
- presorted_keys,
- -1.0);
- }
+ if (path == NULL)
+ continue;
if (parse->hasAggs)
add_partial_path(partially_grouped_rel, (Path *)
--
2.43.0
[text/plain] 0002-Explore-alternative-orderings-of-group-by-pathkeys-d.patch (31.5K, ../../[email protected]/3-0002-Explore-alternative-orderings-of-group-by-pathkeys-d.patch)
download | inline diff:
From 886227bd33a368df1b564121506842be39d31f3d Mon Sep 17 00:00:00 2001
From: "Andrey V. Lepikhov" <[email protected]>
Date: Tue, 9 Jan 2024 12:32:15 +0700
Subject: [PATCH 2/2] Explore alternative orderings of group-by pathkeys during
optimization.
When evaluating a query with a multi-column GROUP BY clause, we can minimize
sort operations or avoid them if we synchronize the order of GROUP BY clauses
with the ORDER BY sort clause or sort order, which comes from the underlying
query tree. Grouping does not imply any ordering, so we can compare
the keys in arbitrary order, and a Hash Agg leverages this. But for Group Agg,
we simply compared keys in the order specified in the query. This commit
explores alternative ordering of the keys, trying to find a cheaper one.
The ordering of group keys may interact with other parts of the query, some of
which may not be known while planning the grouping. For example, there may be
an explicit ORDER BY clause or some other ordering-dependent operation higher up
in the query, and using the same ordering may allow using either incremental
sort or even eliminating the sort entirely.
The patch always keeps the ordering specified in the query, assuming the user
might have additional insights.
This introduces a new GUC enable_group_by_reordering so that the optimization
may be disabled if needed.
---
src/backend/optimizer/path/equivclass.c | 13 +-
src/backend/optimizer/path/pathkeys.c | 222 ++++++++++++++++++
src/backend/optimizer/plan/planner.c | 214 +++++++++++------
src/backend/utils/misc/guc_tables.c | 10 +
src/backend/utils/misc/postgresql.conf.sample | 1 +
src/include/nodes/pathnodes.h | 10 +
src/include/optimizer/paths.h | 3 +
src/test/regress/expected/aggregates.out | 132 +++++++++++
src/test/regress/expected/sysviews.out | 3 +-
src/test/regress/sql/aggregates.sql | 47 ++++
10 files changed, 586 insertions(+), 69 deletions(-)
diff --git a/src/backend/optimizer/path/equivclass.c b/src/backend/optimizer/path/equivclass.c
index e86dfeaecd..7dd14d0a43 100644
--- a/src/backend/optimizer/path/equivclass.c
+++ b/src/backend/optimizer/path/equivclass.c
@@ -652,7 +652,18 @@ get_eclass_for_sort_expr(PlannerInfo *root,
if (opcintype == cur_em->em_datatype &&
equal(expr, cur_em->em_expr))
- return cur_ec; /* Match! */
+ {
+ /*
+ * Match!
+ *
+ * Copy the sortref if it wasn't set yet. That may happen if
+ * the ec was constructed from WHERE clause, i.e. it doesn't
+ * have a target reference at all.
+ */
+ if (cur_ec->ec_sortref == 0 && sortref > 0)
+ cur_ec->ec_sortref = sortref;
+ return cur_ec;
+ }
}
}
diff --git a/src/backend/optimizer/path/pathkeys.c b/src/backend/optimizer/path/pathkeys.c
index ca94a31f71..8bc4eb9f03 100644
--- a/src/backend/optimizer/path/pathkeys.c
+++ b/src/backend/optimizer/path/pathkeys.c
@@ -28,6 +28,8 @@
#include "partitioning/partbounds.h"
#include "utils/lsyscache.h"
+/* Consider reordering of GROUP BY keys? */
+bool enable_group_by_reordering = true;
static bool pathkey_is_redundant(PathKey *new_pathkey, List *pathkeys);
static bool matches_boolean_partition_clause(RestrictInfo *rinfo,
@@ -350,6 +352,173 @@ pathkeys_contained_in(List *keys1, List *keys2)
return false;
}
+/*
+ * group_keys_reorder_by_pathkeys
+ * Reorder GROUP BY keys to match pathkeys of input path.
+ *
+ * Function returns new lists (pathkeys and clauses), original GROUP BY lists
+ * stay untouched.
+ *
+ * Returns the number of GROUP BY keys with a matching pathkey.
+ */
+static int
+group_keys_reorder_by_pathkeys(List *pathkeys, List **group_pathkeys,
+ List **group_clauses,
+ int num_groupby_pathkeys)
+{
+ List *new_group_pathkeys = NIL,
+ *new_group_clauses = NIL;
+ ListCell *lc;
+ int n;
+
+ if (pathkeys == NIL || *group_pathkeys == NIL)
+ return 0;
+
+ /*
+ * Walk the pathkeys (determining ordering of the input path) and see if
+ * there's a matching GROUP BY key. If we find one, we append it to the
+ * list, and do the same for the clauses.
+ *
+ * Once we find the first pathkey without a matching GROUP BY key, the
+ * rest of the pathkeys are useless and can't be used to evaluate the
+ * grouping, so we abort the loop and ignore the remaining pathkeys.
+ *
+ * XXX Pathkeys are built in a way to allow simply comparing pointers.
+ */
+ foreach(lc, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(lc);
+ SortGroupClause *sgc;
+
+ if (foreach_current_index(lc) >= num_groupby_pathkeys ||
+ !list_member_ptr(*group_pathkeys, pathkey))
+ /* */
+ break;
+
+ /* abort on first mismatch */
+ sgc = get_sortgroupref_clause_noerr(pathkey->pk_eclass->ec_sortref,
+ *group_clauses);
+ if (!sgc)
+ /* The grouping clause is not cover this pathkey */
+ break;
+
+ new_group_pathkeys = lappend(new_group_pathkeys, pathkey);
+ new_group_clauses = lappend(new_group_clauses, sgc);
+ }
+
+ /* remember the number of pathkeys with a matching GROUP BY key */
+ n = list_length(new_group_pathkeys);
+
+ /* append the remaining group pathkeys (will be treated as not sorted) */
+ *group_pathkeys = list_concat_unique_ptr(new_group_pathkeys,
+ *group_pathkeys);
+ *group_clauses = list_concat_unique_ptr(new_group_clauses,
+ *group_clauses);
+
+ return n;
+}
+
+/*
+ * get_useful_group_keys_orderings
+ * Determine which orderings of GROUP BY keys are potentially interesting.
+ *
+ * Returns list of PathKeyInfo items, each representing an interesting ordering
+ * of GROUP BY keys. Each item stores pathkeys and clauses in matching order.
+ *
+ * The function considers (and keeps) multiple group by orderings:
+ *
+ * - the original ordering, as specified by the GROUP BY clause
+ *
+ * - GROUP BY keys reordered to match path ordering (as much as possible), with
+ * the tail reordered to minimize the sort cost
+ *
+ * - GROUP BY keys to match target ORDER BY clause (as much as possible)
+ *
+ * There are other potentially interesting orderings. Just to memorize:
+ * 1. It might be best to match the first ORDER BY key, order the remaining keys
+ * differently and then rely on the incremental sort to fix this).
+ * 2. Place integer fixed-length columns in the front of text/bytea columns
+ * 3. Rearrange columns according to ndistinct estimations.
+ */
+List *
+get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys)
+{
+ Query *parse = root->parse;
+ List *infos = NIL;
+ PathKeyInfo *info;
+
+ List *pathkeys = root->group_pathkeys;
+ List *clauses = root->processed_groupClause;
+
+ /* always return at least the original pathkeys/clauses */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+ infos = lappend(infos, info);
+
+ /*
+ * Should we try generating alternative orderings of the group keys? If
+ * not, we produce only the order specified in the query, i.e. the
+ * optimization is effectively disabled.
+ */
+ if (!enable_group_by_reordering)
+ return infos;
+
+ /* for grouping sets we can't do any reordering */
+ if (parse->groupingSets)
+ return infos;
+
+ /*
+ * If the path is sorted in some way, try reordering the group keys to
+ * match as much of the ordering as possible - we get this sort for free
+ * (mostly).
+ *
+ * We must not do this when there are no grouping sets, because those use
+ * more complex logic to decide the ordering.
+ *
+ * XXX Isn't this somewhat redundant with presorted_keys? Actually, it's
+ * more a complement, because it allows benefiting from incremental sort
+ * as much as possible.
+ */
+ if (path_pathkeys)
+ {
+ group_keys_reorder_by_pathkeys(path_pathkeys, &pathkeys, &clauses,
+ root->num_groupby_pathkeys);
+
+ /*
+ * reorder the tail to minimize sort cost
+ *
+ * XXX Ignore the return value - there may be nothing to reorder, But we
+ * still want to keep the keys reordered to path_pathkeys.
+ */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ /*
+ * Try reordering pathkeys to minimize the sort cost (this time consider
+ * the ORDER BY clause).
+ */
+ if (root->sort_pathkeys)
+ {
+ group_keys_reorder_by_pathkeys(root->sort_pathkeys, &pathkeys, &clauses,
+ root->num_groupby_pathkeys);
+
+ /* keep the group keys reordered to match ordering of input path */
+ info = makeNode(PathKeyInfo);
+ info->pathkeys = pathkeys;
+ info->clauses = clauses;
+
+ infos = lappend(infos, info);
+ }
+
+ return infos;
+}
+
/*
* pathkeys_count_contained_in
* Same as pathkeys_contained_in, but also sets length of longest
@@ -1939,6 +2108,54 @@ pathkeys_useful_for_ordering(PlannerInfo *root, List *pathkeys)
return n_common_pathkeys;
}
+/*
+ * pathkeys_useful_for_grouping
+ * Count the number of pathkeys that are useful for grouping (instead of
+ * explicit sort)
+ *
+ * Group pathkeys could be reordered to benefit from the ordering. The
+ * ordering may not be "complete" and may require incremental sort, but that's
+ * fine. So we simply count prefix pathkeys with a matching group key, and
+ * stop once we find the first pathkey without a match.
+ *
+ * So e.g. with pathkeys (a,b,c) and group keys (a,b,e) this determines (a,b)
+ * pathkeys are useful for grouping, and we might do incremental sort to get
+ * path ordered by (a,b,e).
+ *
+ * This logic is necessary to retain paths with ordering not matching grouping
+ * keys directly, without the reordering.
+ *
+ * Returns the length of pathkey prefix with matching group keys.
+ */
+static int
+pathkeys_useful_for_grouping(PlannerInfo *root, List *pathkeys)
+{
+ ListCell *key;
+ int n = 0;
+
+ /* no special ordering requested for grouping */
+ if (root->group_pathkeys == NIL)
+ return 0;
+
+ /* unordered path */
+ if (pathkeys == NIL)
+ return 0;
+
+ /* walk the pathkeys and search for matching group key */
+ foreach(key, pathkeys)
+ {
+ PathKey *pathkey = (PathKey *) lfirst(key);
+
+ /* no matching group key, we're done */
+ if (!list_member_ptr(root->group_pathkeys, pathkey))
+ break;
+
+ n++;
+ }
+
+ return n;
+}
+
/*
* truncate_useless_pathkeys
* Shorten the given pathkey list to just the useful pathkeys.
@@ -1953,6 +2170,9 @@ truncate_useless_pathkeys(PlannerInfo *root,
nuseful = pathkeys_useful_for_merging(root, rel, pathkeys);
nuseful2 = pathkeys_useful_for_ordering(root, pathkeys);
+ if (nuseful2 > nuseful)
+ nuseful = nuseful2;
+ nuseful2 = pathkeys_useful_for_grouping(root, pathkeys);
if (nuseful2 > nuseful)
nuseful = nuseful2;
@@ -1988,6 +2208,8 @@ has_useful_pathkeys(PlannerInfo *root, RelOptInfo *rel)
{
if (rel->joininfo != NIL || rel->has_eclass_joins)
return true; /* might be able to use pathkeys for merging */
+ if (root->group_pathkeys != NIL)
+ return true; /* might be able to use pathkeys for grouping */
if (root->query_pathkeys != NIL)
return true; /* might be able to use them for ordering */
return false; /* definitely useless */
diff --git a/src/backend/optimizer/plan/planner.c b/src/backend/optimizer/plan/planner.c
index fcf647940e..90b0f5b6c5 100644
--- a/src/backend/optimizer/plan/planner.c
+++ b/src/backend/optimizer/plan/planner.c
@@ -2805,8 +2805,9 @@ remove_useless_groupby_columns(PlannerInfo *root)
*
* In principle it might be interesting to consider other orderings of the
* GROUP BY elements, which could match the sort ordering of other
- * possible plans (eg an indexscan) and thereby reduce cost. We don't
- * bother with that, though. Hashed grouping will frequently win anyway.
+ * possible plans (eg an indexscan) and thereby reduce cost. However, we
+ * don't yet have sufficient information to do that here, so that's left until
+ * later in planning. See get_useful_group_keys_orderings().
*
* Note: we need no comparable processing of the distinctClause because
* the parser already enforced that that matches ORDER BY.
@@ -6811,12 +6812,12 @@ done:
static Path *
try_presorting(PlannerInfo *root, RelOptInfo *grouped_rel,
- Path *path, Path *cheapest_path)
+ Path *path, Path *cheapest_path, PathKeyInfo *info)
{
bool is_sorted;
int presorted_keys;
- is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
+ is_sorted = pathkeys_count_contained_in(info->pathkeys,
path->pathkeys,
&presorted_keys);
@@ -6842,13 +6843,13 @@ try_presorting(PlannerInfo *root, RelOptInfo *grouped_rel,
path = (Path *) create_sort_path(root,
grouped_rel,
path,
- root->group_pathkeys,
+ info->pathkeys,
-1.0);
else
path = (Path *) create_incremental_sort_path(root,
grouped_rel,
path,
- root->group_pathkeys,
+ info->pathkeys,
presorted_keys,
-1.0);
}
@@ -6885,56 +6886,74 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
+
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
+
+ Assert(list_length(pathkey_orderings) > 0);
+
+ foreach(lc2, pathkey_orderings)
+ {
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
- path = try_presorting(root, grouped_rel, path, cheapest_path);
+ path = try_presorting(root, grouped_rel, path, cheapest_path, info);
- if (path == NULL)
+ if (path == NULL)
continue;
- /* Now decide what to stick atop it */
- if (parse->groupingSets)
- {
- consider_groupingsets_paths(root, grouped_rel,
- path, true, can_hash,
- gd, agg_costs, dNumGroups);
- }
- else if (parse->hasAggs)
- {
- /*
- * We have aggregation, possibly with plain GROUP BY. Make an
- * AggPath.
- */
- add_path(grouped_rel, (Path *)
+ /* Now decide what to stick atop it */
+ if (parse->groupingSets)
+ {
+ consider_groupingsets_paths(root, grouped_rel,
+ path, true, can_hash,
+ gd, agg_costs, dNumGroups);
+ }
+ else if (parse->hasAggs)
+ {
+ /*
+ * We have aggregation, possibly with plain GROUP BY. Make an
+ * AggPath.
+ */
+ add_path(grouped_rel, (Path *)
create_agg_path(root,
grouped_rel,
path,
grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_SIMPLE,
- root->processed_groupClause,
+ info->clauses,
havingQual,
agg_costs,
dNumGroups));
- }
- else if (parse->groupClause)
- {
- /*
- * We have GROUP BY without aggregation or grouping sets. Make
- * a GroupPath.
- */
- add_path(grouped_rel, (Path *)
+ }
+ else if (parse->groupClause)
+ {
+ /*
+ * We have GROUP BY without aggregation or grouping sets. Make
+ * a GroupPath.
+ */
+ add_path(grouped_rel, (Path *)
create_group_path(root,
grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
havingQual,
dNumGroups));
- }
- else
- {
- /* Other cases should have been handled above */
- Assert(false);
+ }
+ else
+ {
+ /* Other cases should have been handled above */
+ Assert(false);
+ }
}
}
@@ -6946,35 +6965,55 @@ add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
{
foreach(lc, partially_grouped_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
- path = try_presorting(root, grouped_rel, path,
- partially_grouped_rel->cheapest_total_path);
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
- if (path == NULL)
- continue;
+ Assert(list_length(pathkey_orderings) > 0);
- if (parse->hasAggs)
- add_path(grouped_rel, (Path *)
+ /* process all potentially interesting grouping reorderings */
+ foreach(lc2, pathkey_orderings)
+ {
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
+
+ path = try_presorting(root, grouped_rel, path,
+ partially_grouped_rel->cheapest_total_path,
+ info);
+
+ if (path == NULL)
+ continue;
+
+ if (parse->hasAggs)
+ add_path(grouped_rel, (Path *)
create_agg_path(root,
grouped_rel,
path,
grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_FINAL_DESERIAL,
- root->processed_groupClause,
+ info->clauses,
havingQual,
agg_final_costs,
dNumGroups));
- else
- add_path(grouped_rel, (Path *)
+ else
+ add_path(grouped_rel, (Path *)
create_group_path(root,
grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
havingQual,
dNumGroups));
+ }
}
}
}
@@ -7177,34 +7216,53 @@ create_partial_grouping_paths(PlannerInfo *root,
*/
foreach(lc, input_rel->pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
- path = try_presorting(root, partially_grouped_rel, path,
- cheapest_total_path);
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
- if (path == NULL)
- continue;
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach(lc2, pathkey_orderings)
+ {
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
+
+ path = try_presorting(root, partially_grouped_rel, path,
+ cheapest_total_path, info);
- if (parse->hasAggs)
- add_path(partially_grouped_rel, (Path *)
+ if (path == NULL)
+ continue;
+
+ if (parse->hasAggs)
+ add_path(partially_grouped_rel, (Path *)
create_agg_path(root,
partially_grouped_rel,
path,
partially_grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_INITIAL_SERIAL,
- root->processed_groupClause,
+ info->clauses,
NIL,
agg_partial_costs,
dNumPartialGroups));
- else
- add_path(partially_grouped_rel, (Path *)
+ else
+ add_path(partially_grouped_rel, (Path *)
create_group_path(root,
partially_grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
NIL,
dNumPartialGroups));
+ }
}
}
@@ -7213,34 +7271,54 @@ create_partial_grouping_paths(PlannerInfo *root,
/* Similar to above logic, but for partial paths. */
foreach(lc, input_rel->partial_pathlist)
{
+ ListCell *lc2;
Path *path = (Path *) lfirst(lc);
+ Path *path_save = path;
+ List *pathkey_orderings = NIL;
- path = try_presorting(root, partially_grouped_rel, path,
- cheapest_partial_path);
+ /* generate alternative group orderings that might be useful */
+ pathkey_orderings = get_useful_group_keys_orderings(root,
+ path->rows,
+ path->pathkeys);
- if (path == NULL)
- continue;
+ Assert(list_length(pathkey_orderings) > 0);
+
+ /* process all potentially interesting grouping reorderings */
+ foreach(lc2, pathkey_orderings)
+ {
+ PathKeyInfo *info = (PathKeyInfo *) lfirst(lc2);
+
+
+ /* restore the path (we replace it in the loop) */
+ path = path_save;
- if (parse->hasAggs)
- add_partial_path(partially_grouped_rel, (Path *)
+ path = try_presorting(root, partially_grouped_rel, path,
+ cheapest_partial_path, info);
+
+ if (path == NULL)
+ continue;
+
+ if (parse->hasAggs)
+ add_partial_path(partially_grouped_rel, (Path *)
create_agg_path(root,
partially_grouped_rel,
path,
partially_grouped_rel->reltarget,
parse->groupClause ? AGG_SORTED : AGG_PLAIN,
AGGSPLIT_INITIAL_SERIAL,
- root->processed_groupClause,
+ info->clauses,
NIL,
agg_partial_costs,
dNumPartialPartialGroups));
- else
- add_partial_path(partially_grouped_rel, (Path *)
+ else
+ add_partial_path(partially_grouped_rel, (Path *)
create_group_path(root,
partially_grouped_rel,
path,
- root->processed_groupClause,
+ info->clauses,
NIL,
dNumPartialPartialGroups));
+ }
}
}
@@ -7354,6 +7432,8 @@ gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
* We can also skip the entire loop when we only have a single-item
* group_pathkeys because then we can't possibly have a presorted prefix
* of the list without having the list be fully sorted.
+ *
+ * XXX Shouldn't this also consider the group-key-reordering?
*/
if (!enable_incremental_sort || list_length(root->group_pathkeys) == 1)
return;
diff --git a/src/backend/utils/misc/guc_tables.c b/src/backend/utils/misc/guc_tables.c
index e53ebc6dc2..7fe58518d7 100644
--- a/src/backend/utils/misc/guc_tables.c
+++ b/src/backend/utils/misc/guc_tables.c
@@ -1050,6 +1050,16 @@ struct config_bool ConfigureNamesBool[] =
true,
NULL, NULL, NULL
},
+ {
+ {"enable_group_by_reordering", PGC_USERSET, QUERY_TUNING_METHOD,
+ gettext_noop("Enables reordering of GROUP BY keys."),
+ NULL,
+ GUC_EXPLAIN
+ },
+ &enable_group_by_reordering,
+ true,
+ NULL, NULL, NULL
+ },
{
{"geqo", PGC_USERSET, QUERY_TUNING_GEQO,
gettext_noop("Enables genetic query optimization."),
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index b2809c711a..dc82fceb05 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -399,6 +399,7 @@
#enable_seqscan = on
#enable_sort = on
#enable_tidscan = on
+#enable_group_by_reordering = on
# - Planner Cost Constants -
diff --git a/src/include/nodes/pathnodes.h b/src/include/nodes/pathnodes.h
index b9713ec9aa..0a251984c1 100644
--- a/src/include/nodes/pathnodes.h
+++ b/src/include/nodes/pathnodes.h
@@ -1456,6 +1456,16 @@ typedef struct PathKey
bool pk_nulls_first; /* do NULLs come before normal values? */
} PathKey;
+/*
+ * Combines information about pathkeys and the associated clauses.
+ */
+typedef struct PathKeyInfo
+{
+ NodeTag type;
+ List *pathkeys;
+ List *clauses;
+} PathKeyInfo;
+
/*
* VolatileFunctionStatus -- allows nodes to cache their
* contain_volatile_functions properties. VOLATILITY_UNKNOWN means not yet
diff --git a/src/include/optimizer/paths.h b/src/include/optimizer/paths.h
index efd4abc28f..e80156c49f 100644
--- a/src/include/optimizer/paths.h
+++ b/src/include/optimizer/paths.h
@@ -24,6 +24,7 @@ extern PGDLLIMPORT bool enable_geqo;
extern PGDLLIMPORT int geqo_threshold;
extern PGDLLIMPORT int min_parallel_table_scan_size;
extern PGDLLIMPORT int min_parallel_index_scan_size;
+extern PGDLLIMPORT bool enable_group_by_reordering;
/* Hook for plugins to get control in set_rel_pathlist() */
typedef void (*set_rel_pathlist_hook_type) (PlannerInfo *root,
@@ -204,6 +205,8 @@ typedef enum
extern PathKeysComparison compare_pathkeys(List *keys1, List *keys2);
extern bool pathkeys_contained_in(List *keys1, List *keys2);
extern bool pathkeys_count_contained_in(List *keys1, List *keys2, int *n_common);
+extern List *get_useful_group_keys_orderings(PlannerInfo *root, double nrows,
+ List *path_pathkeys);
extern Path *get_cheapest_path_for_pathkeys(List *paths, List *pathkeys,
Relids required_outer,
CostSelector cost_criterion,
diff --git a/src/test/regress/expected/aggregates.out b/src/test/regress/expected/aggregates.out
index f635c5a1af..423c8ec3b6 100644
--- a/src/test/regress/expected/aggregates.out
+++ b/src/test/regress/expected/aggregates.out
@@ -2728,6 +2728,138 @@ SELECT balk(hundred) FROM tenk1;
(1 row)
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+CREATE TABLE btg AS SELECT
+ i % 100 AS x,
+ i % 100 AS y,
+ 'abc' || i % 10 AS z,
+ i AS w
+FROM generate_series(1,10000) AS i;
+CREATE INDEX abc ON btg(x,y);
+ANALYZE btg;
+-- GROUP BY optimization by reorder columns by frequency
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+-- Utilize index scan ordering to avoid a Sort operation
+EXPLAIN (COSTS OFF) SELECT count(*) FROM btg GROUP BY x,y;
+ QUERY PLAN
+----------------------------------------
+ GroupAggregate
+ Group Key: x, y
+ -> Index Only Scan using abc on btg
+(3 rows)
+
+EXPLAIN (COSTS OFF) SELECT count(*) FROM btg GROUP BY y,x;
+ QUERY PLAN
+----------------------------------------
+ GroupAggregate
+ Group Key: x, y
+ -> Index Only Scan using abc on btg
+(3 rows)
+
+-- Engage incremental sort
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY x,y,z,w;
+ QUERY PLAN
+-----------------------------------------
+ Group
+ Group Key: x, y, z, w
+ -> Incremental Sort
+ Sort Key: x, y, z, w
+ Presorted Key: x, y
+ -> Index Scan using abc on btg
+(6 rows)
+
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY z,y,w,x;
+ QUERY PLAN
+-----------------------------------------
+ Group
+ Group Key: x, y, z, w
+ -> Incremental Sort
+ Sort Key: x, y, z, w
+ Presorted Key: x, y
+ -> Index Scan using abc on btg
+(6 rows)
+
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY w,z,x,y;
+ QUERY PLAN
+-----------------------------------------
+ Group
+ Group Key: x, y, w, z
+ -> Incremental Sort
+ Sort Key: x, y, w, z
+ Presorted Key: x, y
+ -> Index Scan using abc on btg
+(6 rows)
+
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY w,x,z,y;
+ QUERY PLAN
+-----------------------------------------
+ Group
+ Group Key: x, y, w, z
+ -> Incremental Sort
+ Sort Key: x, y, w, z
+ Presorted Key: x, y
+ -> Index Scan using abc on btg
+(6 rows)
+
+-- Subqueries
+explain (COSTS OFF) SELECT x,y
+FROM (SELECT * FROM btg ORDER BY x,y,w,z) AS q1
+GROUP BY (w,x,z,y);
+ QUERY PLAN
+----------------------------------------------
+ Group
+ Group Key: btg.x, btg.y, btg.w, btg.z
+ -> Incremental Sort
+ Sort Key: btg.x, btg.y, btg.w, btg.z
+ Presorted Key: btg.x, btg.y
+ -> Index Scan using abc on btg
+(6 rows)
+
+explain (COSTS OFF) SELECT x,y
+FROM (SELECT * FROM btg ORDER BY x,y,w,z LIMIT 100) AS q1
+GROUP BY (w,x,z,y);
+ QUERY PLAN
+----------------------------------------------------
+ Group
+ Group Key: btg.x, btg.y, btg.w, btg.z
+ -> Limit
+ -> Incremental Sort
+ Sort Key: btg.x, btg.y, btg.w, btg.z
+ Presorted Key: btg.x, btg.y
+ -> Index Scan using abc on btg
+(7 rows)
+
+-- Should work with and without GROUP-BY optimization
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY w,x,z,y ORDER BY y,x,z,w;
+ QUERY PLAN
+------------------------------
+ Group
+ Group Key: y, x, z, w
+ -> Sort
+ Sort Key: y, x, z, w
+ -> Seq Scan on btg
+(5 rows)
+
+-- Utilize incremental sort to make the ORDER BY rule a bit cheaper
+explain (COSTS OFF) SELECT x,w FROM btg GROUP BY w,x,y,z ORDER BY x*x,z;
+ QUERY PLAN
+-----------------------------------------------
+ Sort
+ Sort Key: ((x * x)), z
+ -> Group
+ Group Key: x, y, w, z
+ -> Incremental Sort
+ Sort Key: x, y, w, z
+ Presorted Key: x, y
+ -> Index Scan using abc on btg
+(8 rows)
+
+DROP TABLE btg;
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
diff --git a/src/test/regress/expected/sysviews.out b/src/test/regress/expected/sysviews.out
index 271313ebf8..9be7aca2b8 100644
--- a/src/test/regress/expected/sysviews.out
+++ b/src/test/regress/expected/sysviews.out
@@ -114,6 +114,7 @@ select name, setting from pg_settings where name like 'enable%';
enable_async_append | on
enable_bitmapscan | on
enable_gathermerge | on
+ enable_group_by_reordering | on
enable_hashagg | on
enable_hashjoin | on
enable_incremental_sort | on
@@ -133,7 +134,7 @@ select name, setting from pg_settings where name like 'enable%';
enable_seqscan | on
enable_sort | on
enable_tidscan | on
-(22 rows)
+(23 rows)
-- There are always wait event descriptions for various types.
select type, count(*) > 0 as ok FROM pg_wait_events
diff --git a/src/test/regress/sql/aggregates.sql b/src/test/regress/sql/aggregates.sql
index cc8f0efad5..b9fcceedd7 100644
--- a/src/test/regress/sql/aggregates.sql
+++ b/src/test/regress/sql/aggregates.sql
@@ -1181,6 +1181,53 @@ SELECT balk(hundred) FROM tenk1;
ROLLBACK;
+-- GROUP BY optimization by reorder columns
+
+CREATE TABLE btg AS SELECT
+ i % 100 AS x,
+ i % 100 AS y,
+ 'abc' || i % 10 AS z,
+ i AS w
+FROM generate_series(1,10000) AS i;
+CREATE INDEX abc ON btg(x,y);
+ANALYZE btg;
+
+-- GROUP BY optimization by reorder columns by frequency
+
+SET enable_hashagg=off;
+SET max_parallel_workers= 0;
+SET max_parallel_workers_per_gather = 0;
+
+-- Utilize index scan ordering to avoid a Sort operation
+EXPLAIN (COSTS OFF) SELECT count(*) FROM btg GROUP BY x,y;
+EXPLAIN (COSTS OFF) SELECT count(*) FROM btg GROUP BY y,x;
+
+-- Engage incremental sort
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY x,y,z,w;
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY z,y,w,x;
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY w,z,x,y;
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY w,x,z,y;
+
+-- Subqueries
+explain (COSTS OFF) SELECT x,y
+FROM (SELECT * FROM btg ORDER BY x,y,w,z) AS q1
+GROUP BY (w,x,z,y);
+explain (COSTS OFF) SELECT x,y
+FROM (SELECT * FROM btg ORDER BY x,y,w,z LIMIT 100) AS q1
+GROUP BY (w,x,z,y);
+
+-- Should work with and without GROUP-BY optimization
+explain (COSTS OFF) SELECT x,y FROM btg GROUP BY w,x,z,y ORDER BY y,x,z,w;
+
+-- Utilize incremental sort to make the ORDER BY rule a bit cheaper
+explain (COSTS OFF) SELECT x,w FROM btg GROUP BY w,x,y,z ORDER BY x*x,z;
+
+DROP TABLE btg;
+
+RESET enable_hashagg;
+RESET max_parallel_workers;
+RESET max_parallel_workers_per_gather;
+
-- Secondly test the case of a parallel aggregate combiner function
-- returning NULL. For that use normal transition function, but a
-- combiner function returning NULL.
--
2.43.0
^ permalink raw reply [nested|flat] 38+ messages in thread
end of thread, other threads:[~2024-01-09 09:01 UTC | newest]
Thread overview: 38+ messages (download: mbox mbox.gz follow: Atom feed)
-- links below jump to the message on this page --
2022-01-21 11:09 Re: POC: GROUP BY optimization Andrey Lepikhov <[email protected]>
2022-01-21 20:34 ` Tomas Vondra <[email protected]>
2022-01-24 05:02 ` Andrey Lepikhov <[email protected]>
2022-02-10 09:00 ` Andrey V. Lepikhov <[email protected]>
2022-03-15 08:26 ` Tomas Vondra <[email protected]>
2022-03-18 07:06 ` Andrey V. Lepikhov <[email protected]>
2022-03-29 00:48 ` Tomas Vondra <[email protected]>
2022-03-29 02:02 ` Zhihong Yu <[email protected]>
2022-03-29 18:00 ` Tomas Vondra <[email protected]>
2022-03-30 23:19 ` Tomas Vondra <[email protected]>
2022-07-13 03:37 ` David Rowley <[email protected]>
2022-09-19 21:16 ` David Rowley <[email protected]>
2023-07-20 06:37 ` Andrey Lepikhov <[email protected]>
2023-07-20 11:46 ` Tomas Vondra <[email protected]>
2023-07-21 04:21 ` Andrey Lepikhov <[email protected]>
2023-07-24 02:10 ` Andrey Lepikhov <[email protected]>
2023-07-24 09:56 ` Tomas Vondra <[email protected]>
2023-07-24 12:04 ` Andrey Lepikhov <[email protected]>
2023-09-19 04:42 ` Andrey Lepikhov <[email protected]>
2023-09-26 05:51 ` Andrey Lepikhov <[email protected]>
2023-10-01 06:32 ` Andrei Lepikhov <[email protected]>
2023-12-21 10:53 ` Alexander Korotkov <[email protected]>
2023-12-26 11:37 ` Andrei Lepikhov <[email protected]>
2024-01-09 09:01 ` Andrei Lepikhov <[email protected]>
2022-07-15 05:18 ` David Rowley <[email protected]>
2022-07-15 19:46 ` Tomas Vondra <[email protected]>
2022-08-02 10:21 ` Michael Paquier <[email protected]>
2022-08-02 11:14 ` David Rowley <[email protected]>
2022-08-17 14:46 ` Tomas Vondra <[email protected]>
2022-08-18 01:32 ` David Rowley <[email protected]>
2022-08-18 07:29 ` Tomas Vondra <[email protected]>
2022-08-30 13:15 ` Jonathan S. Katz <[email protected]>
2022-09-01 13:06 ` Tomas Vondra <[email protected]>
2022-09-01 14:05 ` Jonathan S. Katz <[email protected]>
2022-09-04 22:14 ` Tomas Vondra <[email protected]>
2022-09-05 00:50 ` Jonathan S. Katz <[email protected]>
2022-09-05 06:38 ` Michael Paquier <[email protected]>
2023-07-25 18:18 [PATCH v10 4/4] Remove open-coded binary heap in pg_dump_sort.c. Nathan Bossart <[email protected]>
This inbox is served by agora; see mirroring instructions
for how to clone and mirror all data and code used for this inbox