public inbox for [email protected]  
help / color / mirror / Atom feed
From: Alexander Korotkov <[email protected]>
To: Chao Li <[email protected]>
Cc: Dmitry Koval <[email protected]>
Cc: PostgreSQL-development <[email protected]>
Subject: Re: Fix SPLIT PARTITION bound-overlap bug and other improvements
Date: Tue, 19 May 2026 14:00:52 +0300
Message-ID: <CAPpHfdvn=hPrtMjtvJU-7=HR69ZWDvy2pTE9Prau-dgp5Ua0Jg@mail.gmail.com> (raw)
In-Reply-To: <[email protected]>
References: <[email protected]>
	<[email protected]>
	<[email protected]>
	<CAPpHfdvdevTMpjQe09aWdK_nbrE4+M1z09H2SUe3bgxY3_QUdA@mail.gmail.com>
	<[email protected]>
	<CAPpHfdsLnD3NGnPC32oeg4ZWWX441M=SfH0-7bG652G40TpEEQ@mail.gmail.com>
	<[email protected]>
	<CAPpHfduMLJkOSguaKhm4PEWRYHqqvFDX6bu5+E_LSxuy5YjPEw@mail.gmail.com>
	<[email protected]>

Hi, Chao!

On Tue, May 19, 2026 at 5:50 AM Chao Li <[email protected]> wrote:
> > On May 18, 2026, at 20:04, Alexander Korotkov <[email protected]> wrote:
> >
> > On Mon, May 18, 2026 at 2:57 PM Chao Li <[email protected]> wrote:
> >>> <v3-0003-Clarify-SPLIT-PARTITION-bound-requirements-in-doc.patch><v3-0001-Fix-SPLIT-PARTITION-range-bound-validation-with-D.patch><v3-0002-Fix-SPLIT-PARTITION-hint-for-DEFAULT-partition-bo.patch><v3-0004-Reject-degenerate-SPLIT-PARTITION-with-DEFAULT-pa.patch>
> >>
> >> v3-0001 through v3-0003 look good to me.
> >>
> >> For v3-0004, I have a suspicion, but it's late here and my brain is getting slow, so I would like to study it more tomorrow.
> >
> > Sure, take your time.
> >
> > ------
> > Regards,
> > Alexander Korotkov
> > Supabase
>
> My suspicion was that check_split_partition_not_same_bound() now has two paths. The RANGE path honors collation, while the LIST path does not. So I spent some time creating a test that uses a case-insensitive collation:
> ```
> evantest=# create collation case_insensitive (provider=icu, locale='und-u-ks-level2', deterministic = false);
> CREATE COLLATION
> evantest=# create table t (b text collate case_insensitive) partition by list (b);
> CREATE TABLE
> evantest=# create table tp_ab partition of t for values in ('a', 'b');
> CREATE TABLE
> evantest=# alter table t split partition tp_ab into
> evantest-#   (partition tp_a for values in ('a', 'A'),
> evantest(#   partition tp_default default);
> ERROR:  cannot split partition "tp_ab" only to add a DEFAULT partition
> LINE 2:   (partition tp_a for values in ('a', 'A'),
>                      ^
> DETAIL:  The non-DEFAULT partition would keep the same partition bound.
> HINT:  Use CREATE TABLE ... PARTITION OF ... DEFAULT to add a DEFAULT partition.
> ```
>
> In this test, the split partition’s bound is ('a', 'b'), and the new partition’s bound is ('a', 'A'). Their list lengths are both 2, but the two bounds are actually different, because 'a' and 'A' are considered equal by the collation.
>
> So, in the LIST path, since check_partition_bounds_for_split_list() has already ensured that the new partition’s bound is contained within the split partition’s bound, we need to check the reverse direction as well. Whether the split partition’s bound is also contained in the new partition’s bound. If yes, the two bounds are identical.
>
> See the attached v4 for my changes for 0004. 0001-0003 are unchanged. Since 0001 and 0003 are independent of 0004, maybe they can be pushed first.

I've pushed 0001-0003.  Thank you for discovering the collation issue
in 0004.  Note that original approach of using
partition_bounds_equal() can't handle different collations too (as it
internally uses datumIsEqual()).  I've revised the remaining patch:
made function header comment a bit more detailed and added additional
regression tests.  Please, check.

------
Regards,
Alexander Korotkov
Supabase


Attachments:

  [application/octet-stream] v5-0001-Reject-degenerate-SPLIT-PARTITION-with-DEFAULT-pa.patch (12.2K, 2-v5-0001-Reject-degenerate-SPLIT-PARTITION-with-DEFAULT-pa.patch)
  download | inline diff:
From eb03bd8d346c072af5024899acc79b4991cd45a4 Mon Sep 17 00:00:00 2001
From: Alexander Korotkov <[email protected]>
Date: Mon, 18 May 2026 00:37:52 +0300
Subject: [PATCH v5] Reject degenerate SPLIT PARTITION with DEFAULT partition

ALTER TABLE ... SPLIT PARTITION allows a DEFAULT partition to be created
as one of the replacement partitions when the parent table does not
already have one.  However, it should not allow the degenerate case where
a non-DEFAULT partition keeps exactly the same bound as the split
partition and the command merely adds a DEFAULT partition through the
SPLIT PARTITION path.

Detect that case by comparing the bound of the split partition with the
bound of the only non-DEFAULT replacement partition, and raise an error
when they are the same.  Users should add a DEFAULT partition directly
with CREATE TABLE ... PARTITION OF ... DEFAULT or ALTER TABLE ... ATTACH
PARTITION ... DEFAULT instead.

Author: Chao Li <[email protected]>
Reviewed-by: Alexander Korotkov <[email protected]>
Discussion: https://postgr.es/m/[email protected]
---
 src/backend/partitioning/partbounds.c         | 150 ++++++++++++++++++
 src/test/regress/expected/partition_split.out |  59 +++++++
 src/test/regress/sql/partition_split.sql      |  52 ++++++
 3 files changed, 261 insertions(+)

diff --git a/src/backend/partitioning/partbounds.c b/src/backend/partitioning/partbounds.c
index 7d3580cbc10..eac807ddb77 100644
--- a/src/backend/partitioning/partbounds.c
+++ b/src/backend/partitioning/partbounds.c
@@ -5700,6 +5700,146 @@ check_parent_values_in_new_partitions(Relation parent,
 	}
 }
 
+/*
+ * split_partition_values_contained_in_new_part
+ *
+ * (function for BY LIST partitioning)
+ *
+ * Returns true if all values in the LIST bound of the partition being split
+ * are contained in the specified non-DEFAULT replacement partition's bound.
+ *
+ * The caller must already have verified containment in the other direction,
+ * so this check is sufficient to prove that the two LIST bounds are equal.
+ */
+static bool
+split_partition_values_contained_in_new_part(Relation parent,
+											 Oid splitPartOid,
+											 SinglePartitionSpec *part)
+{
+	PartitionKey key = RelationGetPartitionKey(parent);
+	PartitionDesc partdesc = RelationGetPartitionDesc(parent, false);
+	PartitionBoundInfo boundinfo = partdesc->boundinfo;
+	SinglePartitionSpec *parts[1];
+	Datum		datum = PointerGetDatum(NULL);
+
+	Assert(key->strategy == PARTITION_STRATEGY_LIST);
+
+	parts[0] = part;
+
+	/*
+	 * Special processing for NULL value.  Search for a NULL value if the
+	 * split partition contains it.
+	 */
+	if (partition_bound_accepts_nulls(boundinfo) &&
+		partdesc->oids[boundinfo->null_index] == splitPartOid)
+	{
+		if (!find_value_in_new_partitions_list(&key->partsupfunc[0],
+											   key->partcollation, parts, 1,
+											   datum, true))
+			return false;
+	}
+
+	/*
+	 * Search all values of the split partition in the single non-DEFAULT
+	 * replacement partition.
+	 */
+	for (int i = 0; i < boundinfo->ndatums; i++)
+	{
+		if (partdesc->oids[boundinfo->indexes[i]] == splitPartOid)
+		{
+			datum = boundinfo->datums[i][0];
+
+			if (!find_value_in_new_partitions_list(&key->partsupfunc[0],
+												   key->partcollation, parts, 1,
+												   datum, false))
+				return false;
+		}
+	}
+
+	return true;
+}
+
+/*
+ * check_split_partition_not_same_bound
+ *
+ * Reject splitting a non-DEFAULT partition into one non-DEFAULT partition
+ * with the original bound plus a DEFAULT partition.  That form does not
+ * perform a real split; it merely adds a DEFAULT partition to the parent
+ * table through the split-partition path.  Users should use
+ * CREATE TABLE ... PARTITION OF ... DEFAULT or ALTER TABLE ... ATTACH
+ * PARTITION ... DEFAULT for that.
+ *
+ * Must be called after the per-partition bound validation in
+ * check_partitions_for_split() so that containment of new bounds within the
+ * split partition is already established.  Given containment, RANGE bounds
+ * are equal iff their lower and upper rbounds match; LIST bound sets are
+ * equal iff the split partition's values are also contained in the new
+ * partition (the containment is then bidirectional).  Both checks honor
+ * the partition key collation via the operator-family comparators
+ * (partition_rbound_cmp / find_value_in_new_partitions_list), so e.g.
+ * ('a','b') and ('A','B') under a case-insensitive ICU collation are
+ * correctly recognised as the same bound.
+ */
+static void
+check_split_partition_not_same_bound(Relation parent,
+									 Oid splitPartOid,
+									 SinglePartitionSpec **parts,
+									 int nparts,
+									 ParseState *pstate)
+{
+	PartitionKey key = RelationGetPartitionKey(parent);
+	PartitionBoundSpec *new_spec;
+	PartitionBoundSpec *split_spec;
+
+	if (nparts != 1)
+		return;
+
+	new_spec = parts[0]->bound;
+	split_spec = get_partition_bound_spec(splitPartOid);
+
+	Assert(new_spec->strategy == split_spec->strategy);
+
+	if (key->strategy == PARTITION_STRATEGY_RANGE)
+	{
+		PartitionRangeBound *new_lower;
+		PartitionRangeBound *new_upper;
+		PartitionRangeBound *split_lower;
+		PartitionRangeBound *split_upper;
+
+		new_lower = make_one_partition_rbound(key, -1, new_spec->lowerdatums, true);
+		new_upper = make_one_partition_rbound(key, -1, new_spec->upperdatums, false);
+		split_lower = make_one_partition_rbound(key, -1, split_spec->lowerdatums, true);
+		split_upper = make_one_partition_rbound(key, -1, split_spec->upperdatums, false);
+
+		if (partition_rbound_cmp(key->partnatts, key->partsupfunc,
+								 key->partcollation,
+								 new_lower->datums, new_lower->kind, true,
+								 split_lower) != 0)
+			return;
+		if (partition_rbound_cmp(key->partnatts, key->partsupfunc,
+								 key->partcollation,
+								 new_upper->datums, new_upper->kind, false,
+								 split_upper) != 0)
+			return;
+	}
+	else
+	{
+		Assert(key->strategy == PARTITION_STRATEGY_LIST);
+
+		if (!split_partition_values_contained_in_new_part(parent, splitPartOid,
+														  parts[0]))
+			return;
+	}
+
+	ereport(ERROR,
+			errcode(ERRCODE_INVALID_OBJECT_DEFINITION),
+			errmsg("cannot split partition \"%s\" only to add a DEFAULT partition",
+				   get_rel_name(splitPartOid)),
+			errdetail("The non-DEFAULT partition would keep the same partition bound."),
+			errhint("Use CREATE TABLE ... PARTITION OF ... DEFAULT to add a DEFAULT partition."),
+			parser_errposition(pstate, parts[0]->name->location));
+}
+
 /*
  * check_partitions_for_split
  *
@@ -5871,5 +6011,15 @@ check_partitions_for_split(Relation parent,
 												  new_parts, nparts, pstate);
 	}
 
+	/*
+	 * Reject the degenerate form where the single non-DEFAULT replacement
+	 * partition keeps the bound of the split partition; the command then does
+	 * nothing beyond adding a DEFAULT partition.  Containment was established
+	 * by the per-partition validation above, so an equality check is enough.
+	 */
+	if (!isSplitPartDefault && createDefaultPart)
+		check_split_partition_not_same_bound(parent, splitPartOid, new_parts,
+											 nparts, pstate);
+
 	pfree(new_parts);
 }
diff --git a/src/test/regress/expected/partition_split.out b/src/test/regress/expected/partition_split.out
index 2b9a6aa50ed..2fd9aee1dcc 100644
--- a/src/test/regress/expected/partition_split.out
+++ b/src/test/regress/expected/partition_split.out
@@ -1188,6 +1188,65 @@ SELECT tableoid::regclass, * FROM sales_range ORDER BY tableoid::regclass::text
 
 DROP TABLE sales_range;
 --
+-- Test that SPLIT PARTITION rejects the degenerate case where the only
+-- non-DEFAULT replacement partition keeps the original bound and the command
+-- merely adds a DEFAULT partition.
+--
+CREATE TABLE t (i int) PARTITION BY RANGE (i);
+CREATE TABLE tp_0_50 PARTITION OF t FOR VALUES FROM (0) TO (50);
+INSERT INTO t VALUES (1);
+-- ERROR
+ALTER TABLE t SPLIT PARTITION tp_0_50 INTO
+  (PARTITION tp_0_50 FOR VALUES FROM (0) TO (50),
+   PARTITION tp_default DEFAULT);
+ERROR:  cannot split partition "tp_0_50" only to add a DEFAULT partition
+LINE 2:   (PARTITION tp_0_50 FOR VALUES FROM (0) TO (50),
+                     ^
+DETAIL:  The non-DEFAULT partition would keep the same partition bound.
+HINT:  Use CREATE TABLE ... PARTITION OF ... DEFAULT to add a DEFAULT partition.
+DROP TABLE t;
+--
+-- Test that a LIST split with DEFAULT is not considered degenerate when
+-- only NULL is removed from the explicit replacement partition.
+--
+CREATE TABLE t (i int) PARTITION BY LIST (i);
+CREATE TABLE tp_null_1 PARTITION OF t FOR VALUES IN (NULL, 1);
+ALTER TABLE t SPLIT PARTITION tp_null_1 INTO
+  (PARTITION tp_1 FOR VALUES IN (1),
+   PARTITION tp_default DEFAULT);
+INSERT INTO t VALUES (NULL), (1), (2);
+SELECT tableoid::regclass, i FROM t ORDER BY tableoid::regclass::text COLLATE "C", i NULLS FIRST;
+  tableoid  | i 
+------------+---
+ tp_1       | 1
+ tp_default |  
+ tp_default | 2
+(3 rows)
+
+DROP TABLE t;
+--
+-- Test that the same-bound check for LIST partitioning uses partition
+-- comparison semantics, not raw list length.  The case-insensitive collation
+-- treats 'a' and 'A' as equal, so the non-DEFAULT replacement partition
+-- covers only the 'a' group and the DEFAULT partition covers the rest.
+--
+CREATE COLLATION case_insensitive (provider = icu, locale = 'und-u-ks-level2', deterministic = false);
+CREATE TABLE t (b text COLLATE case_insensitive) PARTITION BY LIST (b);
+CREATE TABLE tp_ab PARTITION OF t FOR VALUES IN ('a', 'b');
+ALTER TABLE t SPLIT PARTITION tp_ab INTO
+  (PARTITION tp_a FOR VALUES IN ('a', 'A'),
+   PARTITION tp_default DEFAULT);
+INSERT INTO t VALUES ('a'), ('A'), ('b'), ('c');
+SELECT tableoid::regclass, count(*) FROM t GROUP BY 1 ORDER BY 1;
+  tableoid  | count 
+------------+-------
+ tp_a       |     2
+ tp_default |     2
+(2 rows)
+
+DROP TABLE t;
+DROP COLLATION case_insensitive;
+--
 -- Test that the explicit partition bound cannot extend outside the split
 -- partition's bound when a DEFAULT partition is specified.
 --
diff --git a/src/test/regress/sql/partition_split.sql b/src/test/regress/sql/partition_split.sql
index d9821c5e2a3..ede89ad0228 100644
--- a/src/test/regress/sql/partition_split.sql
+++ b/src/test/regress/sql/partition_split.sql
@@ -834,6 +834,58 @@ SELECT tableoid::regclass, * FROM sales_range ORDER BY tableoid::regclass::text
 
 DROP TABLE sales_range;
 
+--
+-- Test that SPLIT PARTITION rejects the degenerate case where the only
+-- non-DEFAULT replacement partition keeps the original bound and the command
+-- merely adds a DEFAULT partition.
+--
+CREATE TABLE t (i int) PARTITION BY RANGE (i);
+CREATE TABLE tp_0_50 PARTITION OF t FOR VALUES FROM (0) TO (50);
+INSERT INTO t VALUES (1);
+
+-- ERROR
+ALTER TABLE t SPLIT PARTITION tp_0_50 INTO
+  (PARTITION tp_0_50 FOR VALUES FROM (0) TO (50),
+   PARTITION tp_default DEFAULT);
+
+DROP TABLE t;
+
+--
+-- Test that a LIST split with DEFAULT is not considered degenerate when
+-- only NULL is removed from the explicit replacement partition.
+--
+CREATE TABLE t (i int) PARTITION BY LIST (i);
+CREATE TABLE tp_null_1 PARTITION OF t FOR VALUES IN (NULL, 1);
+
+ALTER TABLE t SPLIT PARTITION tp_null_1 INTO
+  (PARTITION tp_1 FOR VALUES IN (1),
+   PARTITION tp_default DEFAULT);
+
+INSERT INTO t VALUES (NULL), (1), (2);
+SELECT tableoid::regclass, i FROM t ORDER BY tableoid::regclass::text COLLATE "C", i NULLS FIRST;
+
+DROP TABLE t;
+
+--
+-- Test that the same-bound check for LIST partitioning uses partition
+-- comparison semantics, not raw list length.  The case-insensitive collation
+-- treats 'a' and 'A' as equal, so the non-DEFAULT replacement partition
+-- covers only the 'a' group and the DEFAULT partition covers the rest.
+--
+CREATE COLLATION case_insensitive (provider = icu, locale = 'und-u-ks-level2', deterministic = false);
+CREATE TABLE t (b text COLLATE case_insensitive) PARTITION BY LIST (b);
+CREATE TABLE tp_ab PARTITION OF t FOR VALUES IN ('a', 'b');
+
+ALTER TABLE t SPLIT PARTITION tp_ab INTO
+  (PARTITION tp_a FOR VALUES IN ('a', 'A'),
+   PARTITION tp_default DEFAULT);
+
+INSERT INTO t VALUES ('a'), ('A'), ('b'), ('c');
+SELECT tableoid::regclass, count(*) FROM t GROUP BY 1 ORDER BY 1;
+
+DROP TABLE t;
+DROP COLLATION case_insensitive;
+
 --
 -- Test that the explicit partition bound cannot extend outside the split
 -- partition's bound when a DEFAULT partition is specified.
-- 
2.39.5 (Apple Git-154)



reply

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Reply to all the recipients using the --to and --cc options:
  reply via email

  To: [email protected]
  Cc: [email protected], [email protected], [email protected]
  Subject: Re: Fix SPLIT PARTITION bound-overlap bug and other improvements
  In-Reply-To: <CAPpHfdvn=hPrtMjtvJU-7=HR69ZWDvy2pTE9Prau-dgp5Ua0Jg@mail.gmail.com>

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

This inbox is served by agora; see mirroring instructions
for how to clone and mirror all data and code used for this inbox