public inbox for [email protected]
help / color / mirror / Atom feedRe: Hash-based MCV matching for large IN-lists
7+ messages / 2 participants
[nested] [flat]
* Re: Hash-based MCV matching for large IN-lists
@ 2026-02-25 22:45 Ilia Evdokimov <[email protected]>
2026-02-26 08:57 ` Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
0 siblings, 1 reply; 7+ messages in thread
From: Ilia Evdokimov @ 2026-02-25 22:45 UTC (permalink / raw)
To: Zsolt Parragi <[email protected]>; +Cc: Tatsuya Kawata <[email protected]>; David Geier <[email protected]>; Chengpeng Yan <[email protected]>; [email protected] <[email protected]>
I've addressed the review comments mentioned above.
David made a very good observation: for unique columns, where each
iteration effectively returns the same per-element selectivity, there is
no need to iterate at all. In such cases we can reduce the computation
to a closed-form expression, i.e. O(1) instead of running the loop O(N).
I applied this idea to unique columns and cases falling back to
DEFAULT_EQ_SEL. In both cases the loop can be replaced with a
closed-from formula implemented in calculate_combined_selectivity(). The
formula mirros the existing independent/disjoint probability model: ANY
(sel = 1 - (1 - s) ^ length or length * s ), ALL (sel = s ^ length or 1
- length*(1 - s)). It would be good to carefully review that this is
fully equivalent to the current accumulation logic.
I also exprimented with applying the same idea to elements that are not
found in MCV, are not Const, and effectively found in MCV with more than
one count. Those cases can still be accumulated using
accum_scalararray_prob(), but potentially grouped to reduce repeated work.
Overall, the optimization work can be logically split into three parts:
1. Degenerate NULL case O(N) -> O(1) [0]
2. Identical non-NULL per-element selectivity O(N) -> O(1) (can be
split into a separate thread if prederred)
3. MCV matching via hashing O(M*N) -> O(M+N) (current thread)
Feedback on how to best structure or split this work would be appreciated.
About op_is_reserved. It seems we should assign op_is_reserved = true,
because we don't reverse types like eqjoinsel_semi(). If IN-list smaller
than MCV-list we reverse it by fmgr_info(hash_mcv ? hashLeft :
hashRight, &hash_proc). Thanks for this remark.
Thoughts?
--
Best regards,
Ilia Evdokimov,
Tantor Labs LLC,
https://tantorlabs.com/
Attachments:
[text/x-patch] v6-0003-Use-hash-based-MCV-matching-for-ScalarArrayOpExpr.patch (17.5K, 3-v6-0003-Use-hash-based-MCV-matching-for-ScalarArrayOpExpr.patch)
download | inline diff:
From 606d29dbb3e316796df3a581811ca6c98ac5b3a6 Mon Sep 17 00:00:00 2001
From: Evdokimov Ilia <[email protected]>
Date: Wed, 25 Feb 2026 23:34:37 +0300
Subject: [PATCH v6 3/3] Use hash-based MCV matching for ScalarArrayOpExpr
selectivity
When estimating selectivity for ScalarArrayOpExpr (IN / ANY / ALL) with
available MCV statistics, the planner currently matches IN-list elements
against the MCV array using nested loops. For large IN-lists and/or large
MCV lists this leads to O(N*M) planning-time behavior.
This patch adds a hash-based matching strategy, similar to the one used
in join selectivity estimation. When MCV statistics are available and the
operator supports hashing, the smaller of the two inputs (MCV list or
IN-list constant elements) is chosen as the hash table build side, and
the other side is scanned once, reducing complexity to O(N+M).
The hash-based path is restricted to equality and inequality operators
that use eqsel()/neqsel(), and is applied only when suitable hash
functions and MCV statistics are available.
---
src/backend/utils/adt/selfuncs.c | 450 ++++++++++++++++++++++++++++++-
1 file changed, 440 insertions(+), 10 deletions(-)
diff --git a/src/backend/utils/adt/selfuncs.c b/src/backend/utils/adt/selfuncs.c
index f6091a576d8..b6a9cf7844f 100644
--- a/src/backend/utils/adt/selfuncs.c
+++ b/src/backend/utils/adt/selfuncs.c
@@ -146,23 +146,27 @@
/*
* In production builds, switch to hash-based MCV matching when the lists are
* large enough to amortize hash setup cost. (This threshold is compared to
- * the sum of the lengths of the two MCV lists. This is simplistic but seems
+ * the sum of the lengths of the two lists. This is simplistic but seems
* to work well enough.) In debug builds, we use a smaller threshold so that
* the regression tests cover both paths well.
*/
#ifndef USE_ASSERT_CHECKING
-#define EQJOINSEL_MCV_HASH_THRESHOLD 200
+#define MCV_HASH_THRESHOLD 200
#else
-#define EQJOINSEL_MCV_HASH_THRESHOLD 20
+#define MCV_HASH_THRESHOLD 20
#endif
-/* Entries in the simplehash hash table used by eqjoinsel_find_matches */
+/*
+ * Entries in the simplehash hash table used by
+ * eqjoinsel_find_matches and scalararray_mcv_hash_match
+ */
typedef struct MCVHashEntry
{
Datum value; /* the value represented by this entry */
int index; /* its index in the relevant AttStatsSlot */
uint32 hash; /* hash code for the Datum */
char status; /* status code used by simplehash.h */
+ int count; /* number of occurrences of current value in */
} MCVHashEntry;
/* private_data for the simplehash hash table */
@@ -184,6 +188,15 @@ get_relation_stats_hook_type get_relation_stats_hook = NULL;
get_index_stats_hook_type get_index_stats_hook = NULL;
static double eqsel_internal(PG_FUNCTION_ARGS, bool negate);
+static double scalararray_mcv_hash_match(VariableStatData *vardata, Oid operator, Oid collation,
+ Selectivity s2, Datum *elem_values,
+ bool *elem_nulls, int num_elems, bool *elem_const,
+ Oid nominal_element_type, bool useOr, bool isEquality,
+ bool isInequality);
+static void accum_scalararray_prob(double individual_s, int count,
+ bool useOr, bool isEquality,
+ bool isInequality, double nullfrac,
+ double *selec, double *s1disjoint);
static Selectivity calculate_combined_selectivity(Selectivity s2, int num_elems,
bool useOr,
bool isEquality, bool isInequality);
@@ -1951,6 +1964,36 @@ calculate_combined_selectivity(Selectivity s2, int num_elems, bool useOr, bool i
return s1;
}
+/*
+ * Accumulate the selectivity contribution of a single array element
+ * into the running ScalarArrayOpExpr selectivity estimate.
+ */
+static void
+accum_scalararray_prob(double individual_s, int count, bool useOr, bool isEquality,
+ bool isInequality, double nullfrac, double *selec, double *s1disjoint)
+{
+ if (count <= 0)
+ return;
+
+ if (isInequality)
+ individual_s = 1.0 - individual_s - nullfrac;
+
+ CLAMP_PROBABILITY(individual_s);
+
+ if (useOr)
+ {
+ *selec = 1.0 - (1.0 - *selec) * pow(1.0 - individual_s, count);
+ if (isEquality)
+ *s1disjoint += individual_s * count;
+ }
+ else
+ {
+ *selec = (*selec) * pow(individual_s, count);
+ if (isInequality)
+ *s1disjoint += count * (individual_s - 1.0);
+ }
+}
+
/*
* scalararraysel - Selectivity of ScalarArrayOpExpr Node.
*/
@@ -2102,6 +2145,7 @@ scalararraysel(PlannerInfo *root,
Selectivity s2 = -1.0;
Node *other_op = NULL;
bool var_on_left;
+ bool *elem_const = NULL;
/*
* If the clause is of the form "var OP something" or
@@ -2134,16 +2178,25 @@ scalararraysel(PlannerInfo *root,
s2 = 1.0 - DEFAULT_EQ_SEL;
}
- ReleaseVariableStats(vardata);
-
if (s2 >= 0.0)
{
CLAMP_PROBABILITY(s2);
s1 = calculate_combined_selectivity(s2, num_elems, useOr, isEquality, isInequality);
+ ReleaseVariableStats(vardata);
+
return s1;
}
+
+ s1 = scalararray_mcv_hash_match(&vardata, operator, clause->inputcollid, s2,
+ elem_values, elem_nulls, num_elems, elem_const,
+ nominal_element_type, useOr, isEquality, isInequality);
+
+ ReleaseVariableStats(vardata);
+
+ if (s1 >= 0.0)
+ return s1;
}
else
{
@@ -2250,16 +2303,78 @@ scalararraysel(PlannerInfo *root,
* variable, then fall back to default code path to compute
* default selectivity.
*/
- if (!get_restriction_variable(root, clause->args, varRelid,
+ if (get_restriction_variable(root, clause->args, varRelid,
&vardata, &other_op, &var_on_left))
+ {
+ Datum *elem_values;
+ bool *elem_nulls;
+ bool *elem_const;
+ ListCell *lc;
+
+ elem_values = palloc_array(Datum, num_elems);
+ elem_nulls = palloc0_array(bool, num_elems);
+ elem_const = palloc0_array(bool, num_elems);
+
+ /*
+ * Build arrays describing ARRAY[] elements:
+ * - elem_values: Datum value for Const elements
+ * - elem_nulls: whether element is NULL
+ * - elem_const: whether element is a Const node
+ */
+ foreach(lc, arrayexpr->elements)
+ {
+ Node *elem_value = (Node *) lfirst(lc);
+ int i = foreach_current_index(lc);
+
+ if (IsA(elem_value, Const))
+ {
+ elem_values[i] = ((Const *) elem_value)->constvalue;
+ elem_nulls[i] = ((Const *) elem_value)->constisnull;
+ elem_const[i] = true;
+ }
+ else
+ {
+ elem_nulls[i] = false;
+ elem_const[i] = false;
+ }
+
+ /* Selectivity of "WHERE x NOT IN (NULL, ... )" is always 0 */
+ if (!useOr && elem_nulls[i])
+ {
+ pfree(elem_values);
+ pfree(elem_nulls);
+ pfree(elem_const);
+
+ ReleaseVariableStats(vardata);
+
+ return (Selectivity) 0.0;
+ }
+ }
+
+ /* Compute per-element selectivity via eqsel()/neqsel semantics. */
+ s2 = var_eq_non_const(&vardata, operator, clause->inputcollid,
+ other_op, var_on_left, isInequality);
+
+ s1 = scalararray_mcv_hash_match(&vardata, operator, clause->inputcollid, s2,
+ elem_values, elem_nulls, num_elems, elem_const,
+ nominal_element_type, useOr, isEquality, isInequality);
+
+ pfree(elem_values);
+ pfree(elem_nulls);
+ pfree(elem_const);
+
+ ReleaseVariableStats(vardata);
+
+ if (s1 >= 0.0)
+ return s1;
+ }
+ else
{
s2 = (isInequality) ? (1.0 - DEFAULT_EQ_SEL) : DEFAULT_EQ_SEL;
s1 = calculate_combined_selectivity(s2, num_elems, useOr, isEquality, isInequality);
return s1;
}
- else
- ReleaseVariableStats(vardata);
}
/*
@@ -2376,6 +2491,321 @@ scalararraysel(PlannerInfo *root,
return s1;
}
+/*
+ * Estimate selectivity of a ScalarArrayOpExpr (ANY/ALL) using MCV statistics
+ * with hash-based matching.
+ *
+ * This function follows the same probability model as the generic
+ * ScalarArrayOpExpr selectivity code (independent or disjoint probabilities
+ * for OR/AND combinations), but attempts to speed up matching between
+ * IN-list elements and the column's most-common-values (MCV) statistics by
+ * using hashing instead of nested loops.
+ *
+ * MCV statistics are used only to obtain per-value selectivities for
+ * constants that match MCV entries. All probabilities are combined using
+ * the standard ANY/ALL formulas, exactly as in the generic estimator.
+ *
+ * The function may return -1.0 to indicate that hash-based MCV estimation
+ * is not applicable (for example, missing statistics, unsupported operator,
+ * or unavailable hash functions), in which case the caller should fall back
+ * to the generic ScalarArrayOpExpr selectivity estimation.
+ *
+ * Inputs:
+ * vardata: statistics and metadata for the variable being estimated
+ * operator: equality or inequality operator to apply
+ * collation: OID of collation to use
+ * nonconst_sel: selectivity of non-const element
+ * elem_values: array of IN-list element values
+ * elem_nulls: array indicating which IN-list elements are NULL
+ * elem_const: array indicating which IN-list elements are Const nodes.
+ * array is NULL if all elemnets are const.
+ * num_elems: number of IN-list elements
+ * nominal_element_type: type of IN-list elements
+ * useOr: true if elements are combined using OR semantics, false for AND
+ * isEquality: true if the operator behaves like equality
+ * isInequality: true if the operator behaves like inequality
+ *
+ * Result:
+ * Selectivity estimate in the range [0.0, 1.0], or -1.0 if no estimate
+ * could be produced by this function.
+ *
+ * Note:
+ * This function assumes that the operator’s selectivity behavior matches
+ * eqsel()/neqsel semantics. It must not be used for operators with custom
+ * or non-standard selectivity behavior.
+ */
+static double
+scalararray_mcv_hash_match(VariableStatData *vardata, Oid operator, Oid collation, Selectivity nonconst_sel,
+ Datum *elem_values, bool *elem_nulls, int num_elems, bool *elem_const,
+ Oid nominal_element_type, bool useOr, bool isEquality,
+ bool isInequality)
+{
+ Form_pg_statistic stats;
+ AttStatsSlot sslot;
+ FmgrInfo eqproc;
+ double selec = -1.0,
+ s1disjoint,
+ nullfrac = 0.0;
+ Oid hashLeft = InvalidOid,
+ hashRight = InvalidOid,
+ opfuncoid;
+ bool have_mcvs = false;
+
+ opfuncoid = get_opcode(operator);
+ memset(&sslot, 0, sizeof(sslot));
+
+ if (HeapTupleIsValid(vardata->statsTuple))
+ {
+ if (statistic_proc_security_check(vardata, opfuncoid))
+ have_mcvs = get_attstatsslot(&sslot, vardata->statsTuple,
+ STATISTIC_KIND_MCV, InvalidOid,
+ ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
+ }
+
+ if (have_mcvs)
+ {
+ /*
+ * If the MCV list and IN-list are large enough, and the operator
+ * supports hashing, attempt to use hash functions so that MCV–IN
+ * matching can be done in O(N+M) instead of O(N×M).
+ */
+ if (sslot.nvalues + num_elems >= MCV_HASH_THRESHOLD)
+ {
+ fmgr_info(opfuncoid, &eqproc);
+ (void) get_op_hash_functions(operator, &hashLeft, &hashRight);
+ }
+ }
+
+ if (have_mcvs && OidIsValid(hashLeft) && OidIsValid(hashRight))
+ {
+ /* Use a hash table to speed up the matching */
+ LOCAL_FCINFO(fcinfo, 2);
+ LOCAL_FCINFO(hash_fcinfo, 1);
+ MCVHashTable_hash *hashTable;
+ FmgrInfo hash_proc;
+ MCVHashContext hashContext;
+ double sumallcommon = 0.0,
+ nonmcv_selec = 0.0;
+ bool isdefault;
+ bool hash_mcv;
+ double otherdistinct;
+ Datum *arrayHash;
+ Datum *arrayProbe;
+ int nvaluesHash;
+ int nvaluesProbe;
+ int nonmcv_cnt = num_elems;
+ int nonconst_cnt = 0;
+
+ /* Grab the nullfrac for use below. */
+ stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
+ nullfrac = stats->stanullfrac;
+
+ selec = s1disjoint = (useOr ? 0.0 : 1.0);
+
+ InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
+ NULL, NULL);
+ fcinfo->args[0].isnull = false;
+ fcinfo->args[1].isnull = false;
+
+ for (int i = 0; i < sslot.nvalues; i++)
+ sumallcommon += sslot.numbers[i];
+
+ /*
+ * Compute the total probability mass of all non-MCV values. This is
+ * the part of the column distribution not covered by MCVs.
+ */
+ nonmcv_selec = 1.0 - sumallcommon - nullfrac;
+ CLAMP_PROBABILITY(nonmcv_selec);
+
+ /*
+ * Approximate the per-value probability of a non-MCV constant by
+ * dividing the remaining probability mass by the number of other
+ * distinct values.
+ */
+ otherdistinct = get_variable_numdistinct(vardata, &isdefault) - sslot.nnumbers;
+ if (otherdistinct > 1)
+ nonmcv_selec /= otherdistinct;
+
+ if (sslot.nnumbers > 0 && nonmcv_selec > sslot.numbers[sslot.nnumbers - 1])
+ nonmcv_selec = sslot.numbers[sslot.nnumbers - 1];
+
+ /* Make sure we build the hash table on the smaller array. */
+ if (sslot.nvalues <= num_elems)
+ {
+ hash_mcv = true;
+ nvaluesHash = sslot.nvalues;
+ nvaluesProbe = num_elems;
+ arrayHash = sslot.values;
+ arrayProbe = elem_values;
+ }
+ else
+ {
+ hash_mcv = false;
+ nvaluesHash = num_elems;
+ nvaluesProbe = sslot.nvalues;
+ arrayHash = elem_values;
+ arrayProbe = sslot.values;
+ }
+
+ fmgr_info(hash_mcv ? hashLeft : hashRight, &hash_proc);
+ InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
+ NULL, NULL);
+ hash_fcinfo->args[0].isnull = false;
+
+ hashContext.equal_fcinfo = fcinfo;
+ hashContext.hash_fcinfo = hash_fcinfo;
+ hashContext.op_is_reversed = true;
+ hashContext.insert_mode = true;
+
+ get_typlenbyval(hash_mcv ? sslot.valuetype : nominal_element_type,
+ &hashContext.hash_typlen,
+ &hashContext.hash_typbyval);
+
+ hashTable = MCVHashTable_create(CurrentMemoryContext,
+ nvaluesHash,
+ &hashContext);
+
+ /* Build a hash table over the smaller input side. */
+ for (int i = 0; i < nvaluesHash; i++)
+ {
+ bool found = false;
+ MCVHashEntry *entry;
+
+ /*
+ * When hashing IN-list values (hash_mcv == false), we only insert
+ * constant, non-NULL elements. NULL and non-Const elements are
+ * counted separately, because they cannot participate in MCV
+ * matching and must be handled later using generic selectivity
+ * estimation.
+ */
+ if (!hash_mcv)
+ {
+ if (elem_nulls[i])
+ {
+ nonmcv_cnt--;
+ continue;
+ }
+
+ if (elem_const != NULL && !elem_const[i])
+ {
+ nonmcv_cnt--;
+ nonconst_cnt++;
+ continue;
+ }
+ }
+
+ entry = MCVHashTable_insert(hashTable, arrayHash[i], &found);
+
+ /*
+ * entry->count tracks how many times the same value appears, so
+ * that duplicate IN-list elements can be folded into the
+ * probability calculation.
+ */
+ if (likely(!found))
+ {
+ entry->index = i;
+ entry->count = 1;
+ }
+ else
+ entry->count++;
+ }
+
+ hashContext.insert_mode = false;
+ if (hashLeft != hashRight)
+ {
+ fmgr_info(hash_mcv ? hashRight : hashLeft, &hash_proc);
+ /* Resetting hash_fcinfo is probably unnecessary, but be safe */
+ InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
+ NULL, NULL);
+ hash_fcinfo->args[0].isnull = false;
+ }
+
+ for (int i = 0; i < nvaluesProbe; i++)
+ {
+ MCVHashEntry *entry;
+ Selectivity s1;
+ int nvaluesmcv;
+
+ /*
+ * When probing with IN-list elements, ignore NULLs and non-Const
+ * expressions: they cannot be matched against MCVs and will be
+ * accounted for later by generic estimation.
+ */
+ if (hash_mcv)
+ {
+ if (elem_nulls[i])
+ {
+ nonmcv_cnt--;
+ continue;
+ }
+
+ if (elem_const != NULL && !elem_const[i])
+ {
+ nonmcv_cnt--;
+ nonconst_cnt++;
+ continue;
+ }
+ }
+
+ entry = MCVHashTable_lookup(hashTable, arrayProbe[i]);
+
+ /*
+ * If found, obtain its MCV frequency and remember how many values
+ * on the hashed side map to this entry.
+ */
+ if (entry != NULL)
+ {
+ s1 = hash_mcv ? sslot.numbers[entry->index]
+ : sslot.numbers[i];
+
+ nvaluesmcv = entry->count;
+
+ accum_scalararray_prob(s1, nvaluesmcv, useOr, isEquality, isInequality,
+ nullfrac, &selec, &s1disjoint);
+
+ /* Matched values are no longer considered non-MCV */
+ nonmcv_cnt -= nvaluesmcv;
+ }
+ }
+
+ /*
+ * Account for constant IN-list values that did not match any MCV.
+ *
+ * Each such value is assumed to have probability = nonmcv_selec,
+ * derived from the remaining (non-MCV) probability mass.
+ */
+ accum_scalararray_prob(nonmcv_selec, nonmcv_cnt, useOr, isEquality, isInequality,
+ nullfrac, &selec, &s1disjoint);
+
+ /*
+ * Account for non-Const IN-list elements.
+ *
+ * These values cannot be matched against MCVs, so we rely on the
+ * operator's generic selectivity estimator for each of them.
+ */
+ accum_scalararray_prob(nonconst_sel, nonconst_cnt, useOr, isEquality, isInequality,
+ nullfrac, &selec, &s1disjoint);
+
+ /*
+ * For = ANY or <> ALL, if the IN-list elements are assumed distinct,
+ * the events are disjoint and the total probability is the sum of
+ * individual probabilities. Use that estimate if it lies in [0,1].
+ */
+ if ((useOr ? isEquality : isInequality) &&
+ s1disjoint >= 0.0 && s1disjoint <= 1.0)
+ selec = s1disjoint;
+
+ CLAMP_PROBABILITY(selec);
+
+ MCVHashTable_destroy(hashTable);
+ }
+
+ if (have_mcvs)
+ free_attstatsslot(&sslot);
+
+ return selec;
+}
+
/*
* Estimate number of elements in the array yielded by an expression.
*
@@ -2612,7 +3042,7 @@ eqjoinsel(PG_FUNCTION_ARGS)
* If the MCV lists are long enough to justify hashing, try to look up
* hash functions for the join operator.
*/
- if ((sslot1.nvalues + sslot2.nvalues) >= EQJOINSEL_MCV_HASH_THRESHOLD)
+ if ((sslot1.nvalues + sslot2.nvalues) >= MCV_HASH_THRESHOLD)
(void) get_op_hash_functions(operator, &hashLeft, &hashRight);
}
else
--
2.34.1
[text/x-patch] v6-0002-Use-O-1-selectivity-formula-for-eqsel-neqsel-IN-A.patch (6.4K, 4-v6-0002-Use-O-1-selectivity-formula-for-eqsel-neqsel-IN-A.patch)
download | inline diff:
From 7d668752ceb49b901571a96d156e0219da4e7c1f Mon Sep 17 00:00:00 2001
From: Evdokimov Ilia <[email protected]>
Date: Wed, 25 Feb 2026 23:00:32 +0300
Subject: [PATCH v6 2/3] Use O(1) selectivity formula for eqsel/neqsel IN/ALL
Replace per-element iteration in ScalarArrayOpExpr selectivity
estimation with a closed-form probability formula when all elements
share the same eqsel()/neqsel() semantics.
Preserves existing independence/disjoint models while reducing
planning cost for large IN/ALL lists from O(N) to O(1).
Special handling added for unique columns using 1/reltuples.
---
src/backend/utils/adt/selfuncs.c | 157 +++++++++++++++++++++++++++++++
1 file changed, 157 insertions(+)
diff --git a/src/backend/utils/adt/selfuncs.c b/src/backend/utils/adt/selfuncs.c
index eef3f0375a5..f6091a576d8 100644
--- a/src/backend/utils/adt/selfuncs.c
+++ b/src/backend/utils/adt/selfuncs.c
@@ -184,6 +184,9 @@ get_relation_stats_hook_type get_relation_stats_hook = NULL;
get_index_stats_hook_type get_index_stats_hook = NULL;
static double eqsel_internal(PG_FUNCTION_ARGS, bool negate);
+static Selectivity calculate_combined_selectivity(Selectivity s2, int num_elems,
+ bool useOr,
+ bool isEquality, bool isInequality);
static double eqjoinsel_inner(FmgrInfo *eqproc, Oid collation,
Oid hashLeft, Oid hashRight,
VariableStatData *vardata1, VariableStatData *vardata2,
@@ -1893,6 +1896,61 @@ strip_array_coercion(Node *node)
return node;
}
+/*
+ * calculate_combined_selectivity
+ *
+ * Combine selectivities of N identical ScalarArrayOpExpr elements.
+ *
+ * This function assumes that all elements of the IN/ANY or ALL list
+ * have the same per-element selectivity s2, and computes the overall
+ * selectivity without iterating over the elements.
+ *
+ * For OR semantics (x = ANY (...)):
+ * main model : 1 - (1 - s2)^N
+ * disjoint model : N * s2
+ *
+ * For AND semantics (x <> ALL (...)):
+ * main model : s2^N
+ * disjoint model : 1 - N * (1 - s2)
+ *
+ * If the disjoint estimate is within [0,1], it is preferred.
+ * Otherwise, we fall back to the main (independence) model.
+ */
+static Selectivity
+calculate_combined_selectivity(Selectivity s2, int num_elems, bool useOr, bool isEquality, bool isInequality)
+{
+ bool use_disjoint = false;
+ Selectivity s1;
+ Selectivity s1disjoint;
+
+ s1 = s1disjoint = (useOr ? 0.0 : 1.0);
+
+ if (useOr)
+ {
+ if (isEquality)
+ {
+ s1disjoint = s2 * num_elems;
+ if (s1disjoint >= 0.0 && s1disjoint <= 1.0)
+ use_disjoint = true;
+ }
+ s1 = use_disjoint ? s1disjoint : (1.0 - pow(1.0 - s2, num_elems));
+ }
+ else
+ {
+ if (isInequality)
+ {
+ s1disjoint = 1.0 + num_elems * (s2 - 1.0);
+ if (s1disjoint >= 0.0 && s1disjoint <= 1.0)
+ use_disjoint = true;
+ }
+ s1 = use_disjoint ? s1disjoint : pow(s2, num_elems);
+ }
+
+ CLAMP_PROBABILITY(s1);
+
+ return s1;
+}
+
/*
* scalararraysel - Selectivity of ScalarArrayOpExpr Node.
*/
@@ -2030,6 +2088,72 @@ scalararraysel(PlannerInfo *root,
elmlen, elmbyval, elmalign,
&elem_values, &elem_nulls, &num_elems);
+ /*
+ * Try to avoid O(N^2) selectivity calculation for ScalarArrayOpExpr.
+ *
+ * For equality/inequality operators in restriction clauses,
+ * attempt to derive a single per-element selectivity (s2) and
+ * combine it in O(1) time using a closed-form formula instead
+ * of iterating over all elements.
+ */
+ if ((isEquality || isInequality) && !is_join_clause)
+ {
+ VariableStatData vardata;
+ Selectivity s2 = -1.0;
+ Node *other_op = NULL;
+ bool var_on_left;
+
+ /*
+ * If the clause is of the form "var OP something" or
+ * "something OP var", extract statistics for the variable.
+ * Otherwise, fall back to a default per-element estimate.
+ */
+ if (get_restriction_variable(root, clause->args, varRelid, &vardata, &other_op, &var_on_left))
+ {
+ /*
+ * Fast path for unique columns.
+ *
+ * If the variable is known to be unique and the relation
+ * has at least one tuple, equality selectivity is exactly
+ * 1 / reltuples.
+ */
+ if (vardata.isunique && vardata.rel && vardata.rel->tuples >= 1.0)
+ {
+ s2 = 1.0 / vardata.rel->tuples;
+ if (HeapTupleIsValid(vardata.statsTuple))
+ {
+ Form_pg_statistic stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
+ if (isInequality)
+ s2 = 1.0 - s2 - stats->stanullfrac;
+ }
+ }
+ else if (isInequality)
+ {
+ Oid negator = get_negator(operator);
+ if (!OidIsValid(negator))
+ s2 = 1.0 - DEFAULT_EQ_SEL;
+ }
+
+ ReleaseVariableStats(vardata);
+
+ if (s2 >= 0.0)
+ {
+ CLAMP_PROBABILITY(s2);
+
+ s1 = calculate_combined_selectivity(s2, num_elems, useOr, isEquality, isInequality);
+
+ return s1;
+ }
+ }
+ else
+ {
+ s2 = (isInequality) ? (1.0 - DEFAULT_EQ_SEL) : DEFAULT_EQ_SEL;
+ s1 = calculate_combined_selectivity(s2, num_elems, useOr, isEquality, isInequality);
+
+ return s1;
+ }
+ }
+
/*
* For generic operators, we assume the probability of success is
* independent for each array element. But for "= ANY" or "<> ALL",
@@ -2105,6 +2229,39 @@ scalararraysel(PlannerInfo *root,
get_typlenbyval(arrayexpr->element_typeid,
&elmlen, &elmbyval);
+ /*
+ * Try to avoid O(N^2) selectivity calculation for ScalarArrayOpExpr.
+ *
+ * For equality/inequality operators in restriction clauses,
+ * attempt to derive a single per-element selectivity (s2) and
+ * combine it in O(1) time using a closed-form formula instead
+ * of iterating over all elements.
+ */
+ if ((isEquality || isInequality) && !is_join_clause)
+ {
+ VariableStatData vardata;
+ Selectivity s2 = -1.0;
+ Node *other_op = NULL;
+ bool var_on_left;
+ int num_elems = list_length(arrayexpr->elements);
+
+ /*
+ * If expression is not variable = something or something =
+ * variable, then fall back to default code path to compute
+ * default selectivity.
+ */
+ if (!get_restriction_variable(root, clause->args, varRelid,
+ &vardata, &other_op, &var_on_left))
+ {
+ s2 = (isInequality) ? (1.0 - DEFAULT_EQ_SEL) : DEFAULT_EQ_SEL;
+ s1 = calculate_combined_selectivity(s2, num_elems, useOr, isEquality, isInequality);
+
+ return s1;
+ }
+ else
+ ReleaseVariableStats(vardata);
+ }
+
/*
* We use the assumption of disjoint probabilities here too, although
* the odds of equal array elements are rather higher if the elements
--
2.34.1
[text/x-patch] v6-0001-Reduce-planning-time-for-large-NOT-IN-lists-conta.patch (4.9K, 5-v6-0001-Reduce-planning-time-for-large-NOT-IN-lists-conta.patch)
download | inline diff:
From c6cc307e6d7a131ba5fc3c59fc86ba0df6768a43 Mon Sep 17 00:00:00 2001
From: Evdokimov Ilia <[email protected]>
Date: Wed, 25 Feb 2026 22:58:32 +0300
Subject: [PATCH v6 1/3] Reduce planning time for large NOT IN lists containing
NULL
For x <> ALL (...) / x NOT IN (...), the presence of a NULL element
makes the selectivity 0.0.
The planner currently still iterates over all elements and computes
per-element selectivity, even though the final result is known.
Add an early NULL check for constant arrays and immediately return
0.0 under ALL semantics.
This reduces planning time for large NOT IN / <> ALL lists without
changing semantics.
---
src/backend/utils/adt/selfuncs.c | 9 +++++
src/test/regress/expected/expressions.out | 44 +++++++++++++++++++++++
src/test/regress/sql/expressions.sql | 41 +++++++++++++++++++++
3 files changed, 94 insertions(+)
diff --git a/src/backend/utils/adt/selfuncs.c b/src/backend/utils/adt/selfuncs.c
index 29fec655593..eef3f0375a5 100644
--- a/src/backend/utils/adt/selfuncs.c
+++ b/src/backend/utils/adt/selfuncs.c
@@ -2018,6 +2018,11 @@ scalararraysel(PlannerInfo *root,
if (arrayisnull) /* qual can't succeed if null array */
return (Selectivity) 0.0;
arrayval = DatumGetArrayTypeP(arraydatum);
+
+ /* Selectivity of "WHERE x NOT IN (NULL, ... )" is always 0 */
+ if (!useOr && array_contains_nulls(arrayval))
+ return (Selectivity) 0.0;
+
get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
&elmlen, &elmbyval, &elmalign);
deconstruct_array(arrayval,
@@ -2115,6 +2120,10 @@ scalararraysel(PlannerInfo *root,
List *args;
Selectivity s2;
+ /* Selectivity of "WHERE x NOT IN (NULL, ... )" is always 0 */
+ if (!useOr && IsA(elem, Const) && ((Const *) elem)->constisnull)
+ return (Selectivity) 0.0;
+
/*
* Theoretically, if elem isn't of nominal_element_type we should
* insert a RelabelType, but it seems unlikely that any operator
diff --git a/src/test/regress/expected/expressions.out b/src/test/regress/expected/expressions.out
index 9a3c97b15a3..34f14a5775a 100644
--- a/src/test/regress/expected/expressions.out
+++ b/src/test/regress/expected/expressions.out
@@ -426,3 +426,47 @@ select * from inttest where a not in (0::myint,2::myint,3::myint,4::myint,5::myi
(0 rows)
rollback;
+-- Test <> ALL when array initially contained NULL but no longer does
+begin;
+create function check_estimated_rows(text) returns table (estimated int)
+language plpgsql as
+$$
+declare
+ ln text;
+ tmp text[];
+ first_row bool := true;
+begin
+ for ln in
+ execute format('explain %s', $1)
+ loop
+ if first_row then
+ first_row := false;
+ tmp := regexp_match(ln, 'rows=(\d*)');
+ return query select tmp[1]::int;
+ end if;
+ end loop;
+end;
+$$;
+create function replace_elem(arr int[], idx int, val int)
+returns int[] AS $$
+begin
+ arr[idx] := val;
+ return arr;
+end;
+$$ language plpgsql immutable;
+create table notin_test as select generate_series(1, 1000) as x;
+analyze notin_test;
+select * from check_estimated_rows('select * from notin_test where x <> all(array[1,99,3])');
+ estimated
+-----------
+ 997
+(1 row)
+
+-- same array, constructed from an array with a NULL
+select * from check_estimated_rows('select * from notin_test where x <> all(replace_elem(array[1,null,3], 2, 99))');
+ estimated
+-----------
+ 997
+(1 row)
+
+rollback;
diff --git a/src/test/regress/sql/expressions.sql b/src/test/regress/sql/expressions.sql
index e02c21f3368..ca94859bbf8 100644
--- a/src/test/regress/sql/expressions.sql
+++ b/src/test/regress/sql/expressions.sql
@@ -209,3 +209,44 @@ select * from inttest where a not in (1::myint,2::myint,3::myint,4::myint,5::myi
select * from inttest where a not in (0::myint,2::myint,3::myint,4::myint,5::myint, null);
rollback;
+
+-- Test <> ALL when array initially contained NULL but no longer does
+
+begin;
+
+create function check_estimated_rows(text) returns table (estimated int)
+language plpgsql as
+$$
+declare
+ ln text;
+ tmp text[];
+ first_row bool := true;
+begin
+ for ln in
+ execute format('explain %s', $1)
+ loop
+ if first_row then
+ first_row := false;
+ tmp := regexp_match(ln, 'rows=(\d*)');
+ return query select tmp[1]::int;
+ end if;
+ end loop;
+end;
+$$;
+
+create function replace_elem(arr int[], idx int, val int)
+returns int[] AS $$
+begin
+ arr[idx] := val;
+ return arr;
+end;
+$$ language plpgsql immutable;
+
+create table notin_test as select generate_series(1, 1000) as x;
+analyze notin_test;
+
+select * from check_estimated_rows('select * from notin_test where x <> all(array[1,99,3])');
+-- same array, constructed from an array with a NULL
+select * from check_estimated_rows('select * from notin_test where x <> all(replace_elem(array[1,null,3], 2, 99))');
+
+rollback;
\ No newline at end of file
--
2.34.1
^ permalink raw reply [nested|flat] 7+ messages in thread
* Re: Hash-based MCV matching for large IN-lists
2026-02-25 22:45 Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
@ 2026-02-26 08:57 ` Ilia Evdokimov <[email protected]>
2026-03-02 21:37 ` Re: Hash-based MCV matching for large IN-lists Zsolt Parragi <[email protected]>
0 siblings, 1 reply; 7+ messages in thread
From: Ilia Evdokimov @ 2026-02-26 08:57 UTC (permalink / raw)
To: Zsolt Parragi <[email protected]>; +Cc: Tatsuya Kawata <[email protected]>; David Geier <[email protected]>; Chengpeng Yan <[email protected]>; [email protected] <[email protected]>
On 2/26/26 01:45, Ilia Evdokimov wrote:
> About op_is_reserved. It seems we should assign op_is_reserved = true,
> because we don't reverse types like eqjoinsel_semi(). If IN-list
> smaller than MCV-list we reverse it by fmgr_info(hash_mcv ? hashLeft :
> hashRight, &hash_proc). Thanks for this remark.
>
I guess I rushed to conclusions. This assignment op_is_reversed = true
was incorrect. During lookups, simplehash passes: key0 as the value
stored in the hash table, key1 as the probe value. Since MCV entries
correspond to the variable's statistics, the correct argument order
depends on which side we build the hash table on. If we hash MCV values
(hash_mcv = true), then key0 = MCV value, key1 = IN-list value, so we
must call operator(key0, key1). If we hash IN-list elements (hash_mcv =
fasle), then key0 = IN-list value, key1 = MCV value and we must call
operator(key1, key0). Therefore the correct assignment is
hashContext.op_is_reversed = hash_mcv.
If you have another suggestions to v6 patches, send them, and I'll fix
them with hashContext.op_is_reversed = hash_mcv.
--
Best regards,
Ilia Evdokimov,
Tantor Labs LLC,
https://tantorlabs.com/
^ permalink raw reply [nested|flat] 7+ messages in thread
* Re: Hash-based MCV matching for large IN-lists
2026-02-25 22:45 Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
2026-02-26 08:57 ` Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
@ 2026-03-02 21:37 ` Zsolt Parragi <[email protected]>
2026-03-10 14:55 ` Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
0 siblings, 1 reply; 7+ messages in thread
From: Zsolt Parragi @ 2026-03-02 21:37 UTC (permalink / raw)
To: Ilia Evdokimov <[email protected]>; +Cc: David Geier <[email protected]>; Chengpeng Yan <[email protected]>; Tatsuya Kawata <[email protected]>; [email protected] <[email protected]>
Hello!
+ if (vardata.isunique && vardata.rel && vardata.rel->tuples >= 1.0)
+ {
+ s2 = 1.0 / vardata.rel->tuples;
+ if (HeapTupleIsValid(vardata.statsTuple))
+ {
+ Form_pg_statistic stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
+ if (isInequality)
+ s2 = 1.0 - s2 - stats->stanullfrac;
+ }
+ }
Isn't there's a corner case where this if order returns an incorrect
estimate/regression?
See the following test:
CREATE TABLE test AS SELECT generate_series(1, 1000) AS id;
CREATE UNIQUE INDEX ON test(id);
-- no ANALYZE
EXPLAIN SELECT * FROM test WHERE id <> ALL(ARRAY[1, 2, 3]);
-- Actual: rows=1
-- Expected: rows=997
ANALYZE test;
EXPLAIN SELECT * FROM test WHERE id <> ALL(ARRAY[1, 2, 3]);
-- Correct: rows=997
DROP TABLE test;
^ permalink raw reply [nested|flat] 7+ messages in thread
* Re: Hash-based MCV matching for large IN-lists
2026-02-25 22:45 Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
2026-02-26 08:57 ` Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
2026-03-02 21:37 ` Re: Hash-based MCV matching for large IN-lists Zsolt Parragi <[email protected]>
@ 2026-03-10 14:55 ` Ilia Evdokimov <[email protected]>
2026-03-11 08:01 ` Re: Hash-based MCV matching for large IN-lists Zsolt Parragi <[email protected]>
0 siblings, 1 reply; 7+ messages in thread
From: Ilia Evdokimov @ 2026-03-10 14:55 UTC (permalink / raw)
To: Zsolt Parragi <[email protected]>; +Cc: David Geier <[email protected]>; Chengpeng Yan <[email protected]>; Tatsuya Kawata <[email protected]>; [email protected] <[email protected]>
In the thread discussing ALL semantics and NULL [0], the question was
raised about adding a new regression test that checks selectivity
estimation. If the change gets committed, it would make sense to add
tests for this case as well.
Regarding the idea of optimizing the loop when all per-element
selectivities are the same: I ran some quick tests to see how much the
change in the v7-0002 patch affects planning time. Even without that
patch, iterating over an array with 50k elements takes about 30 ms.
```
CREATE TABLE t (val bytea PRIMARY KEY);
INSERT INTO t SELECT int4send(i) FROM generate_series(1,50000) AS i;
ANALYZE t;
SELECT n_distinct FROM pg_stats WHERE tablename = 't';
n_distinct
------------
-1
(1 row)
SELECT string_agg(format('int4send(%s)', i), ',') FROM
generate_series(1,50000) AS i \gset
EXPLAIN (SUMMARY) SELECT * FROM t WHERE val = ANY
(ARRAY[:string_agg]::bytea[]);
..........
Planning Time: 32.816 ms
(3 rows)
```
Given that, I don't see much benefit in adding additional logic here
just to avoid the loop. It would likely introduce extra code complexity
without a manful gain. If there is interest in optimization this case
further, I can revisit it and add the additional patch.
The patch v8 can still be reviewed as-is, and if the selectivity
regression test gets committed [0], I will add corresponding tests for
this change as well.
[0]:
https://www.postgresql.org/message-id/390a46f3-dbc4-4dc1-b49d-5cc61dd36026%40tantorlabs.com
--
Best regards,
Ilia Evdokimov,
Tantor Labs LLC,
https://tantorlabs.com/
Attachments:
[text/x-patch] v8-0001-Use-hash-based-MCV-matching-for-ScalarArrayOpExpr.patch (18.9K, 2-v8-0001-Use-hash-based-MCV-matching-for-ScalarArrayOpExpr.patch)
download | inline diff:
From 22be37fc625921dbb1722a18dde6b6e9da00890a Mon Sep 17 00:00:00 2001
From: Ilia Evdokimov <[email protected]>
Date: Tue, 10 Mar 2026 17:11:16 +0300
Subject: [PATCH v8] Use hash-based MCV matching for ScalarArrayOpExpr
selectivity
When estimating selectivity for ScalarArrayOpExpr (IN / ANY / ALL) with
available MCV statistics, the planner currently matches IN-list elements
against the MCV array using nested loops. For large IN-lists and/or large
MCV lists this leads to O(N*M) planning-time behavior.
This patch adds a hash-based matching strategy, similar to the one used
in join selectivity estimation. When MCV statistics are available and the
operator supports hashing, the smaller of the two inputs (MCV list or
IN-list constant elements) is chosen as the hash table build side, and
the other side is scanned once, reducing complexity to O(N+M).
The hash-based path is restricted to equality and inequality operators
that use eqsel()/neqsel(), and is applied only when suitable hash
functions and MCV statistics are available.
---
src/backend/utils/adt/selfuncs.c | 518 ++++++++++++++++++++++++++++++-
1 file changed, 513 insertions(+), 5 deletions(-)
diff --git a/src/backend/utils/adt/selfuncs.c b/src/backend/utils/adt/selfuncs.c
index d4da0e8dea9..66c3e31eaa5 100644
--- a/src/backend/utils/adt/selfuncs.c
+++ b/src/backend/utils/adt/selfuncs.c
@@ -146,23 +146,27 @@
/*
* In production builds, switch to hash-based MCV matching when the lists are
* large enough to amortize hash setup cost. (This threshold is compared to
- * the sum of the lengths of the two MCV lists. This is simplistic but seems
+ * the sum of the lengths of the two lists. This is simplistic but seems
* to work well enough.) In debug builds, we use a smaller threshold so that
* the regression tests cover both paths well.
*/
#ifndef USE_ASSERT_CHECKING
-#define EQJOINSEL_MCV_HASH_THRESHOLD 200
+#define MCV_HASH_THRESHOLD 200
#else
-#define EQJOINSEL_MCV_HASH_THRESHOLD 20
+#define MCV_HASH_THRESHOLD 20
#endif
-/* Entries in the simplehash hash table used by eqjoinsel_find_matches */
+/*
+ * Entries in the simplehash hash table used by
+ * eqjoinsel_find_matches and scalararray_mcv_hash_match
+ */
typedef struct MCVHashEntry
{
Datum value; /* the value represented by this entry */
int index; /* its index in the relevant AttStatsSlot */
uint32 hash; /* hash code for the Datum */
char status; /* status code used by simplehash.h */
+ int count; /* number of occurrences of current value in */
} MCVHashEntry;
/* private_data for the simplehash hash table */
@@ -184,6 +188,16 @@ get_relation_stats_hook_type get_relation_stats_hook = NULL;
get_index_stats_hook_type get_index_stats_hook = NULL;
static double eqsel_internal(PG_FUNCTION_ARGS, bool negate);
+static double scalararray_mcv_hash_match(VariableStatData *vardata, Oid operator,
+ Oid collation, Selectivity s2,
+ Datum *elem_values, bool *elem_nulls,
+ int num_elems, bool *elem_const,
+ Oid nominal_element_type, bool useOr,
+ bool isEquality, bool isInequality);
+static void accum_scalararray_prob(double individual_s, int count,
+ bool useOr, bool isEquality,
+ bool isInequality, double nullfrac,
+ double *selec, double *s1disjoint);
static double eqjoinsel_inner(FmgrInfo *eqproc, Oid collation,
Oid hashLeft, Oid hashRight,
VariableStatData *vardata1, VariableStatData *vardata2,
@@ -1893,6 +1907,36 @@ strip_array_coercion(Node *node)
return node;
}
+/*
+ * Accumulate the selectivity contribution of a single array element
+ * into the running ScalarArrayOpExpr selectivity estimate.
+ */
+static void
+accum_scalararray_prob(double individual_s, int count, bool useOr, bool isEquality,
+ bool isInequality, double nullfrac, double *selec, double *s1disjoint)
+{
+ if (count <= 0)
+ return;
+
+ if (isInequality)
+ individual_s = 1.0 - individual_s - nullfrac;
+
+ CLAMP_PROBABILITY(individual_s);
+
+ if (useOr)
+ {
+ *selec = 1.0 - (1.0 - *selec) * pow(1.0 - individual_s, count);
+ if (isEquality)
+ *s1disjoint += individual_s * count;
+ }
+ else
+ {
+ *selec = (*selec) * pow(individual_s, count);
+ if (isInequality)
+ *s1disjoint += count * (individual_s - 1.0);
+ }
+}
+
/*
* scalararraysel - Selectivity of ScalarArrayOpExpr Node.
*/
@@ -2025,6 +2069,36 @@ scalararraysel(PlannerInfo *root,
elmlen, elmbyval, elmalign,
&elem_values, &elem_nulls, &num_elems);
+ /* Try to avoid O(N^2) selectivity calculation for ScalarArrayOpExpr */
+ if ((isEquality || isInequality) && !is_join_clause)
+ {
+ VariableStatData vardata;
+ Node *other_op = NULL;
+ bool var_on_left;
+ bool *elem_const = NULL;
+
+ /*
+ * If the clause is of the form "var OP something" or "something
+ * OP var", extract statistics for the variable. Otherwise, fall
+ * back to a default per-element estimate.
+ */
+ if (get_restriction_variable(root, clause->args, varRelid,
+ &vardata, &other_op, &var_on_left))
+ {
+ s1 = scalararray_mcv_hash_match(&vardata, operator,
+ clause->inputcollid, -1.0,
+ elem_values, elem_nulls,
+ num_elems, elem_const,
+ nominal_element_type, useOr,
+ isEquality, isInequality);
+
+ ReleaseVariableStats(vardata);
+
+ if (s1 >= 0.0)
+ return s1;
+ }
+ }
+
/*
* For generic operators, we assume the probability of success is
* independent for each array element. But for "= ANY" or "<> ALL",
@@ -2100,6 +2174,100 @@ scalararraysel(PlannerInfo *root,
get_typlenbyval(arrayexpr->element_typeid,
&elmlen, &elmbyval);
+ /* Try to avoid O(N^2) selectivity calculation for ScalarArrayOpExpr */
+ if ((isEquality || isInequality) && !is_join_clause)
+ {
+ VariableStatData vardata;
+ Node *other_op = NULL;
+ bool var_on_left;
+ int num_elems = list_length(arrayexpr->elements);
+
+ /*
+ * If expression is not variable = something or something =
+ * variable, then fall back to default code path to compute
+ * default selectivity.
+ */
+ if (get_restriction_variable(root, clause->args, varRelid,
+ &vardata, &other_op, &var_on_left))
+ {
+ Selectivity nonconst_sel;
+ Datum *elem_values;
+ bool *elem_nulls;
+ bool *elem_const;
+ ListCell *lc;
+
+ /*
+ * Build arrays describing ARRAY[] elements: - elem_values:
+ * Datum value for Const elements - elem_nulls: whether
+ * element is NULL - elem_const: whether element is a Const
+ * node
+ */
+ elem_values = palloc_array(Datum, num_elems);
+ elem_nulls = palloc0_array(bool, num_elems);
+ elem_const = palloc0_array(bool, num_elems);
+
+ foreach(lc, arrayexpr->elements)
+ {
+ Node *elem_value = (Node *) lfirst(lc);
+ int i = foreach_current_index(lc);
+
+ if (IsA(elem_value, Const))
+ {
+ elem_values[i] = ((Const *) elem_value)->constvalue;
+ elem_nulls[i] = ((Const *) elem_value)->constisnull;
+ elem_const[i] = true;
+ }
+ else
+ {
+ elem_nulls[i] = false;
+ elem_const[i] = false;
+ }
+
+ /*
+ * For ALL semantics, if the array contains NULL, assume
+ * operator is strict. The ScalarArrayOpExpr cannot
+ * evaluate to TRUE, so return zero.
+ */
+ if (!useOr && elem_nulls[i])
+ {
+ pfree(elem_values);
+ pfree(elem_nulls);
+ pfree(elem_const);
+
+ ReleaseVariableStats(vardata);
+
+ return (Selectivity) 0.0;
+ }
+ }
+
+ /*
+ * Compute per-element selectivity via eqsel()/neqsel
+ * semantics.
+ */
+ nonconst_sel = var_eq_non_const(&vardata, operator,
+ clause->inputcollid,
+ other_op, var_on_left,
+ isInequality);
+
+ s1 = scalararray_mcv_hash_match(&vardata, operator,
+ clause->inputcollid,
+ nonconst_sel, elem_values,
+ elem_nulls, num_elems,
+ elem_const,
+ nominal_element_type, useOr,
+ isEquality, isInequality);
+
+ pfree(elem_values);
+ pfree(elem_nulls);
+ pfree(elem_const);
+
+ ReleaseVariableStats(vardata);
+
+ if (s1 >= 0.0)
+ return s1;
+ }
+ }
+
/*
* We use the assumption of disjoint probabilities here too, although
* the odds of equal array elements are rather higher if the elements
@@ -2210,6 +2378,346 @@ scalararraysel(PlannerInfo *root,
return s1;
}
+/*
+ * Estimate selectivity of a ScalarArrayOpExpr (ANY/ALL) using MCV statistics
+ * with hash-based matching.
+ *
+ * This function follows the same probability model as the generic
+ * ScalarArrayOpExpr selectivity code (independent or disjoint probabilities
+ * for OR/AND combinations), but attempts to speed up matching between
+ * IN-list elements and the column's most-common-values (MCV) statistics by
+ * using hashing instead of nested loops.
+ *
+ * MCV statistics are used only to obtain per-value selectivities for
+ * constants that match MCV entries. All probabilities are combined using
+ * the standard ANY/ALL formulas, exactly as in the generic estimator.
+ *
+ * The function may return -1.0 to indicate that hash-based MCV estimation
+ * is not applicable (for example, missing statistics, unsupported operator,
+ * or unavailable hash functions), in which case the caller should fall back
+ * to the generic ScalarArrayOpExpr selectivity estimation.
+ *
+ * Inputs:
+ * vardata: statistics and metadata for the variable being estimated
+ * operator: equality or inequality operator to apply
+ * collation: OID of collation to use
+ * nonconst_sel: selectivity of non-const element
+ * elem_values: array of IN-list element values
+ * elem_nulls: array indicating which IN-list elements are NULL
+ * elem_const: array indicating which IN-list elements are Const nodes.
+ * array is NULL if all elemnets are const.
+ * num_elems: number of IN-list elements
+ * nominal_element_type: type of IN-list elements
+ * useOr: true if elements are combined using OR semantics, false for AND
+ * isEquality: true if the operator behaves like equality
+ * isInequality: true if the operator behaves like inequality
+ *
+ * Result:
+ * Selectivity estimate in the range [0.0, 1.0], or -1.0 if no estimate
+ * could be produced by this function.
+ *
+ * Note:
+ * This function assumes that the operator’s selectivity behavior matches
+ * eqsel()/neqsel semantics. It must not be used for operators with custom
+ * or non-standard selectivity behavior.
+ */
+static double
+scalararray_mcv_hash_match(VariableStatData *vardata, Oid operator,
+ Oid collation, Selectivity nonconst_sel,
+ Datum *elem_values, bool *elem_nulls, int num_elems,
+ bool *elem_const, Oid nominal_element_type,
+ bool useOr, bool isEquality, bool isInequality)
+{
+ Form_pg_statistic stats;
+ AttStatsSlot sslot;
+ FmgrInfo eqproc;
+ double selec = -1.0,
+ s1disjoint,
+ nullfrac = 0.0;
+ Oid hashLeft = InvalidOid,
+ hashRight = InvalidOid,
+ opfuncoid;
+ bool have_mcvs = false;
+
+ /*
+ * If the variable is known to be unique, MCV statistics do not represent
+ * a meaningful frequency distribution, so skip MCV-based estimation.
+ */
+ if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
+ return -1.0;
+
+ /*
+ * For inequality (<>, ALL), we compute probabilities using the negated
+ * equality operator and later transform them as
+ *
+ * p(x <> c) = 1 - p(x = c) - nullfrac
+ */
+ if (isInequality)
+ {
+ operator = get_negator(operator);
+ if (!OidIsValid(operator))
+ return -1.0;
+ }
+
+ opfuncoid = get_opcode(operator);
+ memset(&sslot, 0, sizeof(sslot));
+
+ if (HeapTupleIsValid(vardata->statsTuple))
+ {
+ if (statistic_proc_security_check(vardata, opfuncoid))
+ have_mcvs = get_attstatsslot(&sslot, vardata->statsTuple,
+ STATISTIC_KIND_MCV, InvalidOid,
+ ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
+ }
+
+ if (have_mcvs)
+ {
+ /*
+ * If the MCV list and IN-list are large enough, and the operator
+ * supports hashing, attempt to use hash functions so that MCV–IN
+ * matching can be done in O(N+M) instead of O(N×M).
+ */
+ if (sslot.nvalues + num_elems >= MCV_HASH_THRESHOLD)
+ {
+ fmgr_info(opfuncoid, &eqproc);
+ (void) get_op_hash_functions(operator, &hashLeft, &hashRight);
+ }
+ }
+
+ if (have_mcvs && OidIsValid(hashLeft) && OidIsValid(hashRight))
+ {
+ /* Use a hash table to speed up the matching */
+ LOCAL_FCINFO(fcinfo, 2);
+ LOCAL_FCINFO(hash_fcinfo, 1);
+ MCVHashTable_hash *hashTable;
+ FmgrInfo hash_proc;
+ MCVHashContext hashContext;
+ double sumallcommon = 0.0,
+ nonmcv_selec = 0.0;
+ bool isdefault;
+ bool hash_mcv;
+ double otherdistinct;
+ Datum *arrayHash;
+ Datum *arrayProbe;
+ int nvaluesHash;
+ int nvaluesProbe;
+ int nonmcv_cnt = num_elems;
+ int nonconst_cnt = 0;
+
+ /* Grab the nullfrac for use below. */
+ stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
+ nullfrac = stats->stanullfrac;
+
+ selec = s1disjoint = (useOr ? 0.0 : 1.0);
+
+ InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
+ NULL, NULL);
+ fcinfo->args[0].isnull = false;
+ fcinfo->args[1].isnull = false;
+
+ for (int i = 0; i < sslot.nvalues; i++)
+ sumallcommon += sslot.numbers[i];
+
+ /*
+ * Compute the total probability mass of all non-MCV values. This is
+ * the part of the column distribution not covered by MCVs.
+ */
+ nonmcv_selec = 1.0 - sumallcommon - nullfrac;
+ CLAMP_PROBABILITY(nonmcv_selec);
+
+ /*
+ * Approximate the per-value probability of a non-MCV constant by
+ * dividing the remaining probability mass by the number of other
+ * distinct values.
+ */
+ otherdistinct = get_variable_numdistinct(vardata, &isdefault) - sslot.nnumbers;
+ if (otherdistinct > 1)
+ nonmcv_selec /= otherdistinct;
+
+ if (sslot.nnumbers > 0 &&
+ nonmcv_selec > sslot.numbers[sslot.nnumbers - 1])
+ {
+ nonmcv_selec = sslot.numbers[sslot.nnumbers - 1];
+ }
+
+ /* Make sure we build the hash table on the smaller array. */
+ if (sslot.nvalues <= num_elems)
+ {
+ hash_mcv = true;
+ nvaluesHash = sslot.nvalues;
+ nvaluesProbe = num_elems;
+ arrayHash = sslot.values;
+ arrayProbe = elem_values;
+ }
+ else
+ {
+ hash_mcv = false;
+ nvaluesHash = num_elems;
+ nvaluesProbe = sslot.nvalues;
+ arrayHash = elem_values;
+ arrayProbe = sslot.values;
+ }
+
+ fmgr_info(hash_mcv ? hashLeft : hashRight, &hash_proc);
+ InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
+ NULL, NULL);
+ hash_fcinfo->args[0].isnull = false;
+
+ hashContext.equal_fcinfo = fcinfo;
+ hashContext.hash_fcinfo = hash_fcinfo;
+ hashContext.op_is_reversed = hash_mcv;
+ hashContext.insert_mode = true;
+
+ get_typlenbyval(hash_mcv ? sslot.valuetype : nominal_element_type,
+ &hashContext.hash_typlen,
+ &hashContext.hash_typbyval);
+
+ hashTable = MCVHashTable_create(CurrentMemoryContext,
+ nvaluesHash,
+ &hashContext);
+
+ /* Build a hash table over the smaller input side. */
+ for (int i = 0; i < nvaluesHash; i++)
+ {
+ bool found = false;
+ MCVHashEntry *entry;
+
+ /*
+ * When hashing IN-list values (hash_mcv == false), we only insert
+ * constant, non-NULL elements. NULL and non-Const elements are
+ * counted separately, because they cannot participate in MCV
+ * matching and must be handled later using generic selectivity
+ * estimation.
+ */
+ if (!hash_mcv)
+ {
+ if (elem_nulls[i])
+ {
+ nonmcv_cnt--;
+ continue;
+ }
+
+ if (elem_const != NULL && !elem_const[i])
+ {
+ nonmcv_cnt--;
+ nonconst_cnt++;
+ continue;
+ }
+ }
+
+ entry = MCVHashTable_insert(hashTable, arrayHash[i], &found);
+
+ /*
+ * entry->count tracks how many times the same value appears, so
+ * that duplicate IN-list elements can be folded into the
+ * probability calculation.
+ */
+ if (likely(!found))
+ {
+ entry->index = i;
+ entry->count = 1;
+ }
+ else
+ entry->count++;
+ }
+
+ hashContext.insert_mode = false;
+ if (hashLeft != hashRight)
+ {
+ fmgr_info(hash_mcv ? hashRight : hashLeft, &hash_proc);
+ /* Resetting hash_fcinfo is probably unnecessary, but be safe */
+ InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
+ NULL, NULL);
+ hash_fcinfo->args[0].isnull = false;
+ }
+
+ for (int i = 0; i < nvaluesProbe; i++)
+ {
+ MCVHashEntry *entry;
+ Selectivity s1;
+ int nvaluesmcv;
+
+ /*
+ * When probing with IN-list elements, ignore NULLs and non-Const
+ * expressions: they cannot be matched against MCVs and will be
+ * accounted for later by generic estimation.
+ */
+ if (hash_mcv)
+ {
+ if (elem_nulls[i])
+ {
+ nonmcv_cnt--;
+ continue;
+ }
+
+ if (elem_const != NULL && !elem_const[i])
+ {
+ nonmcv_cnt--;
+ nonconst_cnt++;
+ continue;
+ }
+ }
+
+ entry = MCVHashTable_lookup(hashTable, arrayProbe[i]);
+
+ /*
+ * If found, obtain its MCV frequency and remember how many values
+ * on the hashed side map to this entry.
+ */
+ if (entry != NULL)
+ {
+ s1 = hash_mcv ? sslot.numbers[entry->index]
+ : sslot.numbers[i];
+
+ nvaluesmcv = entry->count;
+
+ accum_scalararray_prob(s1, nvaluesmcv, useOr, isEquality,
+ isInequality, nullfrac, &selec,
+ &s1disjoint);
+
+ /* Matched values are no longer considered non-MCV */
+ nonmcv_cnt -= nvaluesmcv;
+ }
+ }
+
+ /*
+ * Account for constant IN-list values that did not match any MCV.
+ *
+ * Each such value is assumed to have probability = nonmcv_selec,
+ * derived from the remaining (non-MCV) probability mass.
+ */
+ accum_scalararray_prob(nonmcv_selec, nonmcv_cnt, useOr, isEquality,
+ isInequality, nullfrac, &selec, &s1disjoint);
+
+ /*
+ * Account for non-Const IN-list elements.
+ *
+ * These values cannot be matched against MCVs, so we rely on the
+ * operator's generic selectivity estimator for each of them.
+ */
+ accum_scalararray_prob(nonconst_sel, nonconst_cnt, useOr, isEquality,
+ isInequality, nullfrac, &selec, &s1disjoint);
+
+ /*
+ * For = ANY or <> ALL, if the IN-list elements are assumed distinct,
+ * the events are disjoint and the total probability is the sum of
+ * individual probabilities. Use that estimate if it lies in [0,1].
+ */
+ if ((useOr ? isEquality : isInequality) &&
+ s1disjoint >= 0.0 && s1disjoint <= 1.0)
+ selec = s1disjoint;
+
+ CLAMP_PROBABILITY(selec);
+
+ MCVHashTable_destroy(hashTable);
+ }
+
+ if (have_mcvs)
+ free_attstatsslot(&sslot);
+
+ return selec;
+}
+
/*
* Estimate number of elements in the array yielded by an expression.
*
@@ -2446,7 +2954,7 @@ eqjoinsel(PG_FUNCTION_ARGS)
* If the MCV lists are long enough to justify hashing, try to look up
* hash functions for the join operator.
*/
- if ((sslot1.nvalues + sslot2.nvalues) >= EQJOINSEL_MCV_HASH_THRESHOLD)
+ if ((sslot1.nvalues + sslot2.nvalues) >= MCV_HASH_THRESHOLD)
(void) get_op_hash_functions(operator, &hashLeft, &hashRight);
}
else
--
2.34.1
^ permalink raw reply [nested|flat] 7+ messages in thread
* Re: Hash-based MCV matching for large IN-lists
2026-02-25 22:45 Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
2026-02-26 08:57 ` Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
2026-03-02 21:37 ` Re: Hash-based MCV matching for large IN-lists Zsolt Parragi <[email protected]>
2026-03-10 14:55 ` Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
@ 2026-03-11 08:01 ` Zsolt Parragi <[email protected]>
2026-03-20 15:58 ` Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
0 siblings, 1 reply; 7+ messages in thread
From: Zsolt Parragi @ 2026-03-11 08:01 UTC (permalink / raw)
To: Ilia Evdokimov <[email protected]>; +Cc: David Geier <[email protected]>; Chengpeng Yan <[email protected]>; Tatsuya Kawata <[email protected]>; [email protected] <[email protected]>
+ if (elem_nulls[i])
+ {
+ nonmcv_cnt--;
+ continue;
+ }
> The patch v8 can still be reviewed as-is, and if the selectivity
> regression test gets committed [0], I will add corresponding tests for
> this change as well.
Without [0], the const path will return incorrect results for <> ALL
and NULLs. Compared to that, the other path still has special handling
in it:
+ /*
+ * For ALL semantics, if the array contains NULL, assume
+ * operator is strict. The ScalarArrayOpExpr cannot
+ * evaluate to TRUE, so return zero.
+ */
+ nonconst_sel = var_eq_non_const(&vardata, operator,
+ clause->inputcollid,
+ other_op, var_on_left,
+ isInequality);
+ if (isInequality)
+ individual_s = 1.0 - individual_s - nullfrac;
Isn't this the double negation issue again, which was once
mentioned/fixed earlier?
+ int count; /* number of occurrences of current value in */
That's a truncated comment
^ permalink raw reply [nested|flat] 7+ messages in thread
* Re: Hash-based MCV matching for large IN-lists
2026-02-25 22:45 Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
2026-02-26 08:57 ` Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
2026-03-02 21:37 ` Re: Hash-based MCV matching for large IN-lists Zsolt Parragi <[email protected]>
2026-03-10 14:55 ` Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
2026-03-11 08:01 ` Re: Hash-based MCV matching for large IN-lists Zsolt Parragi <[email protected]>
@ 2026-03-20 15:58 ` Ilia Evdokimov <[email protected]>
2026-04-08 16:48 ` Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
0 siblings, 1 reply; 7+ messages in thread
From: Ilia Evdokimov @ 2026-03-20 15:58 UTC (permalink / raw)
To: Zsolt Parragi <[email protected]>; +Cc: David Geier <[email protected]>; Chengpeng Yan <[email protected]>; Tatsuya Kawata <[email protected]>; [email protected] <[email protected]>
On 3/11/26 11:01, Zsolt Parragi wrote:
> + /*
> + * For ALL semantics, if the array contains NULL, assume
> + * operator is strict. The ScalarArrayOpExpr cannot
> + * evaluate to TRUE, so return zero.
> + */
>
>
>
> + nonconst_sel = var_eq_non_const(&vardata, operator,
> + clause->inputcollid,
> + other_op, var_on_left,
> + isInequality);
>
> + if (isInequality)
> + individual_s = 1.0 - individual_s - nullfrac;
>
> Isn't this the double negation issue again, which was once
> mentioned/fixed earlier?
Right. I fixed it by using 'invert' for non-constant case. If there is a
more elegant way to structure this, suggestions are very welcome.
> + int count; /* number of occurrences of current value in */
>
> That's a truncated comment
Fixed.
After the commit c95cd29 I have rebased this patch. During the rebase, I
also add the NUL-handling path. In particular, I added an Assert(useOr)
in the relevant branch to document and enforce the expected execution flow.
Additionally after the 374a639 I prepared a set of regression-style
tests to verify that the selectivity estimates remain unchanged before
and after applying the patch. However, these tests rely on stable row
estimates from EXPLAIN, which are not guaranteed to be consistent across
platforms. For that reason, they are not suitable for inclusion in the
upstream test suite. I will keep these tests locally to validate
correctness before and after the patch.
--
Best regards,
Ilia Evdokimov,
Tantor Labs LLC,
https://tantorlabs.com/
Attachments:
[text/x-patch] v9-0001-Use-hash-based-MCV-matching-for-ScalarArrayOpExpr.patch (18.9K, 2-v9-0001-Use-hash-based-MCV-matching-for-ScalarArrayOpExpr.patch)
download | inline diff:
From 5d93f945a022d38b3dd39c940ba620a3640bb236 Mon Sep 17 00:00:00 2001
From: Evdokimov Ilia <[email protected]>
Date: Fri, 20 Mar 2026 18:18:24 +0300
Subject: [PATCH v9] Use hash-based MCV matching for ScalarArrayOpExpr
selectivity
When estimating selectivity for ScalarArrayOpExpr (IN / ANY / ALL) with
available MCV statistics, the planner currently matches IN-list elements
against the MCV array using nested loops. For large IN-lists and/or large
MCV lists this leads to O(N*M) planning-time behavior.
This patch adds a hash-based matching strategy, similar to the one used
in join selectivity estimation. When MCV statistics are available and the
operator supports hashing, the smaller of the two inputs (MCV list or
IN-list constant elements) is chosen as the hash table build side, and
the other side is scanned once, reducing complexity to O(N+M).
The hash-based path is restricted to equality and inequality operators
that use eqsel()/neqsel(), and is applied only when suitable hash
functions and MCV statistics are available.
---
src/backend/utils/adt/selfuncs.c | 520 ++++++++++++++++++++++++++++++-
1 file changed, 515 insertions(+), 5 deletions(-)
diff --git a/src/backend/utils/adt/selfuncs.c b/src/backend/utils/adt/selfuncs.c
index 86b55c9bb8b..1d812162980 100644
--- a/src/backend/utils/adt/selfuncs.c
+++ b/src/backend/utils/adt/selfuncs.c
@@ -146,23 +146,27 @@
/*
* In production builds, switch to hash-based MCV matching when the lists are
* large enough to amortize hash setup cost. (This threshold is compared to
- * the sum of the lengths of the two MCV lists. This is simplistic but seems
+ * the sum of the lengths of the two lists. This is simplistic but seems
* to work well enough.) In debug builds, we use a smaller threshold so that
* the regression tests cover both paths well.
*/
#ifndef USE_ASSERT_CHECKING
-#define EQJOINSEL_MCV_HASH_THRESHOLD 200
+#define MCV_HASH_THRESHOLD 200
#else
-#define EQJOINSEL_MCV_HASH_THRESHOLD 20
+#define MCV_HASH_THRESHOLD 20
#endif
-/* Entries in the simplehash hash table used by eqjoinsel_find_matches */
+/*
+ * Entries in the simplehash hash table used by
+ * eqjoinsel_find_matches and scalararray_mcv_hash_match
+ */
typedef struct MCVHashEntry
{
Datum value; /* the value represented by this entry */
int index; /* its index in the relevant AttStatsSlot */
uint32 hash; /* hash code for the Datum */
char status; /* status code used by simplehash.h */
+ int count; /* number of occurrences of current value */
} MCVHashEntry;
/* private_data for the simplehash hash table */
@@ -184,6 +188,16 @@ get_relation_stats_hook_type get_relation_stats_hook = NULL;
get_index_stats_hook_type get_index_stats_hook = NULL;
static double eqsel_internal(PG_FUNCTION_ARGS, bool negate);
+static double scalararray_mcv_hash_match(VariableStatData *vardata, Oid operator,
+ Oid collation, Selectivity nonconst_sel,
+ Datum *elem_values, bool *elem_nulls,
+ int num_elems, bool *elem_const,
+ Oid nominal_element_type, bool useOr,
+ bool isEquality, bool isInequality);
+static void accum_scalararray_prob(double s1, int count, bool useOr,
+ bool isEquality, bool isInequality,
+ double nullfrac, bool invert,
+ double *selec, double *s1disjoint);
static double eqjoinsel_inner(FmgrInfo *eqproc, Oid collation,
Oid hashLeft, Oid hashRight,
VariableStatData *vardata1, VariableStatData *vardata2,
@@ -1893,6 +1907,37 @@ strip_array_coercion(Node *node)
return node;
}
+/*
+ * Accumulate the selectivity contribution of a single array element
+ * into the running ScalarArrayOpExpr selectivity estimate.
+ */
+static void
+accum_scalararray_prob(double s1, int count, bool useOr, bool isEquality,
+ bool isInequality, double nullfrac, bool invert,
+ double *selec, double *s1disjoint)
+{
+ if (count <= 0)
+ return;
+
+ if (invert && isInequality)
+ s1 = 1.0 - s1 - nullfrac;
+
+ CLAMP_PROBABILITY(s1);
+
+ if (useOr)
+ {
+ *selec = 1.0 - (1.0 - *selec) * pow(1.0 - s1, count);
+ if (isEquality)
+ *s1disjoint += s1 * count;
+ }
+ else
+ {
+ *selec = (*selec) * pow(s1, count);
+ if (isInequality)
+ *s1disjoint += count * (s1 - 1.0);
+ }
+}
+
/*
* scalararraysel - Selectivity of ScalarArrayOpExpr Node.
*/
@@ -2034,6 +2079,36 @@ scalararraysel(PlannerInfo *root,
elmlen, elmbyval, elmalign,
&elem_values, &elem_nulls, &num_elems);
+ /* Try to avoid O(N^2) selectivity calculation for ScalarArrayOpExpr */
+ if ((isEquality || isInequality) && !is_join_clause)
+ {
+ VariableStatData vardata;
+ Node *other_op = NULL;
+ bool var_on_left;
+ bool *elem_const = NULL;
+
+ /*
+ * If the clause is of the form "var OP something" or "something
+ * OP var", extract statistics for the variable. Otherwise, fall
+ * back to a default per-element estimate.
+ */
+ if (get_restriction_variable(root, clause->args, varRelid,
+ &vardata, &other_op, &var_on_left))
+ {
+ s1 = scalararray_mcv_hash_match(&vardata, operator,
+ clause->inputcollid, -1.0,
+ elem_values, elem_nulls,
+ num_elems, elem_const,
+ nominal_element_type, useOr,
+ isEquality, isInequality);
+
+ ReleaseVariableStats(vardata);
+
+ if (s1 >= 0.0)
+ return s1;
+ }
+ }
+
/*
* For generic operators, we assume the probability of success is
* independent for each array element. But for "= ANY" or "<> ALL",
@@ -2109,6 +2184,100 @@ scalararraysel(PlannerInfo *root,
get_typlenbyval(arrayexpr->element_typeid,
&elmlen, &elmbyval);
+ /* Try to avoid O(N^2) selectivity calculation for ScalarArrayOpExpr */
+ if ((isEquality || isInequality) && !is_join_clause)
+ {
+ VariableStatData vardata;
+ Node *other_op = NULL;
+ bool var_on_left;
+ int num_elems = list_length(arrayexpr->elements);
+
+ /*
+ * If expression is not variable = something or something =
+ * variable, then fall back to default code path to compute
+ * default selectivity.
+ */
+ if (get_restriction_variable(root, clause->args, varRelid,
+ &vardata, &other_op, &var_on_left))
+ {
+ Selectivity nonconst_sel;
+ Datum *elem_values;
+ bool *elem_nulls;
+ bool *elem_const;
+ ListCell *lc;
+
+ /*
+ * Build arrays describing ARRAY[] elements: - elem_values:
+ * Datum value for Const elements - elem_nulls: whether
+ * element is NULL - elem_const: whether element is a Const
+ * node
+ */
+ elem_values = palloc_array(Datum, num_elems);
+ elem_nulls = palloc0_array(bool, num_elems);
+ elem_const = palloc0_array(bool, num_elems);
+
+ foreach(lc, arrayexpr->elements)
+ {
+ Node *elem_value = (Node *) lfirst(lc);
+ int i = foreach_current_index(lc);
+
+ if (IsA(elem_value, Const))
+ {
+ elem_values[i] = ((Const *) elem_value)->constvalue;
+ elem_nulls[i] = ((Const *) elem_value)->constisnull;
+ elem_const[i] = true;
+ }
+ else
+ {
+ elem_nulls[i] = false;
+ elem_const[i] = false;
+ }
+
+ /*
+ * When the array contains a NULL constant, same as var_eq_const,
+ * we assume the operator is strict and nothing will match, thus
+ * return 0.0.
+ */
+ if (!useOr && elem_nulls[i])
+ {
+ pfree(elem_values);
+ pfree(elem_nulls);
+ pfree(elem_const);
+
+ ReleaseVariableStats(vardata);
+
+ return (Selectivity) 0.0;
+ }
+ }
+
+ /*
+ * Compute per-element selectivity via eqsel()/neqsel
+ * semantics.
+ */
+ nonconst_sel = var_eq_non_const(&vardata, operator,
+ clause->inputcollid,
+ other_op, var_on_left,
+ isInequality);
+
+ s1 = scalararray_mcv_hash_match(&vardata, operator,
+ clause->inputcollid,
+ nonconst_sel, elem_values,
+ elem_nulls, num_elems,
+ elem_const,
+ nominal_element_type, useOr,
+ isEquality, isInequality);
+
+ pfree(elem_values);
+ pfree(elem_nulls);
+ pfree(elem_const);
+
+ ReleaseVariableStats(vardata);
+
+ if (s1 >= 0.0)
+ return s1;
+ }
+ }
+
/*
* We use the assumption of disjoint probabilities here too, although
* the odds of equal array elements are rather higher if the elements
@@ -2227,6 +2396,347 @@ scalararraysel(PlannerInfo *root,
return s1;
}
+/*
+ * Estimate selectivity of a ScalarArrayOpExpr (ANY/ALL) using MCV statistics
+ * with hash-based matching.
+ *
+ * This function follows the same probability model as the generic
+ * ScalarArrayOpExpr selectivity code (independent or disjoint probabilities
+ * for OR/AND combinations), but attempts to speed up matching between
+ * IN-list elements and the column's most-common-values (MCV) statistics by
+ * using hashing instead of nested loops.
+ *
+ * MCV statistics are used only to obtain per-value selectivities for
+ * constants that match MCV entries. All probabilities are combined using
+ * the standard ANY/ALL formulas, exactly as in the generic estimator.
+ *
+ * The function may return -1.0 to indicate that hash-based MCV estimation
+ * is not applicable (for example, missing statistics, unsupported operator,
+ * or unavailable hash functions), in which case the caller should fall back
+ * to the generic ScalarArrayOpExpr selectivity estimation.
+ *
+ * Inputs:
+ * vardata: statistics and metadata for the variable being estimated
+ * operator: equality or inequality operator to apply
+ * collation: OID of collation to use
+ * nonconst_sel: selectivity of non-const element
+ * elem_values: array of IN-list element values
+ * elem_nulls: array indicating which IN-list elements are NULL
+ * elem_const: array indicating which IN-list elements are Const nodes.
+ * array is NULL if all elemnets are const.
+ * num_elems: number of IN-list elements
+ * nominal_element_type: type of IN-list elements
+ * useOr: true if elements are combined using OR semantics, false for AND
+ * isEquality: true if the operator behaves like equality
+ * isInequality: true if the operator behaves like inequality
+ *
+ * Result:
+ * Selectivity estimate in the range [0.0, 1.0], or -1.0 if no estimate
+ * could be produced by this function.
+ *
+ * Note:
+ * This function assumes that the operator’s selectivity behavior matches
+ * eqsel()/neqsel semantics. It must not be used for operators with custom
+ * or non-standard selectivity behavior.
+ */
+static double
+scalararray_mcv_hash_match(VariableStatData *vardata, Oid operator,
+ Oid collation, Selectivity nonconst_sel,
+ Datum *elem_values, bool *elem_nulls, int num_elems,
+ bool *elem_const, Oid nominal_element_type,
+ bool useOr, bool isEquality, bool isInequality)
+{
+ Form_pg_statistic stats;
+ AttStatsSlot sslot;
+ FmgrInfo eqproc;
+ double selec = -1.0,
+ s1disjoint,
+ nullfrac = 0.0;
+ Oid hashLeft = InvalidOid,
+ hashRight = InvalidOid,
+ opfuncoid;
+ bool have_mcvs = false;
+
+ /*
+ * If the variable is known to be unique, MCV statistics do not represent
+ * a meaningful frequency distribution, so skip MCV-based estimation.
+ */
+ if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
+ return -1.0;
+
+ /*
+ * For inequality (<>, ALL), we compute probabilities using the negated
+ * equality operator and later transform them as
+ *
+ * p(x <> c) = 1 - p(x = c) - nullfrac
+ */
+ if (isInequality)
+ {
+ operator = get_negator(operator);
+ if (!OidIsValid(operator))
+ return -1.0;
+ }
+
+ opfuncoid = get_opcode(operator);
+ memset(&sslot, 0, sizeof(sslot));
+
+ if (HeapTupleIsValid(vardata->statsTuple))
+ {
+ if (statistic_proc_security_check(vardata, opfuncoid))
+ have_mcvs = get_attstatsslot(&sslot, vardata->statsTuple,
+ STATISTIC_KIND_MCV, InvalidOid,
+ ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
+ }
+
+ if (have_mcvs)
+ {
+ /*
+ * If the MCV list and IN-list are large enough, and the operator
+ * supports hashing, attempt to use hash functions so that MCV–IN
+ * matching can be done in O(N+M) instead of O(N×M).
+ */
+ if (sslot.nvalues + num_elems >= MCV_HASH_THRESHOLD)
+ {
+ fmgr_info(opfuncoid, &eqproc);
+ (void) get_op_hash_functions(operator, &hashLeft, &hashRight);
+ }
+ }
+
+ if (have_mcvs && OidIsValid(hashLeft) && OidIsValid(hashRight))
+ {
+ /* Use a hash table to speed up the matching */
+ LOCAL_FCINFO(fcinfo, 2);
+ LOCAL_FCINFO(hash_fcinfo, 1);
+ MCVHashTable_hash *hashTable;
+ FmgrInfo hash_proc;
+ MCVHashContext hashContext;
+ double sumallcommon = 0.0,
+ nonmcv_selec = 0.0;
+ bool isdefault;
+ bool hash_mcv;
+ double otherdistinct;
+ Datum *arrayHash;
+ Datum *arrayProbe;
+ int nvaluesHash;
+ int nvaluesProbe;
+ int nonmcv_cnt = num_elems;
+ int nonconst_cnt = 0;
+
+ /* Grab the nullfrac for use below. */
+ stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
+ nullfrac = stats->stanullfrac;
+
+ selec = s1disjoint = (useOr ? 0.0 : 1.0);
+
+ InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
+ NULL, NULL);
+ fcinfo->args[0].isnull = false;
+ fcinfo->args[1].isnull = false;
+
+ for (int i = 0; i < sslot.nvalues; i++)
+ sumallcommon += sslot.numbers[i];
+
+ /*
+ * Compute the total probability mass of all non-MCV values. This is
+ * the part of the column distribution not covered by MCVs.
+ */
+ nonmcv_selec = 1.0 - sumallcommon - nullfrac;
+ CLAMP_PROBABILITY(nonmcv_selec);
+
+ /*
+ * Approximate the per-value probability of a non-MCV constant by
+ * dividing the remaining probability mass by the number of other
+ * distinct values.
+ */
+ otherdistinct = get_variable_numdistinct(vardata, &isdefault) - sslot.nnumbers;
+ if (otherdistinct > 1)
+ nonmcv_selec /= otherdistinct;
+
+ if (sslot.nnumbers > 0 && nonmcv_selec > sslot.numbers[sslot.nnumbers - 1])
+ nonmcv_selec = sslot.numbers[sslot.nnumbers - 1];
+
+ /* Make sure we build the hash table on the smaller array. */
+ if (sslot.nvalues <= num_elems)
+ {
+ hash_mcv = true;
+ nvaluesHash = sslot.nvalues;
+ nvaluesProbe = num_elems;
+ arrayHash = sslot.values;
+ arrayProbe = elem_values;
+ }
+ else
+ {
+ hash_mcv = false;
+ nvaluesHash = num_elems;
+ nvaluesProbe = sslot.nvalues;
+ arrayHash = elem_values;
+ arrayProbe = sslot.values;
+ }
+
+ fmgr_info(hash_mcv ? hashLeft : hashRight, &hash_proc);
+ InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
+ NULL, NULL);
+ hash_fcinfo->args[0].isnull = false;
+
+ hashContext.equal_fcinfo = fcinfo;
+ hashContext.hash_fcinfo = hash_fcinfo;
+ hashContext.op_is_reversed = hash_mcv;
+ hashContext.insert_mode = true;
+
+ get_typlenbyval(hash_mcv ? sslot.valuetype : nominal_element_type,
+ &hashContext.hash_typlen,
+ &hashContext.hash_typbyval);
+
+ hashTable = MCVHashTable_create(CurrentMemoryContext,
+ nvaluesHash,
+ &hashContext);
+
+ /* Build a hash table over the smaller input side. */
+ for (int i = 0; i < nvaluesHash; i++)
+ {
+ bool found = false;
+ MCVHashEntry *entry;
+
+ /*
+ * When hashing IN-list values (hash_mcv == false), we only insert
+ * constant, non-NULL elements. NULL and non-Const elements are
+ * counted separately, because they cannot participate in MCV
+ * matching and must be handled later using generic selectivity
+ * estimation.
+ */
+ if (!hash_mcv)
+ {
+ if (elem_nulls[i])
+ {
+ Assert(useOr);
+ nonmcv_cnt--;
+ continue;
+ }
+
+ if (elem_const != NULL && !elem_const[i])
+ {
+ nonmcv_cnt--;
+ nonconst_cnt++;
+ continue;
+ }
+ }
+
+ entry = MCVHashTable_insert(hashTable, arrayHash[i], &found);
+
+ /*
+ * entry->count tracks how many times the same value appears, so
+ * that duplicate IN-list elements can be folded into the
+ * probability calculation.
+ */
+ if (likely(!found))
+ {
+ entry->index = i;
+ entry->count = 1;
+ }
+ else
+ entry->count++;
+ }
+
+ hashContext.insert_mode = false;
+ if (hashLeft != hashRight)
+ {
+ fmgr_info(hash_mcv ? hashRight : hashLeft, &hash_proc);
+ /* Resetting hash_fcinfo is probably unnecessary, but be safe */
+ InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
+ NULL, NULL);
+ hash_fcinfo->args[0].isnull = false;
+ }
+
+ for (int i = 0; i < nvaluesProbe; i++)
+ {
+ MCVHashEntry *entry;
+ Selectivity s1;
+ int nvaluesmcv;
+
+ /*
+ * When probing with IN-list elements, ignore NULLs and non-Const
+ * expressions: they cannot be matched against MCVs and will be
+ * accounted for later by generic estimation.
+ */
+ if (hash_mcv)
+ {
+ if (elem_nulls[i])
+ {
+ Assert(useOr);
+ nonmcv_cnt--;
+ continue;
+ }
+
+ if (elem_const != NULL && !elem_const[i])
+ {
+ nonmcv_cnt--;
+ nonconst_cnt++;
+ continue;
+ }
+ }
+
+ entry = MCVHashTable_lookup(hashTable, arrayProbe[i]);
+
+ /*
+ * If found, obtain its MCV frequency and remember how many values
+ * on the hashed side map to this entry.
+ */
+ if (entry != NULL)
+ {
+ s1 = hash_mcv ? sslot.numbers[entry->index]
+ : sslot.numbers[i];
+
+ nvaluesmcv = entry->count;
+
+ accum_scalararray_prob(s1, nvaluesmcv, useOr, isEquality,
+ isInequality, nullfrac, true, &selec,
+ &s1disjoint);
+
+ /* Matched values are no longer considered non-MCV */
+ nonmcv_cnt -= nvaluesmcv;
+ }
+ }
+
+ /*
+ * Account for constant IN-list values that did not match any MCV.
+ *
+ * Each such value is assumed to have probability = nonmcv_selec,
+ * derived from the remaining (non-MCV) probability mass.
+ */
+ accum_scalararray_prob(nonmcv_selec, nonmcv_cnt, useOr, isEquality,
+ isInequality, nullfrac, true,
+ &selec, &s1disjoint);
+
+ /*
+ * Account for non-Const IN-list elements.
+ *
+ * These values cannot be matched against MCVs, so we rely on the
+ * operator's generic selectivity estimator for each of them.
+ */
+ accum_scalararray_prob(nonconst_sel, nonconst_cnt, useOr, isEquality,
+ isInequality, nullfrac, false,
+ &selec, &s1disjoint);
+
+ /*
+ * For = ANY or <> ALL, if the IN-list elements are assumed distinct,
+ * the events are disjoint and the total probability is the sum of
+ * individual probabilities. Use that estimate if it lies in [0,1].
+ */
+ if ((useOr ? isEquality : isInequality) &&
+ s1disjoint >= 0.0 && s1disjoint <= 1.0)
+ selec = s1disjoint;
+
+ CLAMP_PROBABILITY(selec);
+
+ MCVHashTable_destroy(hashTable);
+ }
+
+ if (have_mcvs)
+ free_attstatsslot(&sslot);
+
+ return selec;
+}
+
/*
* Estimate number of elements in the array yielded by an expression.
*
@@ -2463,7 +2973,7 @@ eqjoinsel(PG_FUNCTION_ARGS)
* If the MCV lists are long enough to justify hashing, try to look up
* hash functions for the join operator.
*/
- if ((sslot1.nvalues + sslot2.nvalues) >= EQJOINSEL_MCV_HASH_THRESHOLD)
+ if ((sslot1.nvalues + sslot2.nvalues) >= MCV_HASH_THRESHOLD)
(void) get_op_hash_functions(operator, &hashLeft, &hashRight);
}
else
--
2.34.1
[text/x-patch] hash_based_any_tests.patch (14.1K, 3-hash_based_any_tests.patch)
download | inline diff:
diff --git a/src/test/regress/expected/planner_est.out b/src/test/regress/expected/planner_est.out
index b62a47552fa..3ef720908f5 100644
--- a/src/test/regress/expected/planner_est.out
+++ b/src/test/regress/expected/planner_est.out
@@ -210,4 +210,213 @@ false, true, false, true);
-> Result (cost=N..N rows=1 width=N)
(4 rows)
+-- Ensure stable and rich MCV statistics
+SET default_statistics_target = 1000;
+CREATE TABLE t_mcv (a int);
+-- Build ~100 MCV values with uniform distribution
+INSERT INTO t_mcv
+SELECT (i % 100)
+FROM generate_series(1, 20000) s(i);
+ANALYZE t_mcv;
+-- =========================================================
+-- CASE 1: Large ANY list (MCV < ANY) → hash on MCV
+-- =========================================================
+-- 1. Single element
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY(SELECT 1));$$,
+false, true, false, true);
+ explain_mask_costs
+--------------------------------------------------
+ Seq Scan on t_mcv (cost=N..N rows=1912 width=N)
+ Filter: (a = ANY ((InitPlan array_1).col1))
+ InitPlan array_1
+ -> Result (cost=N..N rows=1 width=N)
+(4 rows)
+
+-- 2. Multiple elements (all in MCV)
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY(SELECT i FROM generate_series(1,3) s(i)));$$,
+false, true, false, true);
+ explain_mask_costs
+------------------------------------------------------------------------
+ Seq Scan on t_mcv (cost=N..N rows=1912 width=N)
+ Filter: (a = ANY ((InitPlan array_1).col1))
+ InitPlan array_1
+ -> Function Scan on generate_series s (cost=N..N rows=3 width=N)
+(4 rows)
+
+-- 3. Includes non-MCV values
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY[1, 2, 1000]);$$,
+false, true, false, true);
+ explain_mask_costs
+-------------------------------------------------
+ Seq Scan on t_mcv (cost=N..N rows=400 width=N)
+ Filter: (a = ANY ('{1,2,1000}'::integer[]))
+(2 rows)
+
+-- 4. Duplicates
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY(SELECT (i % 3) + 1 FROM generate_series(1,30) s(i)));$$,
+false, true, false, true);
+ explain_mask_costs
+-------------------------------------------------------------------------
+ Seq Scan on t_mcv (cost=N..N rows=1912 width=N)
+ Filter: (a = ANY ((InitPlan array_1).col1))
+ InitPlan array_1
+ -> Function Scan on generate_series s (cost=N..N rows=30 width=N)
+(4 rows)
+
+-- =========================================================
+-- CASE 2: Small ANY list (ANY < MCV) → hash on ANY
+-- =========================================================
+-- 1. Single element
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY[10]);$$,
+false, true, false, true);
+ explain_mask_costs
+-------------------------------------------------
+ Seq Scan on t_mcv (cost=N..N rows=200 width=N)
+ Filter: (a = ANY ('{10}'::integer[]))
+(2 rows)
+
+-- 2. Multiple elements (all in MCV)
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY[10,20,30]);$$,
+false, true, false, true);
+ explain_mask_costs
+-------------------------------------------------
+ Seq Scan on t_mcv (cost=N..N rows=600 width=N)
+ Filter: (a = ANY ('{10,20,30}'::integer[]))
+(2 rows)
+
+-- 3. Includes non-MCV values
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY[10,20,10000]);$$,
+false, true, false, true);
+ explain_mask_costs
+--------------------------------------------------
+ Seq Scan on t_mcv (cost=N..N rows=400 width=N)
+ Filter: (a = ANY ('{10,20,10000}'::integer[]))
+(2 rows)
+
+-- 4. Duplicates
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY[10,10,10,20,20]);$$,
+false, true, false, true);
+ explain_mask_costs
+-----------------------------------------------------
+ Seq Scan on t_mcv (cost=N..N rows=1000 width=N)
+ Filter: (a = ANY ('{10,10,10,20,20}'::integer[]))
+(2 rows)
+
+-- =========================================================
+-- CASE 3: Guaranteed large case → stress hash path
+-- =========================================================
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY(SELECT i % 100 FROM generate_series(1,500) s(i)));$$,
+false, true, false, true);
+ explain_mask_costs
+--------------------------------------------------------------------------
+ Seq Scan on t_mcv (cost=N..N rows=1912 width=N)
+ Filter: (a = ANY ((InitPlan array_1).col1))
+ InitPlan array_1
+ -> Function Scan on generate_series s (cost=N..N rows=500 width=N)
+(4 rows)
+
+-- =========================================================
+-- CASE 4: inequality (<> ALL)
+-- =========================================================
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a <> ALL (ARRAY[1,2,3]);$$,
+false, true, false, true);
+ explain_mask_costs
+---------------------------------------------------
+ Seq Scan on t_mcv (cost=N..N rows=19400 width=N)
+ Filter: (a <> ALL ('{1,2,3}'::integer[]))
+(2 rows)
+
+-- =========================================================
+-- CASE 5: mix const + non-const
+-- =========================================================
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY[1,2] || ARRAY(SELECT i FROM generate_series(3,5) s(i)));$$,
+false, true, false, true);
+ explain_mask_costs
+------------------------------------------------------------------------
+ Seq Scan on t_mcv (cost=N..N rows=1912 width=N)
+ Filter: (a = ANY (('{1,2}'::integer[] || (InitPlan array_1).col1)))
+ InitPlan array_1
+ -> Function Scan on generate_series s (cost=N..N rows=3 width=N)
+(4 rows)
+
+-- =========================================================
+-- CASE 6: NULL handling
+-- =========================================================
+-- ANY with NULL
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY[NULL,1]);$$,
+false, true, false, true);
+ explain_mask_costs
+-------------------------------------------------
+ Seq Scan on t_mcv (cost=N..N rows=200 width=N)
+ Filter: (a = ANY ('{NULL,1}'::integer[]))
+(2 rows)
+
+-- ALL with NULL (should be 0 selectivity)
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ALL (ARRAY[1,NULL]);$$,
+false, true, false, true);
+ explain_mask_costs
+-----------------------------------------------
+ Seq Scan on t_mcv (cost=N..N rows=1 width=N)
+ Filter: (a = ALL ('{1,NULL}'::integer[]))
+(2 rows)
+
+-- =========================================================
+-- CASE 7: Combined all of them
+-- =========================================================
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY[1, 2, 3, 1000, 2000, NULL, 1, 1, 2, (SELECT 4), (SELECT 5), (SELECT 10000), (SELECT 4), (SELECT 4)] || ARRAY( SELECT i % 120 FROM generate_series(1, 500) s(i)));$$,
+false, true, false, true);
+ explain_mask_costs
+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+ Seq Scan on t_mcv (cost=N..N rows=1912 width=N)
+ Filter: (a = ANY ((ARRAY[1, 2, 3, 1000, 2000, NULL::integer, 1, 1, 2, (InitPlan expr_1).col1, (InitPlan expr_2).col1, (InitPlan expr_3).col1, (InitPlan expr_4).col1, (InitPlan expr_5).col1] || (InitPlan array_1).col1)))
+ InitPlan expr_1
+ -> Result (cost=N..N rows=1 width=N)
+ InitPlan expr_2
+ -> Result (cost=N..N rows=1 width=N)
+ InitPlan expr_3
+ -> Result (cost=N..N rows=1 width=N)
+ InitPlan expr_4
+ -> Result (cost=N..N rows=1 width=N)
+ InitPlan expr_5
+ -> Result (cost=N..N rows=1 width=N)
+ InitPlan array_1
+ -> Function Scan on generate_series s (cost=N..N rows=500 width=N)
+(14 rows)
+
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a <> ALL (ARRAY[1, 2, 3, 1000, 2000, NULL, 1, 1, 2, (SELECT 4), (SELECT 5), (SELECT 10000), (SELECT 4), (SELECT 4)] || ARRAY( SELECT i % 120 FROM generate_series(1, 500) s(i)));$$,
+false, true, false, true);
+ explain_mask_costs
+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+ Seq Scan on t_mcv (cost=N..N rows=18088 width=N)
+ Filter: (a <> ALL ((ARRAY[1, 2, 3, 1000, 2000, NULL::integer, 1, 1, 2, (InitPlan expr_1).col1, (InitPlan expr_2).col1, (InitPlan expr_3).col1, (InitPlan expr_4).col1, (InitPlan expr_5).col1] || (InitPlan array_1).col1)))
+ InitPlan expr_1
+ -> Result (cost=N..N rows=1 width=N)
+ InitPlan expr_2
+ -> Result (cost=N..N rows=1 width=N)
+ InitPlan expr_3
+ -> Result (cost=N..N rows=1 width=N)
+ InitPlan expr_4
+ -> Result (cost=N..N rows=1 width=N)
+ InitPlan expr_5
+ -> Result (cost=N..N rows=1 width=N)
+ InitPlan array_1
+ -> Function Scan on generate_series s (cost=N..N rows=500 width=N)
+(14 rows)
+
+DROP TABLE t_mcv;
DROP FUNCTION explain_mask_costs(text, bool, bool, bool, bool);
diff --git a/src/test/regress/sql/planner_est.sql b/src/test/regress/sql/planner_est.sql
index 53210d5baad..ba8f8bd8fb6 100644
--- a/src/test/regress/sql/planner_est.sql
+++ b/src/test/regress/sql/planner_est.sql
@@ -147,4 +147,116 @@ SELECT explain_mask_costs($$
SELECT * FROM tenk1 WHERE unique1 <> ALL (ARRAY[1, 2, 98, (SELECT 99), NULL]);$$,
false, true, false, true);
+-- Ensure stable and rich MCV statistics
+SET default_statistics_target = 1000;
+
+CREATE TABLE t_mcv (a int);
+
+-- Build ~100 MCV values with uniform distribution
+INSERT INTO t_mcv
+SELECT (i % 100)
+FROM generate_series(1, 20000) s(i);
+
+ANALYZE t_mcv;
+
+-- =========================================================
+-- CASE 1: Large ANY list (MCV < ANY) → hash on MCV
+-- =========================================================
+
+-- 1. Single element
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY(SELECT 1));$$,
+false, true, false, true);
+
+-- 2. Multiple elements (all in MCV)
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY(SELECT i FROM generate_series(1,3) s(i)));$$,
+false, true, false, true);
+
+-- 3. Includes non-MCV values
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY[1, 2, 1000]);$$,
+false, true, false, true);
+
+-- 4. Duplicates
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY(SELECT (i % 3) + 1 FROM generate_series(1,30) s(i)));$$,
+false, true, false, true);
+
+-- =========================================================
+-- CASE 2: Small ANY list (ANY < MCV) → hash on ANY
+-- =========================================================
+
+-- 1. Single element
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY[10]);$$,
+false, true, false, true);
+
+-- 2. Multiple elements (all in MCV)
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY[10,20,30]);$$,
+false, true, false, true);
+
+-- 3. Includes non-MCV values
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY[10,20,10000]);$$,
+false, true, false, true);
+
+-- 4. Duplicates
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY[10,10,10,20,20]);$$,
+false, true, false, true);
+
+-- =========================================================
+-- CASE 3: Guaranteed large case → stress hash path
+-- =========================================================
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY(SELECT i % 100 FROM generate_series(1,500) s(i)));$$,
+false, true, false, true);
+
+-- =========================================================
+-- CASE 4: inequality (<> ALL)
+-- =========================================================
+
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a <> ALL (ARRAY[1,2,3]);$$,
+false, true, false, true);
+
+-- =========================================================
+-- CASE 5: mix const + non-const
+-- =========================================================
+
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY[1,2] || ARRAY(SELECT i FROM generate_series(3,5) s(i)));$$,
+false, true, false, true);
+
+-- =========================================================
+-- CASE 6: NULL handling
+-- =========================================================
+
+-- ANY with NULL
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY[NULL,1]);$$,
+false, true, false, true);
+
+-- ALL with NULL (should be 0 selectivity)
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ALL (ARRAY[1,NULL]);$$,
+false, true, false, true);
+
+-- =========================================================
+-- CASE 7: Combined all of them
+-- =========================================================
+
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a = ANY (ARRAY[1, 2, 3, 1000, 2000, NULL, 1, 1, 2, (SELECT 4), (SELECT 5), (SELECT 10000), (SELECT 4), (SELECT 4)] || ARRAY( SELECT i % 120 FROM generate_series(1, 500) s(i)));$$,
+false, true, false, true);
+
+SELECT explain_mask_costs($$
+SELECT * FROM t_mcv WHERE a <> ALL (ARRAY[1, 2, 3, 1000, 2000, NULL, 1, 1, 2, (SELECT 4), (SELECT 5), (SELECT 10000), (SELECT 4), (SELECT 4)] || ARRAY( SELECT i % 120 FROM generate_series(1, 500) s(i)));$$,
+false, true, false, true);
+
+DROP TABLE t_mcv;
+
+
DROP FUNCTION explain_mask_costs(text, bool, bool, bool, bool);
^ permalink raw reply [nested|flat] 7+ messages in thread
* Re: Hash-based MCV matching for large IN-lists
2026-02-25 22:45 Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
2026-02-26 08:57 ` Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
2026-03-02 21:37 ` Re: Hash-based MCV matching for large IN-lists Zsolt Parragi <[email protected]>
2026-03-10 14:55 ` Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
2026-03-11 08:01 ` Re: Hash-based MCV matching for large IN-lists Zsolt Parragi <[email protected]>
2026-03-20 15:58 ` Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
@ 2026-04-08 16:48 ` Ilia Evdokimov <[email protected]>
0 siblings, 0 replies; 7+ messages in thread
From: Ilia Evdokimov @ 2026-04-08 16:48 UTC (permalink / raw)
To: Zsolt Parragi <[email protected]>; David Geier <[email protected]>; Chengpeng Yan <[email protected]>; Tatsuya Kawata <[email protected]>; +Cc: [email protected] <[email protected]>
I rebased the previous patch after it was marked as "Need rebase"
I also initialized the 'elem_cost' array to all 'true' values to
simplify the code and avoid confusion, and rewrote
accum_scalararray_prob() to improve readability.
--
Best regards,
Ilia Evdokimov,
Tantor Labs LLC,
https://tantorlabs.com/
Attachments:
[text/x-patch] v10-0001-Use-hash-based-MCV-matching-for-ScalarArrayOpExp.patch (19.4K, 2-v10-0001-Use-hash-based-MCV-matching-for-ScalarArrayOpExp.patch)
download | inline diff:
From 9b931efdcbe405eab1d58f2a215cb296ab145a59 Mon Sep 17 00:00:00 2001
From: Ilia Evdokimov <[email protected]>
Date: Wed, 8 Apr 2026 19:45:00 +0300
Subject: [PATCH v10] Use hash-based MCV matching for ScalarArrayOpExpr
selectivity
When estimating selectivity for ScalarArrayOpExpr (IN / ANY / ALL) with
available MCV statistics, the planner currently matches IN-list elements
against the MCV array using nested loops. For large IN-lists and/or large
MCV lists this leads to O(N*M) planning-time behavior.
This patch adds a hash-based matching strategy, similar to the one used
in join selectivity estimation. When MCV statistics are available and the
operator supports hashing, the smaller of the two inputs (MCV list or
IN-list constant elements) is chosen as the hash table build side, and
the other side is scanned once, reducing complexity to O(N+M).
The hash-based path is restricted to equality and inequality operators
that use eqsel()/neqsel(), and is applied only when suitable hash
functions and MCV statistics are available.
---
src/backend/utils/adt/selfuncs.c | 547 ++++++++++++++++++++++++++++++-
1 file changed, 542 insertions(+), 5 deletions(-)
diff --git a/src/backend/utils/adt/selfuncs.c b/src/backend/utils/adt/selfuncs.c
index 4160d2d6e24..af12071ed0e 100644
--- a/src/backend/utils/adt/selfuncs.c
+++ b/src/backend/utils/adt/selfuncs.c
@@ -146,23 +146,27 @@
/*
* In production builds, switch to hash-based MCV matching when the lists are
* large enough to amortize hash setup cost. (This threshold is compared to
- * the sum of the lengths of the two MCV lists. This is simplistic but seems
+ * the sum of the lengths of the two lists. This is simplistic but seems
* to work well enough.) In debug builds, we use a smaller threshold so that
* the regression tests cover both paths well.
*/
#ifndef USE_ASSERT_CHECKING
-#define EQJOINSEL_MCV_HASH_THRESHOLD 200
+#define MCV_HASH_THRESHOLD 200
#else
-#define EQJOINSEL_MCV_HASH_THRESHOLD 20
+#define MCV_HASH_THRESHOLD 20
#endif
-/* Entries in the simplehash hash table used by eqjoinsel_find_matches */
+/*
+ * Entries in the simplehash hash table used by
+ * eqjoinsel_find_matches and scalararray_mcv_hash_match
+ */
typedef struct MCVHashEntry
{
Datum value; /* the value represented by this entry */
int index; /* its index in the relevant AttStatsSlot */
uint32 hash; /* hash code for the Datum */
char status; /* status code used by simplehash.h */
+ int count; /* number of occurrences of current value */
} MCVHashEntry;
/* private_data for the simplehash hash table */
@@ -184,6 +188,16 @@ get_relation_stats_hook_type get_relation_stats_hook = NULL;
get_index_stats_hook_type get_index_stats_hook = NULL;
static double eqsel_internal(PG_FUNCTION_ARGS, bool negate);
+static double scalararray_mcv_hash_match(VariableStatData *vardata, Oid operator,
+ Oid collation, Selectivity nonconst_sel,
+ Datum *elem_values, bool *elem_nulls,
+ int num_elems, bool *elem_const,
+ Oid nominal_element_type, bool useOr,
+ bool isEquality, bool isInequality);
+static void accum_scalararray_prob(Selectivity elem_sel, int count, bool useOr,
+ bool isEquality, bool isInequality,
+ double nullfrac, bool invert,
+ Selectivity *p_selec, Selectivity *p_selec_disjoint);
static double eqjoinsel_inner(FmgrInfo *eqproc, Oid collation,
Oid hashLeft, Oid hashRight,
VariableStatData *vardata1, VariableStatData *vardata2,
@@ -1893,6 +1907,67 @@ strip_array_coercion(Node *node)
return node;
}
+/*
+ * accum_scalararray_prob - combine selectivity for repeated elements
+ *
+ * Update the running selectivity for a ScalarArrayOpExpr by adding the
+ * contribution of 'count' identical elements with per-element selectivity.
+ *
+ * This is equivalent to applying the per-element update 'count' times:
+ *
+ * OR (ANY): P = P + s - P*s
+ * AND (ALL): P = P * s
+ *
+ * but uses closed-form formulas:
+ *
+ * OR: P = 1 - (1 - P) * (1 - s)^count
+ * AND: P = P * s^count
+ *
+ * The selec_disjoint accumulator tracks the alternative "disjoint events"
+ * estimate used for "= ANY" / "<> ALL".
+ */
+static void
+accum_scalararray_prob(Selectivity elem_sel, int count, bool useOr,
+ bool isEquality, bool isInequality,
+ double nullfrac, bool invert,
+ Selectivity *p_selec, Selectivity *p_selec_disjoint)
+{
+ Selectivity selec;
+ Selectivity disjoint;
+
+ if (count <= 0)
+ return;
+
+ /* Convert to inequality probability if needed */
+ if (invert && isInequality)
+ elem_sel = 1.0 - elem_sel - nullfrac;
+
+ CLAMP_PROBABILITY(elem_sel);
+
+ selec = *p_selec;
+ disjoint = *p_selec_disjoint;
+
+ if (useOr)
+ {
+ /* ANY semantics: probability that at least one element matches */
+ selec = 1.0 - (1.0 - selec) * pow(1.0 - elem_sel, count);
+
+ if (isEquality)
+ disjoint += elem_sel * count;
+ }
+ else
+ {
+ /* ALL semantics: probability that all elements match */
+ selec *= pow(elem_sel, count);
+
+ if (isInequality)
+ disjoint += count * (elem_sel - 1.0);
+ }
+
+ *p_selec = selec;
+ *p_selec_disjoint = disjoint;
+}
+
/*
* scalararraysel - Selectivity of ScalarArrayOpExpr Node.
*/
@@ -2034,6 +2109,45 @@ scalararraysel(PlannerInfo *root,
elmlen, elmbyval, elmalign,
&elem_values, &elem_nulls, &num_elems);
+ /* Try to avoid O(N^2) selectivity calculation */
+ if ((isEquality || isInequality) && !is_join_clause)
+ {
+ VariableStatData vardata;
+ Node *other_op = NULL;
+ bool var_on_left;
+
+ /*
+ * If the clause is of the form "var OP something" or "something
+ * OP var", extract statistics for the variable. Otherwise, fall
+ * back to a default per-element estimate.
+ */
+ if (get_restriction_variable(root, clause->args, varRelid,
+ &vardata, &other_op, &var_on_left))
+ {
+ bool *elem_const = palloc_array(bool, num_elems);
+
+ /*
+ * All elements are constants here, since we deconstructed a
+ * Const array.
+ */
+ memset(elem_const, true, sizeof(bool) * num_elems);
+
+ s1 = scalararray_mcv_hash_match(&vardata, operator,
+ clause->inputcollid, -1.0,
+ elem_values, elem_nulls,
+ num_elems, elem_const,
+ nominal_element_type, useOr,
+ isEquality, isInequality);
+
+ pfree(elem_const);
+
+ ReleaseVariableStats(vardata);
+
+ if (s1 >= 0.0)
+ return s1;
+ }
+ }
+
/*
* For generic operators, we assume the probability of success is
* independent for each array element. But for "= ANY" or "<> ALL",
@@ -2109,6 +2223,95 @@ scalararraysel(PlannerInfo *root,
get_typlenbyval(arrayexpr->element_typeid,
&elmlen, &elmbyval);
+ /* Try to avoid O(N^2) selectivity calculation */
+ if ((isEquality || isInequality) && !is_join_clause)
+ {
+ VariableStatData vardata;
+ Node *other_op = NULL;
+ bool var_on_left;
+ int num_elems = list_length(arrayexpr->elements);
+
+ /*
+ * If expression is not variable = something or something =
+ * variable, then fall back to default code path to compute
+ * default selectivity.
+ */
+ if (get_restriction_variable(root, clause->args, varRelid,
+ &vardata, &other_op, &var_on_left))
+ {
+ Selectivity nonconst_sel;
+ Datum *elem_values;
+ bool *elem_nulls;
+ bool *elem_const;
+ ListCell *lc;
+
+ /* Build arrays describing ARRAY[] elements */
+ elem_values = palloc_array(Datum, num_elems);
+ elem_nulls = palloc0_array(bool, num_elems);
+ elem_const = palloc0_array(bool, num_elems);
+
+ foreach(lc, arrayexpr->elements)
+ {
+ Node *elem_value = (Node *) lfirst(lc);
+ int i = foreach_current_index(lc);
+
+ if (IsA(elem_value, Const))
+ {
+ elem_values[i] = ((Const *) elem_value)->constvalue;
+ elem_nulls[i] = ((Const *) elem_value)->constisnull;
+ elem_const[i] = true;
+ }
+ else
+ {
+ elem_nulls[i] = false;
+ elem_const[i] = false;
+ }
+
+ /*
+ * When the array contains a NULL constant, same as
+ * var_eq_const, we assume the operator is strict and
+ * nothing will match, thus return 0.0.
+ */
+ if (!useOr && elem_nulls[i])
+ {
+ pfree(elem_values);
+ pfree(elem_nulls);
+ pfree(elem_const);
+
+ ReleaseVariableStats(vardata);
+
+ return (Selectivity) 0.0;
+ }
+ }
+
+ /*
+ * Non-Const elements cannot be matched against MCV entries so
+ * estimate their selectivity separately using a fallback.
+ */
+ nonconst_sel = var_eq_non_const(&vardata, operator,
+ clause->inputcollid,
+ other_op, var_on_left,
+ isInequality);
+
+ s1 = scalararray_mcv_hash_match(&vardata, operator,
+ clause->inputcollid,
+ nonconst_sel, elem_values,
+ elem_nulls, num_elems,
+ elem_const,
+ nominal_element_type, useOr,
+ isEquality, isInequality);
+
+ pfree(elem_values);
+ pfree(elem_nulls);
+ pfree(elem_const);
+
+ ReleaseVariableStats(vardata);
+
+ if (s1 >= 0.0)
+ return s1;
+ }
+ }
+
/*
* We use the assumption of disjoint probabilities here too, although
* the odds of equal array elements are rather higher if the elements
@@ -2227,6 +2430,339 @@ scalararraysel(PlannerInfo *root,
return s1;
}
+/*
+ * scalararray_mcv_hash_match - Selectivity of ScalarArrayOpExpr by O(N+M)
+ *
+ * This function matches IN-list elements against MCV entries of the variable,
+ * using either hashing (O(N+M)) or fallback nested-loop logic. For matched
+ * elements, selectivity is taken from MCV frequencies; unmatched elements are
+ * handled using fallback estimates.
+ *
+ * The resulting probabilities are combined using the standard ANY/ALL
+ * selectivity model (independent or disjoint events), identical to the
+ * generic estimator.
+ *
+ * Inputs:
+ * vardata: statistics for the variable
+ * operator: equality or inequality operator
+ * collation: collation to use
+ * nonconst_sel: fallback selectivity for non-Const elements
+ * elem_values: IN-list element values
+ * elem_nulls: NULL flags for elements
+ * elem_const: flags indicating Const elements
+ * num_elems: number of elements
+ * nominal_element_type: element type
+ * useOr: OR (ANY) vs AND (ALL) semantics
+ * isEquality: operator behaves like equality
+ * isInequality: operator behaves like inequality
+ *
+ * Result:
+ * Selectivity in [0,1], or -1.0 if MCV-based estimation is not applicable.
+ *
+ * Note:
+ * Assumes eqsel()/neqsel semantics. Each element is accounted for once,
+ * either via MCV match or fallback estimation.
+ */
+static double
+scalararray_mcv_hash_match(VariableStatData *vardata, Oid operator,
+ Oid collation, Selectivity nonconst_sel,
+ Datum *elem_values, bool *elem_nulls, int num_elems,
+ bool *elem_const, Oid nominal_element_type,
+ bool useOr, bool isEquality, bool isInequality)
+{
+ Form_pg_statistic stats;
+ AttStatsSlot sslot;
+ FmgrInfo eqproc;
+ double selec = -1.0,
+ s1disjoint,
+ nullfrac = 0.0;
+ Oid hashLeft = InvalidOid,
+ hashRight = InvalidOid,
+ opfuncoid;
+ bool have_mcvs = false;
+
+ /*
+ * If the variable is known to be unique, MCV statistics do not represent
+ * a meaningful frequency distribution, so skip MCV-based estimation.
+ */
+ if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
+ return -1.0;
+
+ /*
+ * For inequality (<>, ALL), we compute probabilities using the negated
+ * equality operator and later transform them as
+ *
+ * p(x <> c) = 1 - p(x = c) - nullfrac
+ */
+ if (isInequality)
+ {
+ operator = get_negator(operator);
+ if (!OidIsValid(operator))
+ return -1.0;
+ }
+
+ opfuncoid = get_opcode(operator);
+ memset(&sslot, 0, sizeof(sslot));
+
+ if (HeapTupleIsValid(vardata->statsTuple))
+ {
+ if (statistic_proc_security_check(vardata, opfuncoid))
+ have_mcvs = get_attstatsslot(&sslot, vardata->statsTuple,
+ STATISTIC_KIND_MCV, InvalidOid,
+ ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
+ }
+
+ if (have_mcvs)
+ {
+ /*
+ * If the MCV list and IN-list are large enough, and the operator
+ * supports hashing, attempt to use hash functions so that MCV–IN
+ * matching can be done in O(N+M) instead of O(N×M).
+ */
+ if (sslot.nvalues + num_elems >= MCV_HASH_THRESHOLD)
+ {
+ fmgr_info(opfuncoid, &eqproc);
+ (void) get_op_hash_functions(operator, &hashLeft, &hashRight);
+ }
+ }
+
+ if (have_mcvs && OidIsValid(hashLeft) && OidIsValid(hashRight))
+ {
+ /* Use a hash table to speed up the matching */
+ LOCAL_FCINFO(fcinfo, 2);
+ LOCAL_FCINFO(hash_fcinfo, 1);
+ MCVHashTable_hash *hashTable;
+ FmgrInfo hash_proc;
+ MCVHashContext hashContext;
+ Selectivity nonmcv_selec = 0.0;
+ double sumallcommon = 0.0;
+ bool isdefault;
+ bool hash_mcv;
+ double otherdistinct;
+ Datum *arrayHash;
+ Datum *arrayProbe;
+ int nvaluesHash;
+ int nvaluesProbe;
+ int nonmcv_cnt = num_elems;
+ int nonconst_cnt = 0;
+
+ /* Grab the nullfrac for use below. */
+ stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
+ nullfrac = stats->stanullfrac;
+
+ selec = s1disjoint = (useOr ? 0.0 : 1.0);
+
+ InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
+ NULL, NULL);
+ fcinfo->args[0].isnull = false;
+ fcinfo->args[1].isnull = false;
+
+ for (int i = 0; i < sslot.nvalues; i++)
+ sumallcommon += sslot.numbers[i];
+
+ /*
+ * Compute the total probability mass of all non-MCV values. This is
+ * the part of the column distribution not covered by MCVs.
+ */
+ nonmcv_selec = 1.0 - sumallcommon - nullfrac;
+ CLAMP_PROBABILITY(nonmcv_selec);
+
+ /*
+ * Approximate the per-value probability of a non-MCV constant by
+ * dividing the remaining probability mass by the number of other
+ * distinct values.
+ */
+ otherdistinct = get_variable_numdistinct(vardata, &isdefault) - sslot.nnumbers;
+ if (otherdistinct > 1)
+ nonmcv_selec /= otherdistinct;
+
+ if (sslot.nnumbers > 0 && nonmcv_selec > sslot.numbers[sslot.nnumbers - 1])
+ nonmcv_selec = sslot.numbers[sslot.nnumbers - 1];
+
+ /* Make sure we build the hash table on the smaller array. */
+ if (sslot.nvalues <= num_elems)
+ {
+ hash_mcv = true;
+ nvaluesHash = sslot.nvalues;
+ nvaluesProbe = num_elems;
+ arrayHash = sslot.values;
+ arrayProbe = elem_values;
+ }
+ else
+ {
+ hash_mcv = false;
+ nvaluesHash = num_elems;
+ nvaluesProbe = sslot.nvalues;
+ arrayHash = elem_values;
+ arrayProbe = sslot.values;
+ }
+
+ fmgr_info(hash_mcv ? hashLeft : hashRight, &hash_proc);
+ InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
+ NULL, NULL);
+ hash_fcinfo->args[0].isnull = false;
+
+ hashContext.equal_fcinfo = fcinfo;
+ hashContext.hash_fcinfo = hash_fcinfo;
+ hashContext.op_is_reversed = hash_mcv;
+ hashContext.insert_mode = true;
+
+ get_typlenbyval(hash_mcv ? sslot.valuetype : nominal_element_type,
+ &hashContext.hash_typlen,
+ &hashContext.hash_typbyval);
+
+ hashTable = MCVHashTable_create(CurrentMemoryContext,
+ nvaluesHash,
+ &hashContext);
+
+ /* Build a hash table over the smaller input side. */
+ for (int i = 0; i < nvaluesHash; i++)
+ {
+ bool found = false;
+ MCVHashEntry *entry;
+
+ /*
+ * When hashing IN-list values (hash_mcv == false), we only insert
+ * constant, non-NULL elements. NULL and non-Const elements are
+ * counted separately, because they cannot participate in MCV
+ * matching and must be handled later using generic selectivity
+ * estimation.
+ */
+ if (!hash_mcv)
+ {
+ if (elem_nulls[i])
+ {
+ Assert(useOr);
+ nonmcv_cnt--;
+ continue;
+ }
+
+ if (!elem_const[i])
+ {
+ nonmcv_cnt--;
+ nonconst_cnt++;
+ continue;
+ }
+ }
+
+ entry = MCVHashTable_insert(hashTable, arrayHash[i], &found);
+
+ /*
+ * entry->count tracks how many times the same value appears, so
+ * that duplicate IN-list elements can be folded into the
+ * probability calculation.
+ */
+ if (likely(!found))
+ {
+ entry->index = i;
+ entry->count = 1;
+ }
+ else
+ entry->count++;
+ }
+
+ hashContext.insert_mode = false;
+ if (hashLeft != hashRight)
+ {
+ fmgr_info(hash_mcv ? hashRight : hashLeft, &hash_proc);
+ /* Resetting hash_fcinfo is probably unnecessary, but be safe */
+ InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
+ NULL, NULL);
+ hash_fcinfo->args[0].isnull = false;
+ }
+
+ for (int i = 0; i < nvaluesProbe; i++)
+ {
+ MCVHashEntry *entry;
+ Selectivity s1;
+ int nvaluesmcv;
+
+ /*
+ * When probing with IN-list elements, ignore NULLs and non-Const
+ * expressions: they cannot be matched against MCVs and will be
+ * accounted for later by generic estimation.
+ */
+ if (hash_mcv)
+ {
+ if (elem_nulls[i])
+ {
+ Assert(useOr);
+ nonmcv_cnt--;
+ continue;
+ }
+
+ if (!elem_const[i])
+ {
+ nonmcv_cnt--;
+ nonconst_cnt++;
+ continue;
+ }
+ }
+
+ entry = MCVHashTable_lookup(hashTable, arrayProbe[i]);
+
+ /*
+ * If found, obtain its MCV frequency and remember how many values
+ * on the hashed side map to this entry.
+ */
+ if (entry != NULL)
+ {
+ s1 = hash_mcv ? sslot.numbers[entry->index]
+ : sslot.numbers[i];
+
+ nvaluesmcv = entry->count;
+
+ accum_scalararray_prob(s1, nvaluesmcv, useOr, isEquality,
+ isInequality, nullfrac, true, &selec,
+ &s1disjoint);
+
+ /* Matched values are no longer considered non-MCV */
+ nonmcv_cnt -= nvaluesmcv;
+ }
+ }
+
+ nonmcv_cnt = Max(nonmcv_cnt, 0);
+
+ /*
+ * Account for constant IN-list values that did not match any MCV.
+ *
+ * Each such value is assumed to have probability = nonmcv_selec,
+ * derived from the remaining (non-MCV) probability mass.
+ */
+ accum_scalararray_prob(nonmcv_selec, nonmcv_cnt, useOr, isEquality,
+ isInequality, nullfrac, true,
+ &selec, &s1disjoint);
+
+ /*
+ * Account for non-Const IN-list elements.
+ *
+ * These values cannot be matched against MCVs, so we rely on the
+ * operator's generic selectivity estimator for each of them.
+ */
+ accum_scalararray_prob(nonconst_sel, nonconst_cnt, useOr, isEquality,
+ isInequality, nullfrac, false,
+ &selec, &s1disjoint);
+
+ /*
+ * For = ANY or <> ALL, if the IN-list elements are assumed distinct,
+ * the events are disjoint and the total probability is the sum of
+ * individual probabilities. Use that estimate if it lies in [0,1].
+ */
+ if ((useOr ? isEquality : isInequality) &&
+ s1disjoint >= 0.0 && s1disjoint <= 1.0)
+ selec = s1disjoint;
+
+ CLAMP_PROBABILITY(selec);
+
+ MCVHashTable_destroy(hashTable);
+ }
+
+ if (have_mcvs)
+ free_attstatsslot(&sslot);
+
+ return selec;
+}
+
/*
* Estimate number of elements in the array yielded by an expression.
*
@@ -2463,7 +2999,7 @@ eqjoinsel(PG_FUNCTION_ARGS)
* If the MCV lists are long enough to justify hashing, try to look up
* hash functions for the join operator.
*/
- if ((sslot1.nvalues + sslot2.nvalues) >= EQJOINSEL_MCV_HASH_THRESHOLD)
+ if ((sslot1.nvalues + sslot2.nvalues) >= MCV_HASH_THRESHOLD)
(void) get_op_hash_functions(operator, &hashLeft, &hashRight);
}
else
@@ -3103,6 +3639,7 @@ eqjoinsel_find_matches(FmgrInfo *eqproc, Oid collation,
/*
* Support functions for the hash tables used by eqjoinsel_find_matches
+ * and scalararray_mcv_hash_match
*/
static uint32
hash_mcv(MCVHashTable_hash *tab, Datum key)
--
2.34.1
^ permalink raw reply [nested|flat] 7+ messages in thread
end of thread, other threads:[~2026-04-08 16:48 UTC | newest]
Thread overview: 7+ messages (download: mbox mbox.gz follow: Atom feed)
-- links below jump to the message on this page --
2026-02-25 22:45 Re: Hash-based MCV matching for large IN-lists Ilia Evdokimov <[email protected]>
2026-02-26 08:57 ` Ilia Evdokimov <[email protected]>
2026-03-02 21:37 ` Zsolt Parragi <[email protected]>
2026-03-10 14:55 ` Ilia Evdokimov <[email protected]>
2026-03-11 08:01 ` Zsolt Parragi <[email protected]>
2026-03-20 15:58 ` Ilia Evdokimov <[email protected]>
2026-04-08 16:48 ` Ilia Evdokimov <[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