From 6290ab46da8f19993ad0d48ebc7d5e11a649c8cf Mon Sep 17 00:00:00 2001 From: Mikhail Nikalayeu Date: Thu, 3 Sep 2026 10:53:45 +0200 Subject: [PATCH v4] Fix tuple search during apply after concurrent index DDL. When the apply worker searches the local relation by index, it can take the first match as the row only if the index is the relation's replica identity or primary key. Otherwise every match has to be compared with the search slot, which holds a complete row only under REPLICA IDENTITY FULL. Only the index OID was saved, so the scan worked this out a second time from the catalogs, and the two answers can differ. Apply holds only RowExclusiveLock, which doesn't conflict with DROP INDEX CONCURRENTLY or REINDEX CONCURRENTLY, so either can demote the chosen index in between. Nothing matches, so the change is silently dropped as an update_missing conflict, and assert-enabled builds fail. Fix this by recording the answer as idxisreplident in the relation map entry and passing the entry down to FindReplTupleInLocalRel(), so it is settled once. Oversight in 89e46da5e5. Author: Mikhail Nikalayeu Reviewed-by: Amit Kapila Reviewed-by: vignesh C Reviewed-by: Zhijie Hou Discussion: https://postgr.es/m/CADzfLwUJovFcnknCC9wjZKECX9xecgnGzC2r2TMV8h4QDD_jwQ@mail.gmail.com Backpatch-through: 16, where it was introduced --- src/backend/executor/execReplication.c | 15 +- src/backend/replication/logical/conflict.c | 4 + src/backend/replication/logical/relation.c | 18 +- src/backend/replication/logical/worker.c | 94 ++++---- src/include/executor/executor.h | 2 + src/include/replication/logicalrelation.h | 4 + .../subscription/t/032_subscribe_use_index.pl | 212 ++++++++++++++++++ 7 files changed, 301 insertions(+), 48 deletions(-) diff --git a/src/backend/executor/execReplication.c b/src/backend/executor/execReplication.c index fd9efd94737..25caf3d85c2 100644 --- a/src/backend/executor/execReplication.c +++ b/src/backend/executor/execReplication.c @@ -177,9 +177,15 @@ should_refetch_tuple(TM_Result res, TM_FailureData *tmfd) * * If a matching tuple is found, lock it with lockmode, fill the slot with its * contents, and return true. Return false otherwise. + * + * 'isIdxSafeToSkipDuplicates' specifies whether the first matching tuple + * can be used without comparing it against 'searchslot'. If false, all + * matching tuples are compared against 'searchslot', which must contain + * a complete row. */ bool RelationFindReplTupleByIndex(Relation rel, Oid idxoid, + bool isIdxSafeToSkipDuplicates, LockTupleMode lockmode, TupleTableSlot *searchslot, TupleTableSlot *outslot) @@ -192,13 +198,10 @@ RelationFindReplTupleByIndex(Relation rel, Oid idxoid, Relation idxrel; bool found; TypeCacheEntry **eq = NULL; - bool isIdxSafeToSkipDuplicates; /* Open the index. */ idxrel = index_open(idxoid, RowExclusiveLock); - isIdxSafeToSkipDuplicates = (GetRelationIdentityOrPK(rel) == idxoid); - InitDirtySnapshot(snap); /* Build scan key. */ @@ -629,9 +632,12 @@ RelationFindDeletedTupleInfoSeq(Relation rel, TupleTableSlot *searchslot, /* * Similar to RelationFindDeletedTupleInfoSeq() but using index scan to locate * the deleted tuple. + * + * 'isIdxSafeToSkipDuplicates' works as in RelationFindReplTupleByIndex(). */ bool RelationFindDeletedTupleInfoByIndex(Relation rel, Oid idxoid, + bool isIdxSafeToSkipDuplicates, TupleTableSlot *searchslot, TransactionId oldestxmin, TransactionId *delete_xid, @@ -644,7 +650,6 @@ RelationFindDeletedTupleInfoByIndex(Relation rel, Oid idxoid, IndexScanDesc scan; TupleTableSlot *scanslot; TypeCacheEntry **eq = NULL; - bool isIdxSafeToSkipDuplicates; TupleDesc desc PG_USED_FOR_ASSERTS_ONLY = RelationGetDescr(rel); Assert(equalTupleDescs(desc, searchslot->tts_tupleDescriptor)); @@ -654,8 +659,6 @@ RelationFindDeletedTupleInfoByIndex(Relation rel, Oid idxoid, *delete_time = 0; *delete_origin = InvalidReplOriginId; - isIdxSafeToSkipDuplicates = (GetRelationIdentityOrPK(rel) == idxoid); - scanslot = table_slot_create(rel, NULL); idxrel = index_open(idxoid, RowExclusiveLock); diff --git a/src/backend/replication/logical/conflict.c b/src/backend/replication/logical/conflict.c index b71a0c9e206..90431ffce98 100644 --- a/src/backend/replication/logical/conflict.c +++ b/src/backend/replication/logical/conflict.c @@ -743,6 +743,10 @@ get_tuple_desc(EState *estate, ResultRelInfo *relinfo, ConflictType type, * when applying update or delete, such an index scan may not result * in a unique tuple and we still compare the complete tuple in such * cases, thus such indexes are not used here. + * + * XXX This can disagree with the index the apply worker searched by, + * see FindReplTupleInLocalRel(). It may not even be one that + * ExecOpenIndices() locked. */ Oid replica_index = GetRelationIdentityOrPK(localrel); diff --git a/src/backend/replication/logical/relation.c b/src/backend/replication/logical/relation.c index 8ffd2583afb..6242ce70ad2 100644 --- a/src/backend/replication/logical/relation.c +++ b/src/backend/replication/logical/relation.c @@ -56,7 +56,7 @@ typedef struct LogicalRepPartMapEntry } LogicalRepPartMapEntry; static Oid FindLogicalRepLocalIndex(Relation localrel, LogicalRepRelation *remoterel, - AttrMap *attrMap); + AttrMap *attrMap, bool *idxisreplident); /* * Relcache invalidation callback for our relation map cache. @@ -497,7 +497,8 @@ logicalrep_rel_open(LogicalRepRelId remoteid, LOCKMODE lockmode) * on the relation). */ entry->localindexoid = FindLogicalRepLocalIndex(entry->localrel, remoterel, - entry->attrmap); + entry->attrmap, + &entry->idxisreplident); entry->localrelvalid = true; } @@ -764,7 +765,8 @@ logicalrep_partition_open(LogicalRepRelMapEntry *root, * anything in the LogicalRepPartMapContext (hence CacheMemoryContext). */ entry->localindexoid = FindLogicalRepLocalIndex(partrel, remoterel, - entry->attrmap); + entry->attrmap, + &entry->idxisreplident); entry->localrelvalid = true; @@ -925,13 +927,18 @@ GetRelationIdentityOrPK(Relation rel) /* * Returns the index oid if we can use an index for subscriber. Otherwise, * returns InvalidOid. + * + * '*idxisreplident' is true if the returned index is the relation's replica + * identity or primary key, and false otherwise. */ static Oid FindLogicalRepLocalIndex(Relation localrel, LogicalRepRelation *remoterel, - AttrMap *attrMap) + AttrMap *attrMap, bool *idxisreplident) { Oid idxoid; + *idxisreplident = false; + /* * We never need index oid for partitioned tables, always rely on leaf * partition's index. @@ -944,7 +951,10 @@ FindLogicalRepLocalIndex(Relation localrel, LogicalRepRelation *remoterel, */ idxoid = GetRelationIdentityOrPK(localrel); if (OidIsValid(idxoid)) + { + *idxisreplident = true; return idxoid; + } if (remoterel->replident == REPLICA_IDENTITY_FULL) { diff --git a/src/backend/replication/logical/worker.c b/src/backend/replication/logical/worker.c index 7781bb1c168..ab7c4ced66d 100644 --- a/src/backend/replication/logical/worker.c +++ b/src/backend/replication/logical/worker.c @@ -290,6 +290,7 @@ #include "tcop/tcopprot.h" #include "utils/acl.h" #include "utils/guc.h" +#include "utils/injection_point.h" #include "utils/inval.h" #include "utils/lsyscache.h" #include "utils/memutils.h" @@ -609,19 +610,17 @@ static void apply_handle_insert_internal(ApplyExecutionData *edata, static void apply_handle_update_internal(ApplyExecutionData *edata, ResultRelInfo *relinfo, TupleTableSlot *remoteslot, - LogicalRepTupleData *newtup, - Oid localindexoid); + LogicalRepTupleData *newtup); static void apply_handle_delete_internal(ApplyExecutionData *edata, ResultRelInfo *relinfo, TupleTableSlot *remoteslot, - Oid localindexoid); + LogicalRepRelMapEntry *relmapentry); static bool FindReplTupleInLocalRel(ApplyExecutionData *edata, Relation localrel, - LogicalRepRelation *remoterel, - Oid localidxoid, + LogicalRepRelMapEntry *relmapentry, TupleTableSlot *remoteslot, TupleTableSlot **localslot); static bool FindDeletedTupleInLocalRel(Relation localrel, - Oid localidxoid, + LogicalRepRelMapEntry *relmapentry, TupleTableSlot *remoteslot, TransactionId *delete_xid, ReplOriginId *delete_origin, @@ -2800,11 +2799,8 @@ check_relation_updatable(LogicalRepRelMapEntry *rel) if (rel->updatable) return; - /* - * We are in error mode so it's fine this is somewhat slow. It's better to - * give user correct error. - */ - if (OidIsValid(GetRelationIdentityOrPK(rel->localrel))) + /* Use the entry, so this matches what updatable was decided from. */ + if (rel->idxisreplident) { ereport(ERROR, (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE), @@ -2931,7 +2927,7 @@ apply_handle_update(StringInfo s) remoteslot, &newtup, CMD_UPDATE); else apply_handle_update_internal(edata, edata->targetRelInfo, - remoteslot, &newtup, rel->localindexoid); + remoteslot, &newtup); finish_edata(edata); @@ -2955,8 +2951,7 @@ static void apply_handle_update_internal(ApplyExecutionData *edata, ResultRelInfo *relinfo, TupleTableSlot *remoteslot, - LogicalRepTupleData *newtup, - Oid localindexoid) + LogicalRepTupleData *newtup) { EState *estate = edata->estate; LogicalRepRelMapEntry *relmapentry = edata->targetRel; @@ -2968,11 +2963,12 @@ apply_handle_update_internal(ApplyExecutionData *edata, MemoryContext oldctx; EvalPlanQualInit(&epqstate, estate, NULL, NIL, -1, NIL); + + INJECTION_POINT("apply-update-before-open-indices", NULL); + ExecOpenIndices(relinfo, false); - found = FindReplTupleInLocalRel(edata, localrel, - &relmapentry->remoterel, - localindexoid, + found = FindReplTupleInLocalRel(edata, localrel, relmapentry, remoteslot, &localslot); /* @@ -3026,7 +3022,7 @@ apply_handle_update_internal(ApplyExecutionData *edata, * Detecting whether the tuple was recently deleted or never existed * is crucial to avoid misleading the user during conflict handling. */ - if (FindDeletedTupleInLocalRel(localrel, localindexoid, remoteslot, + if (FindDeletedTupleInLocalRel(localrel, relmapentry, remoteslot, &conflicttuple.xmin, &conflicttuple.origin, &conflicttuple.ts) && @@ -3128,7 +3124,7 @@ apply_handle_delete(StringInfo s) ExecOpenIndices(relinfo, false); apply_handle_delete_internal(edata, relinfo, - remoteslot, rel->localindexoid); + remoteslot, rel); ExecCloseIndices(relinfo); } @@ -3154,11 +3150,10 @@ static void apply_handle_delete_internal(ApplyExecutionData *edata, ResultRelInfo *relinfo, TupleTableSlot *remoteslot, - Oid localindexoid) + LogicalRepRelMapEntry *relmapentry) { EState *estate = edata->estate; Relation localrel = relinfo->ri_RelationDesc; - LogicalRepRelation *remoterel = &edata->targetRel->remoterel; EPQState epqstate; TupleTableSlot *localslot; ConflictTupleInfo conflicttuple = {0}; @@ -3171,7 +3166,7 @@ apply_handle_delete_internal(ApplyExecutionData *edata, !localrel->rd_rel->relhasindex || RelationGetIndexList(localrel) == NIL); - found = FindReplTupleInLocalRel(edata, localrel, remoterel, localindexoid, + found = FindReplTupleInLocalRel(edata, localrel, relmapentry, remoteslot, &localslot); /* If found delete it. */ @@ -3216,16 +3211,20 @@ apply_handle_delete_internal(ApplyExecutionData *edata, * the corresponding local relation using either replica identity index, * primary key, index or if needed, sequential scan. * + * 'relmapentry' is the relation map entry for 'localrel'. It tells which + * index to use, if any, and whether that index is the relation's replica + * identity or primary key. + * * Local tuple, if found, is returned in '*localslot'. */ static bool FindReplTupleInLocalRel(ApplyExecutionData *edata, Relation localrel, - LogicalRepRelation *remoterel, - Oid localidxoid, + LogicalRepRelMapEntry *relmapentry, TupleTableSlot *remoteslot, TupleTableSlot **localslot) { EState *estate = edata->estate; + Oid localidxoid = relmapentry->localindexoid; bool found; /* @@ -3237,22 +3236,39 @@ FindReplTupleInLocalRel(ApplyExecutionData *edata, Relation localrel, *localslot = table_slot_create(localrel, &estate->es_tupleTable); Assert(OidIsValid(localidxoid) || - (remoterel->replident == REPLICA_IDENTITY_FULL)); + (relmapentry->remoterel.replident == REPLICA_IDENTITY_FULL)); if (OidIsValid(localidxoid)) { #ifdef USE_ASSERT_CHECKING Relation idxrel = index_open(localidxoid, AccessShareLock); - /* Index must be PK, RI, or usable for REPLICA IDENTITY FULL tables */ - Assert(GetRelationIdentityOrPK(localrel) == localidxoid || - (remoterel->replident == REPLICA_IDENTITY_FULL && - IsIndexUsableForReplicaIdentityFull(idxrel, - edata->targetRel->attrmap))); + if (relmapentry->idxisreplident) + { + /* + * We cannot assert this is still the replica identity or primary + * key. DROP INDEX CONCURRENTLY and REINDEX CONCURRENTLY clear + * indisvalid and indisreplident without conflicting with our + * RowExclusiveLock, so GetRelationIdentityOrPK() may no longer + * return it. Unique and non-partial is what the scan actually + * relies on, and no DDL can take those away. + */ + Assert(idxrel->rd_index->indisunique); + Assert(heap_attisnull(idxrel->rd_indextuple, + Anum_pg_index_indpred, NULL)); + } + else + { + /* Otherwise every match is compared, so we need a whole row. */ + Assert(relmapentry->remoterel.replident == REPLICA_IDENTITY_FULL); + Assert(IsIndexUsableForReplicaIdentityFull(idxrel, + relmapentry->attrmap)); + } index_close(idxrel, AccessShareLock); #endif found = RelationFindReplTupleByIndex(localrel, localidxoid, + relmapentry->idxisreplident, LockTupleExclusive, remoteslot, *localslot); } @@ -3309,16 +3325,21 @@ IsIndexUsableForFindingDeletedTuple(Oid localindexoid, * The search is performed using either the replica identity index, primary * key, other available index, or a sequential scan if necessary. * + * 'relmapentry' is the relation map entry for 'localrel', as in + * FindReplTupleInLocalRel(). + * * Returns true if the deleted tuple is found. If found, the transaction ID, * origin, and commit timestamp of the deletion are stored in '*delete_xid', * '*delete_origin', and '*delete_time' respectively. */ static bool -FindDeletedTupleInLocalRel(Relation localrel, Oid localidxoid, +FindDeletedTupleInLocalRel(Relation localrel, + LogicalRepRelMapEntry *relmapentry, TupleTableSlot *remoteslot, TransactionId *delete_xid, ReplOriginId *delete_origin, TimestampTz *delete_time) { + Oid localidxoid = relmapentry->localindexoid; TransactionId oldestxmin; /* @@ -3383,6 +3404,7 @@ FindDeletedTupleInLocalRel(Relation localrel, Oid localidxoid, if (OidIsValid(localidxoid) && IsIndexUsableForFindingDeletedTuple(localidxoid, oldestxmin)) return RelationFindDeletedTupleInfoByIndex(localrel, localidxoid, + relmapentry->idxisreplident, remoteslot, oldestxmin, delete_xid, delete_origin, delete_time); @@ -3484,8 +3506,7 @@ apply_handle_tuple_routing(ApplyExecutionData *edata, case CMD_DELETE: apply_handle_delete_internal(edata, partrelinfo, - remoteslot_part, - part_entry->localindexoid); + remoteslot_part, part_entry); break; case CMD_UPDATE: @@ -3505,9 +3526,7 @@ apply_handle_tuple_routing(ApplyExecutionData *edata, ConflictTupleInfo conflicttuple = {0}; /* Get the matching local tuple from the partition. */ - found = FindReplTupleInLocalRel(edata, partrel, - &part_entry->remoterel, - part_entry->localindexoid, + found = FindReplTupleInLocalRel(edata, partrel, part_entry, remoteslot_part, &localslot); if (!found) { @@ -3519,8 +3538,7 @@ apply_handle_tuple_routing(ApplyExecutionData *edata, * never existed is crucial to avoid misleading the user * during conflict handling. */ - if (FindDeletedTupleInLocalRel(partrel, - part_entry->localindexoid, + if (FindDeletedTupleInLocalRel(partrel, part_entry, remoteslot_part, &conflicttuple.xmin, &conflicttuple.origin, diff --git a/src/include/executor/executor.h b/src/include/executor/executor.h index 8bb6c7bda2f..19132fb0d23 100644 --- a/src/include/executor/executor.h +++ b/src/include/executor/executor.h @@ -775,6 +775,7 @@ extern void check_exclusion_constraint(Relation heap, Relation index, * prototypes from functions in execReplication.c */ extern bool RelationFindReplTupleByIndex(Relation rel, Oid idxoid, + bool isIdxSafeToSkipDuplicates, LockTupleMode lockmode, TupleTableSlot *searchslot, TupleTableSlot *outslot); @@ -787,6 +788,7 @@ extern bool RelationFindDeletedTupleInfoSeq(Relation rel, ReplOriginId *delete_origin, TimestampTz *delete_time); extern bool RelationFindDeletedTupleInfoByIndex(Relation rel, Oid idxoid, + bool isIdxSafeToSkipDuplicates, TupleTableSlot *searchslot, TransactionId oldestxmin, TransactionId *delete_xid, diff --git a/src/include/replication/logicalrelation.h b/src/include/replication/logicalrelation.h index efe0f9d6031..4063118d2b9 100644 --- a/src/include/replication/logicalrelation.h +++ b/src/include/replication/logicalrelation.h @@ -33,6 +33,10 @@ typedef struct LogicalRepRelMapEntry AttrMap *attrmap; /* map of local attributes to remote ones */ bool updatable; /* Can apply updates/deletes? */ Oid localindexoid; /* which index to use, or InvalidOid if none */ + bool idxisreplident; /* is it the relation's replica identity or + * primary key, rather than an index usable + * for a REPLICA IDENTITY FULL remote + * relation? */ /* Sync state. */ char state; diff --git a/src/test/subscription/t/032_subscribe_use_index.pl b/src/test/subscription/t/032_subscribe_use_index.pl index 1ccd36ac227..21491c68528 100644 --- a/src/test/subscription/t/032_subscribe_use_index.pl +++ b/src/test/subscription/t/032_subscribe_use_index.pl @@ -606,6 +606,218 @@ $node_subscriber->safe_psql('postgres', "DROP TABLE test_replica_id_full"); # Testcase end: Subscription can use hash index # ============================================================================= +# ============================================================================= +# Testcase start: Subscription keeps using an index that concurrent DDL has +# demoted from replica identity +# +# REINDEX CONCURRENTLY and DROP INDEX CONCURRENTLY hold locks that do not +# conflict with the apply worker's, so either can commit after the worker has +# taken the index out of its relation map entry. The index goes on finding +# the row, so the change must still be applied through it, not dropped as a +# missing-tuple conflict. + +SKIP: +{ + skip 'Injection points not supported by this build', 8 + unless $ENV{enable_injection_points} eq 'yes'; + skip 'Extension injection_points not installed', 8 + unless $node_subscriber->check_extension('injection_points'); + + $node_subscriber->safe_psql('postgres', + 'CREATE EXTENSION injection_points'); + + # create tables pub and sub + $node_publisher->safe_psql('postgres', + "CREATE TABLE test_reindex (x int PRIMARY KEY, y int)"); + $node_subscriber->safe_psql('postgres', + "CREATE TABLE test_reindex (x int PRIMARY KEY, y int)"); + + # insert some initial data + $node_publisher->safe_psql('postgres', + "INSERT INTO test_reindex SELECT i, i FROM generate_series(1,20) i"); + + # create pub/sub + $node_publisher->safe_psql('postgres', + "CREATE PUBLICATION tap_pub_reindex FOR TABLE test_reindex"); + $node_subscriber->safe_psql('postgres', + "CREATE SUBSCRIPTION tap_sub_reindex CONNECTION '$publisher_connstr application_name=$appname' PUBLICATION tap_pub_reindex" + ); + + # wait for initial table synchronization to finish + $node_subscriber->wait_for_subscription_sync($node_publisher, $appname); + + my $old_index_oid = $node_subscriber->safe_psql('postgres', + "SELECT 'test_reindex_pkey'::regclass::oid"); + + # The rebuild goes first and stops at the swap. The other order does + # not work: REINDEX CONCURRENTLY waits for older snapshots, so a worker + # paused mid-transaction would block it rather than race it. + my $reindex = $node_subscriber->background_psql('postgres'); + $reindex->query_safe( + q[ + SELECT injection_points_set_local(); + SELECT injection_points_attach('reindex-relation-concurrently-before-swap', 'wait'); + ]); + $reindex->query_until( + qr/starting_reindex/, q[ + \echo starting_reindex + REINDEX INDEX CONCURRENTLY test_reindex_pkey; + ]); + $node_subscriber->wait_for_event('client backend', + 'reindex-relation-concurrently-before-swap'); + + # Now let the worker take the index and stop before opening the + # relation's indexes. The point is attached server-wide: the apply + # worker is not a session this test can attach anything in. + $node_subscriber->safe_psql('postgres', + "SELECT injection_points_attach('apply-update-before-open-indices', 'wait')" + ); + + $node_publisher->safe_psql('postgres', + "UPDATE test_reindex SET y = 99 WHERE x = 7"); + + $node_subscriber->wait_for_event( + 'logical replication apply worker', + 'apply-update-before-open-indices'); + + # Let the rebuild swap the index. Only the swap is needed: dropping the + # old index waits for the apply worker's lock anyway. + $node_subscriber->safe_psql('postgres', + "SELECT injection_points_wakeup('reindex-relation-concurrently-before-swap')" + ); + $node_subscriber->poll_query_until('postgres', + "SELECT 'test_reindex_pkey'::regclass::oid <> $old_index_oid") + or die "timed out waiting for the identity index to be swapped"; + + # Release the worker. The index it holds is no longer the identity. + $node_subscriber->safe_psql( + 'postgres', + "SELECT injection_points_wakeup('apply-update-before-open-indices'); + SELECT injection_points_detach('apply-update-before-open-indices');" + ); + ok($reindex->quit, 'REINDEX CONCURRENTLY completes'); + + $node_publisher->wait_for_catchup($appname); + + # The update must have been applied, not dropped as a missing tuple. + $result = $node_subscriber->safe_psql('postgres', + "SELECT y FROM test_reindex WHERE x = 7"); + is($result, qq(99), 'update applied through the index in force now'); + + # And the worker is still alive to apply the next one. + $node_publisher->safe_psql('postgres', + "UPDATE test_reindex SET y = 123 WHERE x = 8"); + $node_publisher->wait_for_catchup($appname); + $result = $node_subscriber->safe_psql('postgres', + "SELECT y FROM test_reindex WHERE x = 8"); + is($result, qq(123), 'replication continues'); + + # The same window without a rebuild: DROP INDEX CONCURRENTLY clears + # indisvalid and indisreplident and commits that before waiting for the + # lock apply holds, leaving relreplident set to 'i' with no index + # claiming to be that identity. The drop gets no further while apply + # holds the table, so the index is still complete and maintained. + $node_publisher->safe_psql( + 'postgres', q[ + CREATE TABLE test_dropri (x int NOT NULL, y int); + CREATE UNIQUE INDEX test_dropri_ri ON test_dropri (x); + ALTER TABLE test_dropri REPLICA IDENTITY USING INDEX test_dropri_ri; + INSERT INTO test_dropri SELECT i, i FROM generate_series(1,20) i; + CREATE PUBLICATION tap_pub_dropri FOR TABLE test_dropri; + ]); + $node_subscriber->safe_psql( + 'postgres', q[ + CREATE TABLE test_dropri (x int NOT NULL, y int); + CREATE UNIQUE INDEX test_dropri_ri ON test_dropri (x); + ALTER TABLE test_dropri REPLICA IDENTITY USING INDEX test_dropri_ri; + ]); + $node_subscriber->safe_psql('postgres', + "CREATE SUBSCRIPTION tap_sub_dropri CONNECTION '$publisher_connstr application_name=dropri' PUBLICATION tap_pub_dropri" + ); + $node_subscriber->wait_for_subscription_sync($node_publisher, 'dropri'); + + $node_subscriber->safe_psql('postgres', + "SELECT injection_points_attach('apply-update-before-open-indices', 'wait')" + ); + $node_publisher->safe_psql('postgres', + "UPDATE test_dropri SET y = 99 WHERE x = 7"); + $node_subscriber->wait_for_event( + 'logical replication apply worker', + 'apply-update-before-open-indices'); + + # This commits the loss of the replica identity, then parks waiting for + # the apply worker's lock on the table. + my $log_offset = -s $node_subscriber->logfile; + my $drop = $node_subscriber->background_psql('postgres'); + $drop->query_until( + qr/starting_drop/, q[ + \echo starting_drop + DROP INDEX CONCURRENTLY test_dropri_ri; + ]); + $node_subscriber->poll_query_until('postgres', + "SELECT count(*) = 0 FROM pg_index" + . " WHERE indrelid = 'test_dropri'::regclass AND indisreplident") + or die "timed out waiting for the identity index to be invalidated"; + + $node_subscriber->safe_psql( + 'postgres', + "SELECT injection_points_detach('apply-update-before-open-indices'); + SELECT injection_points_wakeup('apply-update-before-open-indices');" + ); + # The straddling change goes through, found by the demoted index. + $node_publisher->wait_for_catchup('dropri'); + $result = $node_subscriber->safe_psql('postgres', + "SELECT y FROM test_dropri WHERE x = 7"); + is($result, qq(99), 'change straddling the drop is applied'); + ok($drop->quit, 'DROP INDEX CONCURRENTLY completes'); + + # From the next change on the entry is rebuilt, finds no replica + # identity, and apply stops with the usual error. + $node_publisher->safe_psql('postgres', + "UPDATE test_dropri SET y = 123 WHERE x = 8"); + ok( $node_subscriber->poll_query_until( + 'postgres', q[ + SELECT apply_error_count > 0 FROM pg_stat_subscription_stats + WHERE subname = 'tap_sub_dropri']), + 'later changes wait for a replica identity'); + like( + slurp_file($node_subscriber->logfile, $log_offset), + qr/logical replication target relation "public\.test_dropri" has neither REPLICA IDENTITY index nor PRIMARY KEY/, + 'and say why'); + + # Give the relation a replica identity again and they resume. + $node_subscriber->safe_psql( + 'postgres', q[ + CREATE UNIQUE INDEX test_dropri_ri2 ON test_dropri (x); + ALTER TABLE test_dropri REPLICA IDENTITY USING INDEX test_dropri_ri2; + ]); + $node_publisher->wait_for_catchup('dropri'); + $result = $node_subscriber->safe_psql('postgres', + "SELECT y FROM test_dropri WHERE x = 8"); + is($result, qq(123), 'replication resumes'); + + $node_publisher->safe_psql('postgres', "DROP PUBLICATION tap_pub_dropri"); + $node_publisher->safe_psql('postgres', "DROP TABLE test_dropri"); + $node_subscriber->safe_psql('postgres', + "DROP SUBSCRIPTION tap_sub_dropri"); + $node_subscriber->safe_psql('postgres', "DROP TABLE test_dropri"); + + # cleanup pub + $node_publisher->safe_psql('postgres', + "DROP PUBLICATION tap_pub_reindex"); + $node_publisher->safe_psql('postgres', "DROP TABLE test_reindex"); + # cleanup sub + $node_subscriber->safe_psql('postgres', + "DROP SUBSCRIPTION tap_sub_reindex"); + $node_subscriber->safe_psql('postgres', "DROP TABLE test_reindex"); + $node_subscriber->safe_psql('postgres', + "DROP EXTENSION injection_points"); +} + +# Testcase end: Subscription keeps using an index that concurrent DDL has +# demoted from replica identity +# ============================================================================= + $node_subscriber->stop('fast'); $node_publisher->stop('fast'); -- 2.54.0