Refresh snapshot periodically during index validation

Enhances validation phase of concurrently built indexes by periodically refreshing snapshots rather than using a single reference snapshot. This addresses issues with xmin propagation during long-running validations.

The validation now takes a fresh snapshot every few pages, allowing the xmin horizon to advance. This restores feature of commit d9d0762, which was reverted in commit e28bb88. New STIR-based approach is not depends on single reference snapshot anymore.

Author: michail-nikolaev

doc/src/sgml/ref/create_index.sgml

@@ -881,9 +881,14 @@ Indexes:
   </para>
 
   <para>
-   Like any long-running transaction, <command>CREATE INDEX</command> on a
-   table can affect which tuples can be removed by concurrent
-   <command>VACUUM</command> on any other table.
+   Due to the improved implementation using periodically refreshed snapshots and
+   auxiliary indexes, concurrent index builds have minimal impact on concurrent
+   <command>VACUUM</command> operations. The system automatically advances its
+   internal transaction horizon during the build process, allowing
+   <command>VACUUM</command> to remove dead tuples on other tables without
+   having to wait for the entire index build to complete. Only during very brief
+   periods when snapshots are being refreshed might there be any temporary effect
+   on concurrent <command>VACUUM</command> operations.
   </para>
 
   <para>

doc/src/sgml/ref/reindex.sgml

@@ -502,10 +502,13 @@ Indexes:
    </para>
 
    <para>
-    Like any long-running transaction, <command>REINDEX</command> on a table
-    can affect which tuples can be removed by concurrent
-    <command>VACUUM</command> on any other table.
-   </para>
+    <command>REINDEX CONCURRENTLY</command> has minimal
+    impact on which tuples can be removed by concurrent <command>VACUUM</command>
+    operations on other tables. This is achieved through periodic snapshot
+    refreshes and the use of auxiliary indexes during the rebuild process,
+    allowing the system to advance its transaction horizon regularly rather than
+    maintaining a single long-running snapshot.
+  </para>
 
    <para>
     <command>REINDEX SYSTEM</command> does not support

src/backend/access/heap/README.HOT

@@ -401,12 +401,12 @@ use the key value from the live tuple.
 We mark the index open for inserts (but still not ready for reads) then
 we again wait for transactions which have the table open.  Then validate
 the index.  This searches for tuples missing from the index in auxiliary
-index, and inserts any missing ones if them visible to reference snapshot.
+index, and inserts any missing ones if them visible to fresh snapshot.
 Again, the index entries have to have TIDs equal to HOT-chain root TIDs, but
 the value to be inserted is the one from the live tuple.
 
 Then we wait until every transaction that could have a snapshot older than
-the second reference snapshot is finished.  This ensures that nobody is
+the latest used snapshot is finished.  This ensures that nobody is
 alive any longer who could need to see any tuples that might be missing
 from the index, as well as ensuring that no one can see any inconsistent
 rows in a broken HOT chain (the first condition is stronger than the

src/backend/access/heap/heapam_handler.c

@@ -2034,23 +2034,26 @@ heapam_index_validate_scan_read_stream_next(
 	return result;
 }
 
-static void
+static TransactionId
 heapam_index_validate_scan(Relation heapRelation,
 						   Relation indexRelation,
 						   IndexInfo *indexInfo,
-						   Snapshot snapshot,
 						   ValidateIndexState *state,
 						   ValidateIndexState *auxState)
 {
+	TransactionId limitXmin;
+
 	Datum		values[INDEX_MAX_KEYS];
 	bool		isnull[INDEX_MAX_KEYS];
 
+	Snapshot		snapshot;
 	TupleTableSlot  *slot;
 	EState			*estate;
 	ExprContext		*econtext;
 	BufferAccessStrategy bstrategy = GetAccessStrategy(BAS_BULKREAD);
 
-	int				num_to_check;
+	int				num_to_check,
+					page_read_counter = 1; /* set to 1 to skip snapshot reset at start */
 	Tuplestorestate *tuples_for_check;
 	ValidateIndexScanState callback_private_data;
 
@@ -2061,14 +2064,16 @@ heapam_index_validate_scan(Relation heapRelation,
 	/* Use 10% of memory for tuple store. */
 	int		store_work_mem_part = maintenance_work_mem / 10;
 
-	/*
-	 * Encode TIDs as int8 values for the sort, rather than directly sorting
-	 * item pointers.  This can be significantly faster, primarily because TID
-	 * is a pass-by-reference type on all platforms, whereas int8 is
-	 * pass-by-value on most platforms.
-	 */
+	PushActiveSnapshot(GetTransactionSnapshot());
+
 	tuples_for_check =  tuplestore_begin_datum(INT8OID, false, false, store_work_mem_part);
 
+	PopActiveSnapshot();
+	InvalidateCatalogSnapshot();
+
+	Assert(!HaveRegisteredOrActiveSnapshot());
+	Assert(!TransactionIdIsValid(MyProc->xmin));
+
 	/*
 	 * sanity checks
 	 */
@@ -2084,6 +2089,29 @@ heapam_index_validate_scan(Relation heapRelation,
 
 	state->tuplesort = auxState->tuplesort = NULL;
 
+	/*
+	 * Now take the first snapshot that will be used by to filter candidate
+	 * tuples. We are going to replace it by newer snapshot every so often
+	 * to propagate horizon.
+	 *
+	 * Beware!  There might still be snapshots in use that treat some transaction
+	 * as in-progress that our temporary snapshot treats as committed.
+	 *
+	 * If such a recently-committed transaction deleted tuples in the table,
+	 * we will not include them in the index; yet those transactions which
+	 * see the deleting one as still-in-progress will expect such tuples to
+	 * be there once we mark the index as valid.
+	 *
+	 * We solve this by waiting for all endangered transactions to exit before
+	 * we mark the index as valid, for that reason limitXmin is supported.
+	 *
+	 * We also set ActiveSnapshot to this snap, since functions in indexes may
+	 * need a snapshot.
+	 */
+	snapshot = RegisterSnapshot(GetLatestSnapshot());
+	PushActiveSnapshot(snapshot);
+	limitXmin = snapshot->xmin;
+
 	estate = CreateExecutorState();
 	econtext = GetPerTupleExprContext(estate);
 	slot = MakeSingleTupleTableSlot(RelationGetDescr(heapRelation),
@@ -2117,6 +2145,7 @@ heapam_index_validate_scan(Relation heapRelation,
 
 		LockBuffer(buf, BUFFER_LOCK_SHARE);
 		block_number = BufferGetBlockNumber(buf);
+		page_read_counter++;
 
 		i = 0;
 		while ((off = tuples[i]) != InvalidOffsetNumber)
@@ -2172,6 +2201,20 @@ heapam_index_validate_scan(Relation heapRelation,
 		}
 
 		ReleaseBuffer(buf);
+#define VALIDATE_INDEX_RESET_SNAPSHOT_EACH_N_PAGE 4096
+		if (page_read_counter % VALIDATE_INDEX_RESET_SNAPSHOT_EACH_N_PAGE == 0)
+		{
+			PopActiveSnapshot();
+			UnregisterSnapshot(snapshot);
+			/* to make sure we propagate xmin */
+			InvalidateCatalogSnapshot();
+			Assert(!TransactionIdIsValid(MyProc->xmin));
+
+			snapshot = RegisterSnapshot(GetLatestSnapshot());
+			PushActiveSnapshot(snapshot);
+			/* xmin should not go backwards, but just for the case*/
+			limitXmin = TransactionIdNewer(limitXmin, snapshot->xmin);
+		}
 	}
 
 	ExecDropSingleTupleTableSlot(slot);
@@ -2181,9 +2224,21 @@ heapam_index_validate_scan(Relation heapRelation,
 	read_stream_end(read_stream);
 	tuplestore_end(tuples_for_check);
 
+	/*
+	 * Drop the latest snapshot.  We must do this before waiting out other
+	 * snapshot holders, else we will deadlock against other processes also
+	 * doing CREATE INDEX CONCURRENTLY, which would see our snapshot as one
+	 * they must wait for.
+	 */
+	PopActiveSnapshot();
+	UnregisterSnapshot(snapshot);
+	InvalidateCatalogSnapshot();
+	Assert(MyProc->xmin == InvalidTransactionId);
 	/* These may have been pointing to the now-gone estate */
 	indexInfo->ii_ExpressionsState = NIL;
 	indexInfo->ii_PredicateState = NULL;
+
+	return limitXmin;
 }
 
 /*

src/backend/access/nbtree/nbtsort.c

@@ -445,7 +445,7 @@ _bt_spools_heapscan(Relation heap, Relation index, BTBuildState *buildstate,
 	 * dead tuples) won't get very full, so we give it only work_mem.
 	 *
 	 * In case of concurrent build dead tuples are not need to be put into index
-	 * since we wait for all snapshots older than reference snapshot during the
+	 * since we wait for all snapshots older than latest snapshot during the
 	 * validation phase.
 	 */
 	if (indexInfo->ii_Unique && !indexInfo->ii_Concurrent)

src/backend/access/spgist/spgvacuum.c

@@ -191,14 +191,16 @@ vacuumLeafPage(spgBulkDeleteState *bds, Relation index, Buffer buffer,
 			 * Add target TID to pending list if the redirection could have
 			 * happened since VACUUM started.  (If xid is invalid, assume it
 			 * must have happened before VACUUM started, since REINDEX
-			 * CONCURRENTLY locks out VACUUM.)
+			 * CONCURRENTLY locks out VACUUM, if myXmin is invalid it is
+			 * validation scan.)
 			 *
 			 * Note: we could make a tighter test by seeing if the xid is
 			 * "running" according to the active snapshot; but snapmgr.c
 			 * doesn't currently export a suitable API, and it's not entirely
 			 * clear that a tighter test is worth the cycles anyway.
 			 */
-			if (TransactionIdFollowsOrEquals(dt->xid, bds->myXmin))
+			if (!TransactionIdIsValid(bds->myXmin) ||
+					TransactionIdFollowsOrEquals(dt->xid, bds->myXmin))
 				spgAddPendingTID(bds, &dt->pointer);
 		}
 		else
@@ -808,7 +810,6 @@ spgvacuumscan(spgBulkDeleteState *bds)
 	/* Finish setting up spgBulkDeleteState */
 	initSpGistState(&bds->spgstate, index);
 	bds->pendingList = NULL;
-	bds->myXmin = GetActiveSnapshot()->xmin;
 	bds->lastFilledBlock = SPGIST_LAST_FIXED_BLKNO;
 
 	/*
@@ -959,6 +960,10 @@ spgbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 	bds.stats = stats;
 	bds.callback = callback;
 	bds.callback_state = callback_state;
+	if (info->validate_index)
+		bds.myXmin = InvalidTransactionId;
+	else
+		bds.myXmin = GetActiveSnapshot()->xmin;
 
 	spgvacuumscan(&bds);
 
@@ -999,6 +1004,7 @@ spgvacuumcleanup(IndexVacuumInfo *info, IndexBulkDeleteResult *stats)
 		bds.stats = stats;
 		bds.callback = dummy_callback;
 		bds.callback_state = NULL;
+		bds.myXmin = GetActiveSnapshot()->xmin;
 
 		spgvacuumscan(&bds);
 	}

src/backend/catalog/index.c

@@ -3536,8 +3536,9 @@ IndexCheckExclusion(Relation heapRelation,
  * insert their new tuples into it. At the same moment we clear "indisready" for
  * auxiliary index, since it is no more required to be updated.
  *
- * We then take a new reference snapshot, any tuples that are valid according
- * to this snap, but are not in the index, must be added to the index.
+ * We then take a new snapshot, any tuples that are valid according
+ * to this snap, but are not in the index, must be added to the index. In
+ * order to propagate xmin we reset that snapshot every few so often.
  * (Any tuples committed live after the snap will be inserted into the
  * index by their originating transaction.  Any tuples committed dead before
  * the snap need not be indexed, because we will wait out all transactions
@@ -3550,7 +3551,7 @@ IndexCheckExclusion(Relation heapRelation,
  * TIDs of both auxiliary and target indexes, and doing a "merge join" against
  * the TID lists to see which tuples from auxiliary index are missing from the
  * target index.  Thus we will ensure that all tuples valid according to the
- * reference snapshot are in the index. Notice we need to do bulkdelete in the
+ * latest snapshot are in the index. Notice we need to do bulkdelete in the
  * particular order: auxiliary first, target last.
  *
  * Building a unique index this way is tricky: we might try to insert a
@@ -3571,13 +3572,14 @@ IndexCheckExclusion(Relation heapRelation,
  *
  * Also, some actions to concurrent drop the auxiliary index are performed.
  */
-void
-validate_index(Oid heapId, Oid indexId, Oid auxIndexId, Snapshot snapshot)
+TransactionId
+validate_index(Oid heapId, Oid indexId, Oid auxIndexId)
 {
 	Relation	heapRelation,
 				indexRelation,
 				auxIndexRelation;
 	IndexInfo  *indexInfo;
+	TransactionId limitXmin;
 	IndexVacuumInfo ivinfo, auxivinfo;
 	ValidateIndexState state, auxState;
 	Oid			save_userid;
@@ -3627,8 +3629,12 @@ validate_index(Oid heapId, Oid indexId, Oid auxIndexId, Snapshot snapshot)
 	 * Fetch info needed for index_insert.  (You might think this should be
 	 * passed in from DefineIndex, but its copy is long gone due to having
 	 * been built in a previous transaction.)
+	 *
+	 * We might need snapshot for index expressions or predicates.
 	 */
+	PushActiveSnapshot(GetTransactionSnapshot());
 	indexInfo = BuildIndexInfo(indexRelation);
+	PopActiveSnapshot();
 
 	/* mark build is concurrent just for consistency */
 	indexInfo->ii_Concurrent = true;
@@ -3664,6 +3670,9 @@ validate_index(Oid heapId, Oid indexId, Oid auxIndexId, Snapshot snapshot)
 										   NULL, TUPLESORT_NONE);
 	auxState.htups = auxState.itups = auxState.tups_inserted = 0;
 
+	/* tuplesort_begin_datum may require catalog snapshot */
+	InvalidateCatalogSnapshot();
+
 	(void) index_bulk_delete(&auxivinfo, NULL,
 							 validate_index_callback, &auxState);
 	/* If aux index is empty, merge may be skipped */
@@ -3698,6 +3707,9 @@ validate_index(Oid heapId, Oid indexId, Oid auxIndexId, Snapshot snapshot)
 											NULL, TUPLESORT_NONE);
 	state.htups = state.itups = state.tups_inserted = 0;
 
+	/* tuplesort_begin_datum may require catalog snapshot */
+	InvalidateCatalogSnapshot();
+
 	/* ambulkdelete updates progress metrics */
 	(void) index_bulk_delete(&ivinfo, NULL,
 							 validate_index_callback, &state);
@@ -3717,19 +3729,24 @@ validate_index(Oid heapId, Oid indexId, Oid auxIndexId, Snapshot snapshot)
 		pgstat_progress_update_multi_param(3, progress_index, progress_vals);
 	}
 	tuplesort_performsort(state.tuplesort);
+	/* tuplesort_performsort may require catalog snapshot */
+	InvalidateCatalogSnapshot();
+
 	tuplesort_performsort(auxState.tuplesort);
+	/* tuplesort_performsort may require catalog snapshot */
+	InvalidateCatalogSnapshot();
+	Assert(!TransactionIdIsValid(MyProc->xmin));
 
 	/*
 	 * Now merge both indexes
 	 */
 	pgstat_progress_update_param(PROGRESS_CREATEIDX_PHASE,
 								 PROGRESS_CREATEIDX_PHASE_VALIDATE_IDXMERGE);
-	table_index_validate_scan(heapRelation,
-							  indexRelation,
-							  indexInfo,
-							  snapshot,
-							  &state,
-							  &auxState);
+	limitXmin = table_index_validate_scan(heapRelation,
+										  indexRelation,
+										  indexInfo,
+										  &state,
+										  &auxState);
 
 	/* Tuple sort closed by table_index_validate_scan */
 	Assert(state.tuplesort == NULL && auxState.tuplesort == NULL);
@@ -3752,6 +3769,9 @@ validate_index(Oid heapId, Oid indexId, Oid auxIndexId, Snapshot snapshot)
 	index_close(auxIndexRelation, NoLock);
 	index_close(indexRelation, NoLock);
 	table_close(heapRelation, NoLock);
+
+	Assert(!TransactionIdIsValid(MyProc->xmin));
+	return limitXmin;
 }
 
 /*