public inbox for [email protected]
help / color / mirror / Atom feedFrom: Antonin Houska <[email protected]>
Subject: [PATCH 2/2] Provide the executor with information on updated columns.
Date: Wed, 1 Jul 2026 15:18:10 +0200
When replaying data changes, REPACK calls ExecInsertIndexTuples() for each
INSERT and UPDATE. In the latter case, the function needs to determine the
value of the 'indexUnchanged' hint for the index AM. (The hint is always false
for INSERT.) Therefore, REPACK is supposed to specify which columns are
changed by the UPDATE.
So far, REPACK missed to specify the set of updated columns, so the
'indexUnchanged' hint could incorrectly evaluate to true. Although it should
not affect correctness, it can make the index AM use optimizations that are
not appropriate. Ideally, we should compute the set of updated columns for
each individual UPDATE, however the comparison of the old and new tuple might
add too much overhead.
This patch initializes the set as if all columns were updated. Thus
'indexUnchanged' always evaluates to false. This way the index AM never uses
the related optimizations. It might result in worse structure of the index,
however it seems better to not use the optimizations than to misuse them.
---
src/backend/commands/repack.c | 40 ++++++++++++++++++++++++++++++++++-
1 file changed, 39 insertions(+), 1 deletion(-)
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index c9b0c047477..f37bb9a21af 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -62,6 +62,7 @@
#include "libpq/pqmq.h"
#include "miscadmin.h"
#include "optimizer/optimizer.h"
+#include "parser/parse_relation.h"
#include "pgstat.h"
#include "replication/logicalrelation.h"
#include "storage/bufmgr.h"
@@ -3005,13 +3006,50 @@ static void
initialize_change_context(ChangeContext *chgcxt,
Relation relation, Oid ident_index_id)
{
+ Bitmapset *updatedCols = NULL;
+ RangeTblEntry *rte;
+ List *perminfos = NIL;
+ RTEPermissionInfo *perminfo;
+
chgcxt->cc_rel = relation;
/* Only initialize fields needed by ExecInsertIndexTuples(). */
chgcxt->cc_estate = CreateExecutorState();
+ /*
+ * Initialize updatedCols.
+ *
+ * The point is that ExecInsertIndexTuples() should not pass the
+ * indexUnchanged hint to the index AM unless there's a reason to do
+ * so. For simplicity, we consider all columns updated.
+ *
+ * XXX Should we spend more effort to compare the old and new tuple when
+ * replaying UPDATE, or at least exclude unchanged TOAST values, like we
+ * do in logicalrep_write_tuple()?
+ */
+ for (int i = 0; i < RelationGetDescr(relation)->natts; i++)
+ updatedCols = bms_add_member(updatedCols,
+ i + 1 - FirstLowInvalidHeapAttributeNumber);
+
+ /*
+ * In this case, RTE only needs to have ->perminfoindex initialized, but
+ * there's no reason to not set the fields whose values we have at hand.
+ */
+ rte = makeNode(RangeTblEntry);
+ rte->rtekind = RTE_RELATION;
+ rte->relid = RelationGetRelid(relation);
+ rte->relkind = RelationGetForm(relation)->relkind;
+ /* Create the RTEPermissionInfo instance (and set ->perminfoindex). */
+ addRTEPermissionInfo(&perminfos, rte);
+ /* Make updatedCols available to the executor functions. */
+ perminfo = getRTEPermissionInfo(perminfos, rte);
+ perminfo->updatedCols = updatedCols;
+
+ ExecInitRangeTable(chgcxt->cc_estate, list_make1(rte), perminfos,
+ bms_make_singleton(1));
+
chgcxt->cc_rri = makeNode(ResultRelInfo);
- InitResultRelInfo(chgcxt->cc_rri, relation, 0, NULL, 0);
+ InitResultRelInfo(chgcxt->cc_rri, relation, 1, NULL, 0);
ExecOpenIndices(chgcxt->cc_rri, false);
/*
--
2.52.0
--=-=-=--
view thread (14+ messages) latest in thread
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]
Subject: Re: [PATCH 2/2] Provide the executor with information on updated columns.
In-Reply-To: <no-message-id-1869138@localhost>
* 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