Fixed locking for non-MVCC snapshots

This reverts commit dbef8d1ad1.

CHANGELOG.md

@@ -1,3 +1,8 @@
+## 0.5.1 (unreleased)
+
+- Improved performance of index scans for IVFFlat after updates and deletes
+- Fixed locking for index scans for HNSW
+
 ## 0.5.0 (2023-08-28)
 
 - Added HNSW index type

Makefile

@@ -8,7 +8,7 @@ HEADERS = src/vector.h
 
 TESTS = $(wildcard test/sql/*.sql)
 REGRESS = $(patsubst test/sql/%.sql,%,$(TESTS))
-REGRESS_OPTS = --inputdir=test --load-extension=vector
+REGRESS_OPTS = --inputdir=test --load-extension=$(EXTENSION)
 
 OPTFLAGS = -march=native
 

Makefile.win

@@ -5,7 +5,7 @@ OBJS = src\hnsw.obj src\hnswbuild.obj src\hnswinsert.obj src\hnswscan.obj src\hn
 HEADERS = src\vector.h
 
 REGRESS = btree cast copy functions input ivfflat_cosine ivfflat_ip ivfflat_l2 ivfflat_options ivfflat_unlogged
-REGRESS_OPTS = --inputdir=test --load-extension=vector
+REGRESS_OPTS = --inputdir=test --load-extension=$(EXTENSION)
 
 # For /arch flags
 # https://learn.microsoft.com/en-us/cpp/build/reference/arch-minimum-cpu-architecture

README.md

@@ -595,12 +595,18 @@ pgvector is available on [these providers](https://github.com/pgvector/pgvector/
 
 ## Upgrading
 
-Install the latest version and run:
+Install the latest version. Then in each database you want to upgrade, run:
 
 ```sql
 ALTER EXTENSION vector UPDATE;
 ```
 
+You can check the version in the current database with:
+
+```sql
+SELECT extversion FROM pg_extension WHERE extname = 'vector';
+```
+
 ## Upgrade Notes
 
 ### 0.4.0

src/hnsw.c

@@ -91,7 +91,7 @@ hnswcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 	MemSet(&costs, 0, sizeof(costs));
 
 	index = index_open(path->indexinfo->indexoid, NoLock);
-	m = HnswGetM(index);
+	HnswGetMetaPageInfo(index, &m, NULL);
 	index_close(index, NoLock);
 
 	/* Approximate entry level */
@@ -196,7 +196,7 @@ hnswhandler(PG_FUNCTION_ARGS)
 	amroutine->aminsert = hnswinsert;
 	amroutine->ambulkdelete = hnswbulkdelete;
 	amroutine->amvacuumcleanup = hnswvacuumcleanup;
-	amroutine->amcanreturn = NULL;	/* tuple not included in heapsort */
+	amroutine->amcanreturn = NULL;
 	amroutine->amcostestimate = hnswcostestimate;
 	amroutine->amoptions = hnswoptions;
 	amroutine->amproperty = NULL;	/* TODO AMPROP_DISTANCE_ORDERABLE */

src/hnsw.h

@@ -4,7 +4,6 @@
 #include "postgres.h"
 
 #include "access/generic_xlog.h"
-#include "access/parallel.h"
 #include "access/reloptions.h"
 #include "nodes/execnodes.h"
 #include "port.h"				/* for random() */
@@ -15,10 +14,6 @@
 #error "Requires PostgreSQL 11+"
 #endif
 
-#if PG_VERSION_NUM < 120000
-#include "access/relscan.h"
-#endif
-
 #define HNSW_MAX_DIM 2000
 
 /* Support functions */
@@ -137,49 +132,6 @@ typedef struct HnswOptions
 	int			efConstruction; /* size of dynamic candidate list */
 }			HnswOptions;
 
-typedef struct HnswSpool
-{
-	Relation	heap;
-	Relation	index;
-}			HnswSpool;
-
-typedef struct HnswShared
-{
-	/* Immutable state */
-	Oid			heaprelid;
-	Oid			indexrelid;
-	bool		isconcurrent;
-	int			scantuplesortstates;
-
-	/* Worker progress */
-	ConditionVariable workersdonecv;
-
-	/* Mutex for mutable state */
-	slock_t		mutex;
-
-	/* Mutable state */
-	int			nparticipantsdone;
-	double		reltuples;
-	double		indtuples;
-
-#if PG_VERSION_NUM < 120000
-	ParallelHeapScanDescData heapdesc;	/* must come last */
-#endif
-}			HnswShared;
-
-#if PG_VERSION_NUM >= 120000
-#define ParallelTableScanFromHnswShared(shared) \
-	(ParallelTableScanDesc) ((char *) (shared) + BUFFERALIGN(sizeof(HnswShared)))
-#endif
-
-typedef struct HnswLeader
-{
-	ParallelContext *pcxt;
-	int			nparticipanttuplesorts;
-	HnswShared *hnswshared;
-	Snapshot	snapshot;
-}			HnswLeader;
-
 typedef struct HnswBuildState
 {
 	/* Info */
@@ -213,9 +165,6 @@ typedef struct HnswBuildState
 
 	/* Memory */
 	MemoryContext tmpCtx;
-
-	/* Parallel builds */
-	HnswLeader *hnswleader;
 }			HnswBuildState;
 
 typedef struct HnswMetaPageData
@@ -269,7 +218,6 @@ typedef HnswNeighborTupleData * HnswNeighborTuple;
 typedef struct HnswScanOpaqueData
 {
 	bool		first;
-	Buffer		buf;
 	List	   *w;
 	MemoryContext tmpCtx;
 
@@ -317,8 +265,9 @@ Buffer		HnswNewBuffer(Relation index, ForkNumber forkNum);
 void		HnswInitPage(Buffer buf, Page page);
 void		HnswInitRegisterPage(Relation index, Buffer *buf, Page *page, GenericXLogState **state);
 void		HnswInit(void);
-List	   *HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, bool inserting, HnswElement skipElement);
+List	   *HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement);
 HnswElement HnswGetEntryPoint(Relation index);
+void		HnswGetMetaPageInfo(Relation index, int *m, HnswElement * entryPoint);
 HnswElement HnswInitElement(ItemPointer tid, int m, double ml, int maxLevel);
 void		HnswFreeElement(HnswElement element);
 HnswElement HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno);
@@ -335,8 +284,7 @@ void		HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool
 void		HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index, FmgrInfo *procinfo, Oid collation, bool loadVec);
 void		HnswSetElementTuple(HnswElementTuple etup, HnswElement element);
 void		HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int *updateIdx, Relation index, FmgrInfo *procinfo, Oid collation);
-void		HnswLoadNeighbors(HnswElement element, Relation index);
-PGDLLEXPORT void HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc);
+void		HnswLoadNeighbors(HnswElement element, Relation index, int m);
 
 /* Index access methods */
 IndexBuildResult *hnswbuild(Relation heap, Relation index, IndexInfo *indexInfo);

src/hnswbuild.c

@@ -2,15 +2,12 @@
 
 #include <math.h>
 
-#include "access/parallel.h"
-#include "access/xact.h"
 #include "catalog/index.h"
 #include "hnsw.h"
 #include "miscadmin.h"
 #include "lib/pairingheap.h"
 #include "nodes/pg_list.h"
 #include "storage/bufmgr.h"
-#include "tcop/tcopprot.h"
 #include "utils/memutils.h"
 
 #if PG_VERSION_NUM >= 140000
@@ -38,23 +35,6 @@
 #define UpdateProgress(index, val) ((void)val)
 #endif
 
-#if PG_VERSION_NUM >= 140000
-#include "utils/backend_status.h"
-#include "utils/wait_event.h"
-#endif
-
-#if PG_VERSION_NUM >= 120000
-#include "access/table.h"
-#include "optimizer/optimizer.h"
-#else
-#include "access/heapam.h"
-#include "optimizer/planner.h"
-#include "pgstat.h"
-#endif
-
-#define PARALLEL_KEY_HNSW_SHARED		UINT64CONST(0xA000000000000001)
-#define PARALLEL_KEY_QUERY_TEXT			UINT64CONST(0xA000000000000002)
-
 /*
  * Create the metapage
  */
@@ -371,7 +351,6 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 
 		oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
 
-		/* TODO Fix progress for parallel builds */
 		if (HnswInsertTuple(buildstate->index, values, isnull, tid, buildstate->heap))
 			UpdateProgress(PROGRESS_CREATEIDX_TUPLES_DONE, ++buildstate->indtuples);
 
@@ -469,8 +448,6 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 	buildstate->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
 											   "Hnsw build temporary context",
 											   ALLOCSET_DEFAULT_SIZES);
-
-	buildstate->hnswleader = NULL;
 }
 
 /*
@@ -483,373 +460,21 @@ FreeBuildState(HnswBuildState * buildstate)
 	MemoryContextDelete(buildstate->tmpCtx);
 }
 
-/*
- * Within leader, wait for end of heap scan
- */
-static double
-ParallelHeapScan(HnswBuildState * buildstate)
-{
-	HnswShared *hnswshared = buildstate->hnswleader->hnswshared;
-	int			nparticipanttuplesorts;
-	double		reltuples;
-
-	nparticipanttuplesorts = buildstate->hnswleader->nparticipanttuplesorts;
-	for (;;)
-	{
-		SpinLockAcquire(&hnswshared->mutex);
-		if (hnswshared->nparticipantsdone == nparticipanttuplesorts)
-		{
-			buildstate->indtuples = hnswshared->indtuples;
-			reltuples = hnswshared->reltuples;
-			SpinLockRelease(&hnswshared->mutex);
-			break;
-		}
-		SpinLockRelease(&hnswshared->mutex);
-
-		ConditionVariableSleep(&hnswshared->workersdonecv,
-							   WAIT_EVENT_PARALLEL_CREATE_INDEX_SCAN);
-	}
-
-	ConditionVariableCancelSleep();
-
-	return reltuples;
-}
-
-/*
- * Perform a worker's portion of a parallel insert
- */
-static void
-HnswParallelScanAndInsert(HnswSpool * hnswspool, HnswShared * hnswshared, bool progress)
-{
-	HnswBuildState buildstate;
-#if PG_VERSION_NUM >= 120000
-	TableScanDesc scan;
-#else
-	HeapScanDesc scan;
-#endif
-	double		reltuples;
-	IndexInfo  *indexInfo;
-
-	/* Join parallel scan */
-	indexInfo = BuildIndexInfo(hnswspool->index);
-	indexInfo->ii_Concurrent = hnswshared->isconcurrent;
-	InitBuildState(&buildstate, hnswspool->heap, hnswspool->index, indexInfo, MAIN_FORKNUM);
-	/* TODO Support in-memory builds */
-	buildstate.maxInMemoryElements = 0;
-	buildstate.flushed = true;
-#if PG_VERSION_NUM >= 120000
-	scan = table_beginscan_parallel(hnswspool->heap,
-									ParallelTableScanFromHnswShared(hnswshared));
-	reltuples = table_index_build_scan(hnswspool->heap, hnswspool->index, indexInfo,
-									   true, progress, BuildCallback,
-									   (void *) &buildstate, scan);
-#else
-	scan = heap_beginscan_parallel(hnswspool->heap, &hnswshared->heapdesc);
-	reltuples = IndexBuildHeapScan(hnswspool->heap, hnswspool->index, indexInfo,
-								   true, BuildCallback,
-								   (void *) &buildstate, scan);
-#endif
-
-	/* Record statistics */
-	SpinLockAcquire(&hnswshared->mutex);
-	hnswshared->nparticipantsdone++;
-	hnswshared->reltuples += reltuples;
-	hnswshared->indtuples += buildstate.indtuples;
-	SpinLockRelease(&hnswshared->mutex);
-
-	/* Log statistics */
-	if (progress)
-		ereport(DEBUG1, (errmsg("leader processed " INT64_FORMAT " tuples", (int64) reltuples)));
-	else
-		ereport(DEBUG1, (errmsg("worker processed " INT64_FORMAT " tuples", (int64) reltuples)));
-
-	/* Notify leader */
-	ConditionVariableSignal(&hnswshared->workersdonecv);
-
-	FreeBuildState(&buildstate);
-}
-
-/*
- * Perform work within a launched parallel process
- */
-void
-HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc)
-{
-	char	   *sharedquery;
-	HnswSpool  *hnswspool;
-	HnswShared *hnswshared;
-	Relation	heapRel;
-	Relation	indexRel;
-	LOCKMODE	heapLockmode;
-	LOCKMODE	indexLockmode;
-
-	/* Set debug_query_string for individual workers first */
-	sharedquery = shm_toc_lookup(toc, PARALLEL_KEY_QUERY_TEXT, true);
-	debug_query_string = sharedquery;
-
-	/* Report the query string from leader */
-	pgstat_report_activity(STATE_RUNNING, debug_query_string);
-
-	/* Look up shared state */
-	hnswshared = shm_toc_lookup(toc, PARALLEL_KEY_HNSW_SHARED, false);
-
-	/* Open relations using lock modes known to be obtained by index.c */
-	if (!hnswshared->isconcurrent)
-	{
-		heapLockmode = ShareLock;
-		indexLockmode = AccessExclusiveLock;
-	}
-	else
-	{
-		heapLockmode = ShareUpdateExclusiveLock;
-		indexLockmode = RowExclusiveLock;
-	}
-
-	/* Open relations within worker */
-#if PG_VERSION_NUM >= 120000
-	heapRel = table_open(hnswshared->heaprelid, heapLockmode);
-#else
-	heapRel = heap_open(hnswshared->heaprelid, heapLockmode);
-#endif
-	indexRel = index_open(hnswshared->indexrelid, indexLockmode);
-
-	/* Initialize worker's own spool */
-	hnswspool = (HnswSpool *) palloc0(sizeof(HnswSpool));
-	hnswspool->heap = heapRel;
-	hnswspool->index = indexRel;
-
-	/* Perform inserts */
-	HnswParallelScanAndInsert(hnswspool, hnswshared, false);
-
-	/* Close relations within worker */
-	index_close(indexRel, indexLockmode);
-#if PG_VERSION_NUM >= 120000
-	table_close(heapRel, heapLockmode);
-#else
-	heap_close(heapRel, heapLockmode);
-#endif
-}
-
-/*
- * End parallel build
- */
-static void
-HnswEndParallel(HnswLeader * hnswleader)
-{
-	/* Shutdown worker processes */
-	WaitForParallelWorkersToFinish(hnswleader->pcxt);
-
-	/* Free last reference to MVCC snapshot, if one was used */
-	if (IsMVCCSnapshot(hnswleader->snapshot))
-		UnregisterSnapshot(hnswleader->snapshot);
-	DestroyParallelContext(hnswleader->pcxt);
-	ExitParallelMode();
-}
-
-/*
- * Return size of shared memory required for parallel index build
- */
-static Size
-ParallelEstimateShared(Relation heap, Snapshot snapshot)
-{
-#if PG_VERSION_NUM >= 120000
-	return add_size(BUFFERALIGN(sizeof(HnswShared)), table_parallelscan_estimate(heap, snapshot));
-#else
-	if (!IsMVCCSnapshot(snapshot))
-	{
-		Assert(snapshot == SnapshotAny);
-		return sizeof(HnswShared);
-	}
-
-	return add_size(offsetof(HnswShared, heapdesc) +
-					offsetof(ParallelHeapScanDescData, phs_snapshot_data),
-					EstimateSnapshotSpace(snapshot));
-#endif
-}
-
-/*
- * Within leader, participate as a parallel worker
- */
-static void
-HnswLeaderParticipateAsWorker(HnswBuildState * buildstate)
-{
-	HnswLeader *hnswleader = buildstate->hnswleader;
-	HnswSpool  *leaderworker;
-
-	/* Allocate memory and initialize private spool */
-	leaderworker = (HnswSpool *) palloc0(sizeof(HnswSpool));
-	leaderworker->heap = buildstate->heap;
-	leaderworker->index = buildstate->index;
-
-	/* Perform work common to all participants */
-	HnswParallelScanAndInsert(leaderworker, hnswleader->hnswshared, true);
-}
-
-/*
- * Begin parallel build
- */
-static void
-HnswBeginParallel(HnswBuildState * buildstate, bool isconcurrent, int request)
-{
-	ParallelContext *pcxt;
-	int			scantuplesortstates;
-	Snapshot	snapshot;
-	Size		esthnswshared;
-	HnswShared *hnswshared;
-	HnswLeader *hnswleader = (HnswLeader *) palloc0(sizeof(HnswLeader));
-	bool		leaderparticipates = true;
-	int			querylen;
-
-#ifdef DISABLE_LEADER_PARTICIPATION
-	leaderparticipates = false;
-#endif
-
-	/* Enter parallel mode and create context */
-	EnterParallelMode();
-	Assert(request > 0);
-#if PG_VERSION_NUM >= 120000
-	pcxt = CreateParallelContext("vector", "HnswParallelBuildMain", request);
-#else
-	pcxt = CreateParallelContext("vector", "HnswParallelBuildMain", request, true);
-#endif
-
-	scantuplesortstates = leaderparticipates ? request + 1 : request;
-
-	/* Get snapshot for table scan */
-	if (!isconcurrent)
-		snapshot = SnapshotAny;
-	else
-		snapshot = RegisterSnapshot(GetTransactionSnapshot());
-
-	/* Estimate size of workspaces */
-	esthnswshared = ParallelEstimateShared(buildstate->heap, snapshot);
-	shm_toc_estimate_chunk(&pcxt->estimator, esthnswshared);
-	shm_toc_estimate_keys(&pcxt->estimator, 1);
-
-	/* Finally, estimate PARALLEL_KEY_QUERY_TEXT space */
-	if (debug_query_string)
-	{
-		querylen = strlen(debug_query_string);
-		shm_toc_estimate_chunk(&pcxt->estimator, querylen + 1);
-		shm_toc_estimate_keys(&pcxt->estimator, 1);
-	}
-	else
-		querylen = 0;			/* keep compiler quiet */
-
-	/* Everyone's had a chance to ask for space, so now create the DSM */
-	InitializeParallelDSM(pcxt);
-
-	/* If no DSM segment was available, back out (do serial build) */
-	if (pcxt->seg == NULL)
-	{
-		if (IsMVCCSnapshot(snapshot))
-			UnregisterSnapshot(snapshot);
-		DestroyParallelContext(pcxt);
-		ExitParallelMode();
-		return;
-	}
-
-	/* Store shared build state, for which we reserved space */
-	hnswshared = (HnswShared *) shm_toc_allocate(pcxt->toc, esthnswshared);
-	/* Initialize immutable state */
-	hnswshared->heaprelid = RelationGetRelid(buildstate->heap);
-	hnswshared->indexrelid = RelationGetRelid(buildstate->index);
-	hnswshared->isconcurrent = isconcurrent;
-	hnswshared->scantuplesortstates = scantuplesortstates;
-	ConditionVariableInit(&hnswshared->workersdonecv);
-	SpinLockInit(&hnswshared->mutex);
-	/* Initialize mutable state */
-	hnswshared->nparticipantsdone = 0;
-	hnswshared->reltuples = 0;
-	hnswshared->indtuples = 0;
-#if PG_VERSION_NUM >= 120000
-	table_parallelscan_initialize(buildstate->heap,
-								  ParallelTableScanFromHnswShared(hnswshared),
-								  snapshot);
-#else
-	heap_parallelscan_initialize(&hnswshared->heapdesc, buildstate->heap, snapshot);
-#endif
-
-	shm_toc_insert(pcxt->toc, PARALLEL_KEY_HNSW_SHARED, hnswshared);
-
-	/* Store query string for workers */
-	if (debug_query_string)
-	{
-		char	   *sharedquery;
-
-		sharedquery = (char *) shm_toc_allocate(pcxt->toc, querylen + 1);
-		memcpy(sharedquery, debug_query_string, querylen + 1);
-		shm_toc_insert(pcxt->toc, PARALLEL_KEY_QUERY_TEXT, sharedquery);
-	}
-
-	/* Launch workers, saving status for leader/caller */
-	LaunchParallelWorkers(pcxt);
-	hnswleader->pcxt = pcxt;
-	hnswleader->nparticipanttuplesorts = pcxt->nworkers_launched;
-	if (leaderparticipates)
-		hnswleader->nparticipanttuplesorts++;
-	hnswleader->hnswshared = hnswshared;
-	hnswleader->snapshot = snapshot;
-
-	/* If no workers were successfully launched, back out (do serial build) */
-	if (pcxt->nworkers_launched == 0)
-	{
-		HnswEndParallel(hnswleader);
-		return;
-	}
-
-	/* Log participants */
-	ereport(DEBUG1, (errmsg("using %d parallel workers", pcxt->nworkers_launched)));
-
-	/* Save leader state now that it's clear build will be parallel */
-	buildstate->hnswleader = hnswleader;
-
-	/* Join heap scan ourselves */
-	if (leaderparticipates)
-		HnswLeaderParticipateAsWorker(buildstate);
-
-	/* Wait for all launched workers */
-	WaitForParallelWorkersToAttach(pcxt);
-}
-
 /*
  * Build graph
  */
 static void
 BuildGraph(HnswBuildState * buildstate, ForkNumber forkNum)
 {
-	int			parallel_workers = 0;
-
 	UpdateProgress(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_HNSW_PHASE_LOAD);
 
-	/* Calculate parallel workers */
-	parallel_workers = plan_create_index_workers(RelationGetRelid(buildstate->heap), RelationGetRelid(buildstate->index));
-
-	/* Attempt to launch parallel worker scan when required */
-	if (parallel_workers > 0)
-	{
-		/* TODO Support in-memory builds */
-		FlushPages(buildstate);
-		HnswBeginParallel(buildstate, buildstate->indexInfo->ii_Concurrent, parallel_workers);
-	}
-
-	/* Add tuples to sort */
-	if (buildstate->hnswleader)
-		buildstate->reltuples = ParallelHeapScan(buildstate);
-	else
-	{
 #if PG_VERSION_NUM >= 120000
-		buildstate->reltuples = table_index_build_scan(buildstate->heap, buildstate->index, buildstate->indexInfo,
-													   true, true, BuildCallback, (void *) buildstate, NULL);
+	buildstate->reltuples = table_index_build_scan(buildstate->heap, buildstate->index, buildstate->indexInfo,
+												   true, true, BuildCallback, (void *) buildstate, NULL);
 #else
-		buildstate->reltuples = IndexBuildHeapScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
-												   true, BuildCallback, (void *) buildstate, NULL);
+	buildstate->reltuples = IndexBuildHeapScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
+											   true, BuildCallback, (void *) buildstate, NULL);
 #endif
-	}
-
-	/* End parallel build */
-	if (buildstate->hnswleader)
-		HnswEndParallel(buildstate->hnswleader);
 }
 
 /*

src/hnswinsert.c

@@ -329,7 +329,7 @@ HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswE
 
 			/* Get latest neighbors since they may have changed */
 			/* Do not lock yet since selecting neighbors can take time */
-			HnswLoadNeighbors(hc->element, index);
+			HnswLoadNeighbors(hc->element, index, m);
 
 			/*
 			 * Could improve performance for vacuuming by checking neighbors
@@ -492,9 +492,8 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 	FmgrInfo   *normprocinfo;
 	HnswElement entryPoint;
 	HnswElement element;
-	int			m = HnswGetM(index);
+	int			m;
 	int			efConstruction = HnswGetEfConstruction(index);
-	double		ml = HnswGetMl(m);
 	FmgrInfo   *procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
 	Oid			collation = index->rd_indcollation[0];
 	HnswElement dup;
@@ -511,10 +510,6 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 			return false;
 	}
 
-	/* Create an element */
-	element = HnswInitElement(heap_tid, m, ml, HnswGetMaxLevel(m));
-	element->vec = DatumGetVector(value);
-
 	/*
 	 * Get a shared lock. This allows vacuum to ensure no in-flight inserts
 	 * before repairing graph. Use a page lock so it does not interfere with
@@ -522,8 +517,12 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 	 */
 	LockPage(index, HNSW_UPDATE_LOCK, lockmode);
 
-	/* Get entry point */
-	entryPoint = HnswGetEntryPoint(index);
+	/* Get m and entry point */
+	HnswGetMetaPageInfo(index, &m, &entryPoint);
+
+	/* Create an element */
+	element = HnswInitElement(heap_tid, m, HnswGetMl(m), HnswGetMaxLevel(m));
+	element->vec = DatumGetVector(value);
 
 	/* Prevent concurrent inserts when likely updating entry point */
 	if (entryPoint == NULL || element->level > entryPoint->level)

src/hnswscan.c

@@ -19,7 +19,11 @@ GetScanItems(IndexScanDesc scan, Datum q)
 	Oid			collation = so->collation;
 	List	   *ep;
 	List	   *w;
-	HnswElement entryPoint = HnswGetEntryPoint(index);
+	int			m;
+	HnswElement entryPoint;
+
+	/* Get m and entry point */
+	HnswGetMetaPageInfo(index, &m, &entryPoint);
 
 	if (entryPoint == NULL)
 		return NIL;
@@ -28,11 +32,11 @@ GetScanItems(IndexScanDesc scan, Datum q)
 
 	for (int lc = entryPoint->level; lc >= 1; lc--)
 	{
-		w = HnswSearchLayer(q, ep, 1, lc, index, procinfo, collation, false, NULL);
+		w = HnswSearchLayer(q, ep, 1, lc, index, procinfo, collation, m, false, NULL);
 		ep = w;
 	}
 
-	return HnswSearchLayer(q, ep, hnsw_ef_search, 0, index, procinfo, collation, false, NULL);
+	return HnswSearchLayer(q, ep, hnsw_ef_search, 0, index, procinfo, collation, m, false, NULL);
 }
 
 /*
@@ -58,6 +62,33 @@ GetDimensions(Relation index)
 	return dimensions;
 }
 
+/*
+ * Get scan value
+ */
+static Datum
+GetScanValue(IndexScanDesc scan)
+{
+	HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
+	Datum		value;
+
+	if (scan->orderByData->sk_flags & SK_ISNULL)
+		value = PointerGetDatum(InitVector(GetDimensions(scan->indexRelation)));
+	else
+	{
+		value = scan->orderByData->sk_argument;
+
+		/* Value should not be compressed or toasted */
+		Assert(!VARATT_IS_COMPRESSED(DatumGetPointer(value)));
+		Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));
+
+		/* Fine if normalization fails */
+		if (so->normprocinfo != NULL)
+			HnswNormValue(so->normprocinfo, so->collation, &value, NULL);
+	}
+
+	return value;
+}
+
 /*
  * Prepare for an index scan
  */
@@ -70,7 +101,6 @@ hnswbeginscan(Relation index, int nkeys, int norderbys)
 	scan = RelationGetIndexScan(index, nkeys, norderbys);
 
 	so = (HnswScanOpaque) palloc(sizeof(HnswScanOpaqueData));
-	so->buf = InvalidBuffer;
 	so->first = true;
 	so->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
 									   "Hnsw scan temporary context",
@@ -83,6 +113,12 @@ hnswbeginscan(Relation index, int nkeys, int norderbys)
 
 	scan->opaque = so;
 
+	/*
+	 * Get a shared lock. This allows vacuum to ensure no in-flight scans
+	 * before marking tuples as deleted.
+	 */
+	LockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
+
 	return scan;
 }
 
@@ -130,72 +166,46 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
 		if (scan->orderByData == NULL)
 			elog(ERROR, "cannot scan hnsw index without order");
 
-		if (scan->orderByData->sk_flags & SK_ISNULL)
-			value = PointerGetDatum(InitVector(GetDimensions(scan->indexRelation)));
-		else
-		{
-			value = scan->orderByData->sk_argument;
-
-			/* Value should not be compressed or toasted */
-			Assert(!VARATT_IS_COMPRESSED(DatumGetPointer(value)));
-			Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));
-
-			/* Fine if normalization fails */
-			if (so->normprocinfo != NULL)
-				HnswNormValue(so->normprocinfo, so->collation, &value, NULL);
-		}
-
-		/*
-		 * Get a shared lock. This allows vacuum to ensure no in-flight scans
-		 * before marking tuples as deleted.
-		 */
-		LockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
+		/* Get scan value */
+		value = GetScanValue(scan);
 
 		so->w = GetScanItems(scan, value);
 
-		/* Release shared lock */
-		UnlockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
-
 		so->first = false;
 	}
 
 	while (list_length(so->w) > 0)
 	{
 		HnswCandidate *hc = llast(so->w);
-		ItemPointer tid;
-		BlockNumber indexblkno;
+		ItemPointer heaptid;
 
-		/* Move to next element if no valid heap tids */
+		/* Move to next element if no valid heap TIDs */
 		if (list_length(hc->element->heaptids) == 0)
 		{
 			so->w = list_delete_last(so->w);
 			continue;
 		}
 
-		tid = llast(hc->element->heaptids);
-		indexblkno = hc->element->blkno;
+		heaptid = llast(hc->element->heaptids);
 
 		hc->element->heaptids = list_delete_last(hc->element->heaptids);
 
 		MemoryContextSwitchTo(oldCtx);
 
 #if PG_VERSION_NUM >= 120000
-		scan->xs_heaptid = *tid;
+		scan->xs_heaptid = *heaptid;
 #else
-		scan->xs_ctup.t_self = *tid;
+		scan->xs_ctup.t_self = *heaptid;
 #endif
 
-		/* Unpin buffer */
-		if (BufferIsValid(so->buf))
-			ReleaseBuffer(so->buf);
-
 		/*
-		 * An index scan must maintain a pin on the index page holding the
-		 * item last returned by amgettuple
+		 * Typically, an index scan must maintain a pin on the index page
+		 * holding the item last returned by amgettuple. However, this is not
+		 * needed with the current vacuum strategy, which ensures scans do not
+		 * visit tuples in danger of being marked as deleted.
 		 *
 		 * https://www.postgresql.org/docs/current/index-locking.html
 		 */
-		so->buf = ReadBuffer(scan->indexRelation, indexblkno);
 
 		scan->xs_recheckorderby = false;
 		return true;
@@ -213,9 +223,8 @@ hnswendscan(IndexScanDesc scan)
 {
 	HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
 
-	/* Release pin */
-	if (BufferIsValid(so->buf))
-		ReleaseBuffer(so->buf);
+	/* Release shared lock */
+	UnlockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
 
 	MemoryContextDelete(so->tmpCtx);
 

src/hnswutils.c

@@ -210,25 +210,43 @@ HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno)
 }
 
 /*
- * Get the entry point
+ * Get the metapage info
  */
-HnswElement
-HnswGetEntryPoint(Relation index)
+void
+HnswGetMetaPageInfo(Relation index, int *m, HnswElement * entryPoint)
 {
 	Buffer		buf;
 	Page		page;
 	HnswMetaPage metap;
-	HnswElement entryPoint = NULL;
 
 	buf = ReadBuffer(index, HNSW_METAPAGE_BLKNO);
 	LockBuffer(buf, BUFFER_LOCK_SHARE);
 	page = BufferGetPage(buf);
 	metap = HnswPageGetMeta(page);
 
-	if (BlockNumberIsValid(metap->entryBlkno))
-		entryPoint = HnswInitElementFromBlock(metap->entryBlkno, metap->entryOffno);
+	if (m != NULL)
+		*m = metap->m;
+
+	if (entryPoint != NULL)
+	{
+		if (BlockNumberIsValid(metap->entryBlkno))
+			*entryPoint = HnswInitElementFromBlock(metap->entryBlkno, metap->entryOffno);
+		else
+			*entryPoint = NULL;
+	}
 
 	UnlockReleaseBuffer(buf);
+}
+
+/*
+ * Get the entry point
+ */
+HnswElement
+HnswGetEntryPoint(Relation index)
+{
+	HnswElement entryPoint;
+
+	HnswGetMetaPageInfo(index, NULL, &entryPoint);
 
 	return entryPoint;
 }
@@ -337,10 +355,9 @@ HnswSetNeighborTuple(HnswNeighborTuple ntup, HnswElement e, int m)
  * Load neighbors from page
  */
 static void
-LoadNeighborsFromPage(HnswElement element, Relation index, Page page)
+LoadNeighborsFromPage(HnswElement element, Relation index, Page page, int m)
 {
 	HnswNeighborTuple ntup = (HnswNeighborTuple) PageGetItem(page, PageGetItemId(page, element->neighborOffno));
-	int			m = HnswGetM(index);
 	int			neighborCount = (element->level + 2) * m;
 
 	Assert(HnswIsNeighborTuple(ntup));
@@ -381,7 +398,7 @@ LoadNeighborsFromPage(HnswElement element, Relation index, Page page)
  * Load neighbors
  */
 void
-HnswLoadNeighbors(HnswElement element, Relation index)
+HnswLoadNeighbors(HnswElement element, Relation index, int m)
 {
 	Buffer		buf;
 	Page		page;
@@ -390,7 +407,7 @@ HnswLoadNeighbors(HnswElement element, Relation index)
 	LockBuffer(buf, BUFFER_LOCK_SHARE);
 	page = BufferGetPage(buf);
 
-	LoadNeighborsFromPage(element, index, page);
+	LoadNeighborsFromPage(element, index, page, m);
 
 	UnlockReleaseBuffer(buf);
 }
@@ -543,7 +560,7 @@ AddToVisited(HTAB *v, HnswCandidate * hc, Relation index, bool *found)
  * Algorithm 2 from paper
  */
 List *
-HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, bool inserting, HnswElement skipElement)
+HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement)
 {
 	ListCell   *lc2;
 
@@ -598,7 +615,7 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
 			break;
 
 		if (c->element->neighbors == NULL)
-			HnswLoadNeighbors(c->element, index);
+			HnswLoadNeighbors(c->element, index, m);
 
 		/* Get the neighborhood at layer lc */
 		neighborhood = &c->element->neighbors[lc];
@@ -956,7 +973,7 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
 	/* 1st phase: greedy search to insert level */
 	for (int lc = entryLevel; lc >= level + 1; lc--)
 	{
-		w = HnswSearchLayer(q, ep, 1, lc, index, procinfo, collation, true, skipElement);
+		w = HnswSearchLayer(q, ep, 1, lc, index, procinfo, collation, m, true, skipElement);
 		ep = w;
 	}
 
@@ -974,7 +991,7 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
 		List	   *neighbors;
 		List	   *lw;
 
-		w = HnswSearchLayer(q, ep, efConstruction, lc, index, procinfo, collation, true, skipElement);
+		w = HnswSearchLayer(q, ep, efConstruction, lc, index, procinfo, collation, m, true, skipElement);
 
 		/* Elements being deleted or skipped can help with search */
 		/* but should be removed before selecting neighbors */

src/hnswvacuum.c

@@ -330,7 +330,10 @@ RepairGraph(HnswVacuumState * vacuumstate)
 	BufferAccessStrategy bas = vacuumstate->bas;
 	BlockNumber blkno = HNSW_HEAD_BLKNO;
 
-	/* Wait for inserts to complete */
+	/*
+	 * Wait for inserts to complete. Inserts before this point may have
+	 * neighbors about to be deleted. Inserts after this point will not.
+	 */
 	LockPage(index, HNSW_UPDATE_LOCK, ExclusiveLock);
 	UnlockPage(index, HNSW_UPDATE_LOCK, ExclusiveLock);
 
@@ -443,7 +446,11 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 	Relation	index = vacuumstate->index;
 	BufferAccessStrategy bas = vacuumstate->bas;
 
-	/* Wait for selects to complete */
+	/*
+	 * Wait for index scans to complete. Scans before this point may contain
+	 * tuples about to be deleted. Scans after this point will not, since the
+	 * graph has been repaired.
+	 */
 	LockPage(index, HNSW_SCAN_LOCK, ExclusiveLock);
 	UnlockPage(index, HNSW_SCAN_LOCK, ExclusiveLock);
 
@@ -582,7 +589,6 @@ InitVacuumState(HnswVacuumState * vacuumstate, IndexVacuumInfo *info, IndexBulkD
 	vacuumstate->stats = stats;
 	vacuumstate->callback = callback;
 	vacuumstate->callback_state = callback_state;
-	vacuumstate->m = HnswGetM(index);
 	vacuumstate->efConstruction = HnswGetEfConstruction(index);
 	vacuumstate->bas = GetAccessStrategy(BAS_BULKREAD);
 	vacuumstate->procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
@@ -592,6 +598,9 @@ InitVacuumState(HnswVacuumState * vacuumstate, IndexVacuumInfo *info, IndexBulkD
 												"Hnsw vacuum temporary context",
 												ALLOCSET_DEFAULT_SIZES);
 
+	/* Get m from metapage */
+	HnswGetMetaPageInfo(index, &vacuumstate->m, NULL);
+
 	/* Create hash table */
 	hash_ctl.keysize = sizeof(ItemPointerData);
 	hash_ctl.entrysize = sizeof(ItemPointerData);

src/ivfbuild.c

@@ -10,8 +10,8 @@
 #include "ivfflat.h"
 #include "miscadmin.h"
 #include "storage/bufmgr.h"
-#include "utils/memutils.h"
 #include "tcop/tcopprot.h"
+#include "utils/memutils.h"
 
 #if PG_VERSION_NUM >= 140000
 #include "utils/backend_progress.h"

src/ivfflat.c

@@ -71,7 +71,7 @@ ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 	int			lists;
 	double		ratio;
 	double		spc_seq_page_cost;
-	Relation	indexRel;
+	Relation	index;
 #if PG_VERSION_NUM < 120000
 	List	   *qinfos;
 #endif
@@ -89,9 +89,9 @@ ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 
 	MemSet(&costs, 0, sizeof(costs));
 
-	indexRel = index_open(path->indexinfo->indexoid, NoLock);
-	lists = IvfflatGetLists(indexRel);
-	index_close(indexRel, NoLock);
+	index = index_open(path->indexinfo->indexoid, NoLock);
+	IvfflatGetMetaPageInfo(index, &lists, NULL);
+	index_close(index, NoLock);
 
 	/* Get the ratio of lists that we need to visit */
 	ratio = ((double) ivfflat_probes) / lists;

src/ivfflat.h

@@ -40,6 +40,9 @@
 #define IVFFLAT_METAPAGE_BLKNO	0
 #define IVFFLAT_HEAD_BLKNO		1	/* first list page */
 
+/* Must correspond to page numbers since page lock is used */
+#define IVFFLAT_SCAN_LOCK		0
+
 /* IVFFlat parameters */
 #define IVFFLAT_DEFAULT_LISTS	100
 #define IVFFLAT_MIN_LISTS		1
@@ -244,8 +247,11 @@ typedef struct IvfflatScanList
 typedef struct IvfflatScanOpaqueData
 {
 	int			probes;
+	int			dimensions;
 	bool		first;
+	bool		hasLock;
 	Buffer		buf;
+	ItemPointerData heaptid;
 
 	/* Sorting */
 	Tuplesortstate *sortstate;
@@ -278,6 +284,7 @@ void		IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers);
 FmgrInfo   *IvfflatOptionalProcInfo(Relation rel, uint16 procnum);
 bool		IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
 int			IvfflatGetLists(Relation index);
+void		IvfflatGetMetaPageInfo(Relation index, int *lists, int *dimensions);
 void		IvfflatUpdateList(Relation index, ListInfo listInfo, BlockNumber insertPage, BlockNumber originalInsertPage, BlockNumber startPage, ForkNumber forkNum);
 void		IvfflatCommitBuffer(Buffer buf, GenericXLogState *state);
 void		IvfflatAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **state, ForkNumber forkNum);

src/ivfscan.c

@@ -9,6 +9,7 @@
 #include "miscadmin.h"
 #include "pgstat.h"
 #include "storage/bufmgr.h"
+#include "storage/lmgr.h"
 
 /*
  * Compare list distances
@@ -31,36 +32,36 @@ CompareLists(const pairingheap_node *a, const pairingheap_node *b, void *arg)
 static void
 GetScanLists(IndexScanDesc scan, Datum value)
 {
-	Buffer		cbuf;
-	Page		cpage;
-	IvfflatList list;
-	OffsetNumber offno;
-	OffsetNumber maxoffno;
+	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
 	BlockNumber nextblkno = IVFFLAT_HEAD_BLKNO;
 	int			listCount = 0;
-	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
-	double		distance;
-	IvfflatScanList *scanlist;
 	double		maxDistance = DBL_MAX;
 
 	/* Search all list pages */
 	while (BlockNumberIsValid(nextblkno))
 	{
+		Buffer		cbuf;
+		Page		cpage;
+		OffsetNumber maxoffno;
+
 		cbuf = ReadBuffer(scan->indexRelation, nextblkno);
 		LockBuffer(cbuf, BUFFER_LOCK_SHARE);
 		cpage = BufferGetPage(cbuf);
 
 		maxoffno = PageGetMaxOffsetNumber(cpage);
 
-		for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
+		for (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
 		{
-			list = (IvfflatList) PageGetItem(cpage, PageGetItemId(cpage, offno));
+			IvfflatList list = (IvfflatList) PageGetItem(cpage, PageGetItemId(cpage, offno));
+			double		distance;
 
 			/* Use procinfo from the index instead of scan key for performance */
 			distance = DatumGetFloat8(FunctionCall2Coll(so->procinfo, so->collation, PointerGetDatum(&list->center), value));
 
 			if (listCount < so->probes)
 			{
+				IvfflatScanList *scanlist;
+
 				scanlist = &so->lists[listCount];
 				scanlist->startPage = list->startPage;
 				scanlist->distance = distance;
@@ -75,6 +76,8 @@ GetScanLists(IndexScanDesc scan, Datum value)
 			}
 			else if (distance < maxDistance)
 			{
+				IvfflatScanList *scanlist;
+
 				/* Remove */
 				scanlist = (IvfflatScanList *) pairingheap_remove_first(so->listQueue);
 
@@ -101,14 +104,6 @@ static void
 GetScanItems(IndexScanDesc scan, Datum value)
 {
 	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
-	Buffer		buf;
-	Page		page;
-	IndexTuple	itup;
-	BlockNumber searchPage;
-	OffsetNumber offno;
-	OffsetNumber maxoffno;
-	Datum		datum;
-	bool		isnull;
 	TupleDesc	tupdesc = RelationGetDescr(scan->indexRelation);
 	double		tuples = 0;
 
@@ -128,19 +123,32 @@ GetScanItems(IndexScanDesc scan, Datum value)
 	/* Search closest probes lists */
 	while (!pairingheap_is_empty(so->listQueue))
 	{
-		searchPage = ((IvfflatScanList *) pairingheap_remove_first(so->listQueue))->startPage;
+		BlockNumber searchPage = ((IvfflatScanList *) pairingheap_remove_first(so->listQueue))->startPage;
 
 		/* Search all entry pages for list */
 		while (BlockNumberIsValid(searchPage))
 		{
+			Buffer		buf;
+			Page		page;
+			OffsetNumber maxoffno;
+
 			buf = ReadBufferExtended(scan->indexRelation, MAIN_FORKNUM, searchPage, RBM_NORMAL, bas);
 			LockBuffer(buf, BUFFER_LOCK_SHARE);
 			page = BufferGetPage(buf);
 			maxoffno = PageGetMaxOffsetNumber(page);
 
-			for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
+			for (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
 			{
-				itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, offno));
+				IndexTuple	itup;
+				Datum		datum;
+				bool		isnull;
+				ItemId		itemid = PageGetItemId(page, offno);
+
+				/* Skip dead tuples */
+				if (scan->ignore_killed_tuples && ItemIdIsDead(itemid))
+					continue;
+
+				itup = (IndexTuple) PageGetItem(page, itemid);
 				datum = index_getattr(itup, 1, tupdesc, &isnull);
 
 				/*
@@ -181,26 +189,52 @@ GetScanItems(IndexScanDesc scan, Datum value)
 }
 
 /*
- * Get dimensions from metapage
+ * Mark prior tuple as dead
  */
-static int
-GetDimensions(Relation index)
+static void
+MarkPriorTupleDead(IndexScanDesc scan)
 {
-	Buffer		buf;
+	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
+	Buffer		buf = so->buf;
 	Page		page;
-	IvfflatMetaPage metap;
-	int			dimensions;
+	OffsetNumber maxoffno;
 
-	buf = ReadBuffer(index, IVFFLAT_METAPAGE_BLKNO);
+	/* Safety check */
+	if (!BufferIsValid(so->buf) || !ItemPointerIsValid(&so->heaptid))
+		return;
+
+	/* Only a shared locked is needed for ItemIdMarkDead */
 	LockBuffer(buf, BUFFER_LOCK_SHARE);
 	page = BufferGetPage(buf);
-	metap = IvfflatPageGetMeta(page);
+	maxoffno = PageGetMaxOffsetNumber(page);
 
-	dimensions = metap->dimensions;
+	for (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
+	{
+		ItemId		itemid = PageGetItemId(page, offno);
+		IndexTuple	itup = (IndexTuple) PageGetItem(page, itemid);
 
-	UnlockReleaseBuffer(buf);
+		/*
+		 * Find tuple. Since buffer has been pinned, tuple cannot have been
+		 * vacuumed (and heap TID reused).
+		 */
+		if (ItemPointerEquals(&itup->t_tid, &so->heaptid))
+		{
+			/*
+			 * Make sure tuple has not already been marked dead to avoid extra
+			 * WAL if wal_log_hints or data checksums enabled
+			 */
+			if (!ItemIdIsDead(itemid))
+			{
+				ItemIdMarkDead(itemid);
+				MarkBufferDirtyHint(buf, true);
+			}
 
-	return dimensions;
+			break;
+		}
+	}
+
+	/* Unlock buffer */
+	LockBuffer(buf, BUFFER_LOCK_UNLOCK);
 }
 
 /*
@@ -212,6 +246,7 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 	IndexScanDesc scan;
 	IvfflatScanOpaque so;
 	int			lists;
+	int			dimensions;
 	AttrNumber	attNums[] = {1};
 	Oid			sortOperators[] = {Float8LessOperator};
 	Oid			sortCollations[] = {InvalidOid};
@@ -219,7 +254,9 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 	int			probes = ivfflat_probes;
 
 	scan = RelationGetIndexScan(index, nkeys, norderbys);
-	lists = IvfflatGetLists(scan->indexRelation);
+
+	/* Get lists and dimensions from metapage */
+	IvfflatGetMetaPageInfo(index, &lists, &dimensions);
 
 	if (probes > lists)
 		probes = lists;
@@ -227,7 +264,10 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 	so = (IvfflatScanOpaque) palloc(offsetof(IvfflatScanOpaqueData, lists) + probes * sizeof(IvfflatScanList));
 	so->buf = InvalidBuffer;
 	so->first = true;
+	so->hasLock = false;
+	ItemPointerSetInvalid(&so->heaptid);
 	so->probes = probes;
+	so->dimensions = dimensions;
 
 	/* Set support functions */
 	so->procinfo = index_getprocinfo(index, 1, IVFFLAT_DISTANCE_PROC);
@@ -241,7 +281,7 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 	so->tupdesc = CreateTemplateTupleDesc(3, false);
 #endif
 	TupleDescInitEntry(so->tupdesc, (AttrNumber) 1, "distance", FLOAT8OID, -1, 0);
-	TupleDescInitEntry(so->tupdesc, (AttrNumber) 2, "tid", TIDOID, -1, 0);
+	TupleDescInitEntry(so->tupdesc, (AttrNumber) 2, "heaptid", TIDOID, -1, 0);
 	TupleDescInitEntry(so->tupdesc, (AttrNumber) 3, "indexblkno", INT4OID, -1, 0);
 
 	/* Prep sort */
@@ -274,6 +314,7 @@ ivfflatrescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys, int
 #endif
 
 	so->first = true;
+	ItemPointerSetInvalid(&so->heaptid);
 	pairingheap_reset(so->listQueue);
 
 	if (keys && scan->numberOfKeys > 0)
@@ -308,8 +349,15 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 		if (scan->orderByData == NULL)
 			elog(ERROR, "cannot scan ivfflat index without order");
 
+		/* Get a shared lock for non-MVCC snapshots */
+		if (!so->hasLock && !IsMVCCSnapshot(scan->xs_snapshot))
+		{
+			so->hasLock = true;
+			LockPage(scan->indexRelation, IVFFLAT_SCAN_LOCK, ShareLock);
+		}
+
 		if (scan->orderByData->sk_flags & SK_ISNULL)
-			value = PointerGetDatum(InitVector(GetDimensions(scan->indexRelation)));
+			value = PointerGetDatum(InitVector(so->dimensions));
 		else
 		{
 			value = scan->orderByData->sk_argument;
@@ -331,18 +379,27 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 		if (value != scan->orderByData->sk_argument)
 			pfree(DatumGetPointer(value));
 	}
+	else
+	{
+		/* Mark prior tuple as dead */
+		if (scan->kill_prior_tuple)
+			MarkPriorTupleDead(scan);
+	}
 
 	if (tuplesort_gettupleslot(so->sortstate, true, false, so->slot, NULL))
 	{
-		ItemPointer tid = (ItemPointer) DatumGetPointer(slot_getattr(so->slot, 2, &so->isnull));
+		ItemPointer heaptid = (ItemPointer) DatumGetPointer(slot_getattr(so->slot, 2, &so->isnull));
 		BlockNumber indexblkno = DatumGetInt32(slot_getattr(so->slot, 3, &so->isnull));
 
 #if PG_VERSION_NUM >= 120000
-		scan->xs_heaptid = *tid;
+		scan->xs_heaptid = *heaptid;
 #else
-		scan->xs_ctup.t_self = *tid;
+		scan->xs_ctup.t_self = *heaptid;
 #endif
 
+		/* Keep track of info needed to mark tuple as dead */
+		so->heaptid = *heaptid;
+
 		/* Unpin buffer */
 		if (BufferIsValid(so->buf))
 			ReleaseBuffer(so->buf);
@@ -374,6 +431,10 @@ ivfflatendscan(IndexScanDesc scan)
 	if (BufferIsValid(so->buf))
 		ReleaseBuffer(so->buf);
 
+	/* Release lock */
+	if (so->hasLock)
+		UnlockPage(scan->indexRelation, IVFFLAT_SCAN_LOCK, ShareLock);
+
 	pairingheap_free(so->listQueue);
 	tuplesort_end(so->sortstate);
 

src/ivfutils.c

@@ -172,6 +172,29 @@ IvfflatAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **st
 	*buf = newbuf;
 }
 
+/*
+ * Get the metapage info
+ */
+void
+IvfflatGetMetaPageInfo(Relation index, int *lists, int *dimensions)
+{
+	Buffer		buf;
+	Page		page;
+	IvfflatMetaPage metap;
+
+	buf = ReadBuffer(index, IVFFLAT_METAPAGE_BLKNO);
+	LockBuffer(buf, BUFFER_LOCK_SHARE);
+	page = BufferGetPage(buf);
+	metap = IvfflatPageGetMeta(page);
+
+	*lists = metap->lists;
+
+	if (dimensions != NULL)
+		*dimensions = metap->dimensions;
+
+	UnlockReleaseBuffer(buf);
+}
+
 /*
  * Update the start or insert page of a list
  */

src/ivfvacuum.c

@@ -3,6 +3,7 @@
 #include "commands/vacuum.h"
 #include "ivfflat.h"
 #include "storage/bufmgr.h"
+#include "storage/lmgr.h"
 
 /*
  * Bulk delete tuples from the index
@@ -65,14 +66,10 @@ ivfflatbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 
 				vacuum_delay_point();
 
-				buf = ReadBufferExtended(index, MAIN_FORKNUM, searchPage, RBM_NORMAL, bas);
+				/* Ensure no in-flight index scans for non-MVCC snapshots */
+				LockPage(index, IVFFLAT_SCAN_LOCK, ExclusiveLock);
 
-				/*
-				 * ambulkdelete cannot delete entries from pages that are
-				 * pinned by other backends
-				 *
-				 * https://www.postgresql.org/docs/current/index-locking.html
-				 */
+				buf = ReadBufferExtended(index, MAIN_FORKNUM, searchPage, RBM_NORMAL, bas);
 				LockBufferForCleanup(buf);
 
 				state = GenericXLogStart(index);
@@ -114,6 +111,8 @@ ivfflatbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 					GenericXLogAbort(state);
 
 				UnlockReleaseBuffer(buf);
+
+				UnlockPage(index, IVFFLAT_SCAN_LOCK, ExclusiveLock);
 			}
 
 			/*

src/vector.c

@@ -695,6 +695,8 @@ vector_spherical_distance(PG_FUNCTION_ARGS)
 {
 	Vector	   *a = PG_GETARG_VECTOR_P(0);
 	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	float	   *ax = a->x;
+	float	   *bx = b->x;
 	float		dp = 0.0;
 	double		distance;
 
@@ -702,7 +704,7 @@ vector_spherical_distance(PG_FUNCTION_ARGS)
 
 	/* Auto-vectorized */
 	for (int i = 0; i < a->dim; i++)
-		dp += a->x[i] * b->x[i];
+		dp += ax[i] * bx[i];
 
 	distance = (double) dp;
 

test/expected/functions.out

@@ -106,6 +106,12 @@ SELECT cosine_distance('[1,1]', '[1,1]');
                0
 (1 row)
 
+SELECT cosine_distance('[1,0]', '[0,2]');
+ cosine_distance 
+-----------------
+               1
+(1 row)
+
 SELECT cosine_distance('[1,1]', '[-1,-1]');
  cosine_distance 
 -----------------

test/sql/functions.sql

@@ -25,6 +25,7 @@ SELECT inner_product('[3e38]', '[3e38]');
 SELECT cosine_distance('[1,2]', '[2,4]');
 SELECT cosine_distance('[1,2]', '[0,0]');
 SELECT cosine_distance('[1,1]', '[1,1]');
+SELECT cosine_distance('[1,0]', '[0,2]');
 SELECT cosine_distance('[1,1]', '[-1,-1]');
 SELECT cosine_distance('[1,2]', '[3]');
 SELECT cosine_distance('[1,1]', '[1.1,1.1]');

test/t/001_ivfflat_wal.pl

@@ -19,8 +19,6 @@ sub test_index_replay
 
 	# Wait for replica to catch up
 	my $applname = $node_replica->name;
-
-	my $server_version_num = $node_primary->safe_psql("postgres", "SHOW server_version_num");
 	my $caughtup_query = "SELECT pg_current_wal_lsn() <= replay_lsn FROM pg_stat_replication WHERE application_name = '$applname';";
 	$node_primary->poll_query_until('postgres', $caughtup_query)
 	  or die "Timed out while waiting for replica 1 to catch up";

test/t/003_ivfflat_build_recall.pl

@@ -94,7 +94,7 @@ for my $i (0 .. $#operators)
 	# Test approximate results
 	if ($operator ne "<#>")
 	{
-		# TODO fix test
+		# TODO Fix test (uniform random vectors all have similar inner product)
 		test_recall(1, 0.71, $operator);
 		test_recall(10, 0.95, $operator);
 	}
@@ -115,7 +115,7 @@ for my $i (0 .. $#operators)
 	# Test approximate results
 	if ($operator ne "<#>")
 	{
-		# TODO fix test
+		# TODO Fix test (uniform random vectors all have similar inner product)
 		test_recall(1, 0.71, $operator);
 		test_recall(10, 0.95, $operator);
 	}

test/t/010_hnsw_wal.pl

@@ -19,8 +19,6 @@ sub test_index_replay
 
 	# Wait for replica to catch up
 	my $applname = $node_replica->name;
-
-	my $server_version_num = $node_primary->safe_psql("postgres", "SHOW server_version_num");
 	my $caughtup_query = "SELECT pg_current_wal_lsn() <= replay_lsn FROM pg_stat_replication WHERE application_name = '$applname';";
 	$node_primary->poll_query_until('postgres', $caughtup_query)
 	  or die "Timed out while waiting for replica 1 to catch up";

test/t/012_hnsw_build_recall.pl

@@ -83,31 +83,11 @@ for my $i (0 .. $#operators)
 		push(@expected, $res);
 	}
 
-	# Build index serially
-	$node->safe_psql("postgres", qq(
-		SET max_parallel_maintenance_workers = 0;
-		CREATE INDEX idx ON tst USING hnsw (v $opclass);
-	));
+	# Add index
+	$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v $opclass);");
 
-	# Test approximate results
 	my $min = $operator eq "<#>" ? 0.80 : 0.99;
 	test_recall($min, $operator);
-
-	$node->safe_psql("postgres", "DROP INDEX idx;");
-
-	# Build index in parallel
-	my ($ret, $stdout, $stderr) = $node->psql("postgres", qq(
-		SET client_min_messages = DEBUG;
-		SET min_parallel_table_scan_size = 1;
-		CREATE INDEX idx ON tst USING hnsw (v $opclass);
-	));
-	is($ret, 0, $stderr);
-	like($stderr, qr/using \d+ parallel workers/);
-
-	# Test approximate results
-	test_recall($min, $operator);
-
-	$node->safe_psql("postgres", "DROP INDEX idx;");
 }
 
 done_testing();

test/t/016_hnsw_vacuum_recall.pl

@@ -89,7 +89,7 @@ foreach (@queries)
 test_recall(0.20, $limit, "before vacuum");
 test_recall(0.95, 100, "before vacuum");
 
-# TODO test concurrent inserts with vacuum
+# TODO Test concurrent inserts with vacuum
 $node->safe_psql("postgres", "VACUUM tst;");
 
 test_recall(0.95, $limit, "after vacuum");

test/t/017_ivfflat_insert_recall.pl

@@ -0,0 +1,117 @@
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More;
+
+my $node;
+my @queries = ();
+my @expected;
+my $limit = 20;
+
+sub test_recall
+{
+	my ($probes, $min, $operator) = @_;
+	my $correct = 0;
+	my $total = 0;
+
+	my $explain = $node->safe_psql("postgres", qq(
+		SET enable_seqscan = off;
+		SET ivfflat.probes = $probes;
+		EXPLAIN ANALYZE SELECT i FROM tst ORDER BY v $operator '$queries[0]' LIMIT $limit;
+	));
+	like($explain, qr/Index Scan using idx on tst/);
+
+	for my $i (0 .. $#queries)
+	{
+		my $actual = $node->safe_psql("postgres", qq(
+			SET enable_seqscan = off;
+			SET ivfflat.probes = $probes;
+			SELECT i FROM tst ORDER BY v $operator '$queries[$i]' LIMIT $limit;
+		));
+		my @actual_ids = split("\n", $actual);
+		my %actual_set = map { $_ => 1 } @actual_ids;
+
+		my @expected_ids = split("\n", $expected[$i]);
+
+		foreach (@expected_ids)
+		{
+			if (exists($actual_set{$_}))
+			{
+				$correct++;
+			}
+			$total++;
+		}
+	}
+
+	cmp_ok($correct / $total, ">=", $min, $operator);
+}
+
+# Initialize node
+$node = get_new_node('node');
+$node->init;
+$node->start;
+
+# Create table
+$node->safe_psql("postgres", "CREATE EXTENSION vector;");
+$node->safe_psql("postgres", "CREATE TABLE tst (i serial, v vector(3));");
+
+# Generate queries
+for (1 .. 20)
+{
+	my $r1 = rand();
+	my $r2 = rand();
+	my $r3 = rand();
+	push(@queries, "[$r1,$r2,$r3]");
+}
+
+# Check each index type
+my @operators = ("<->", "<#>", "<=>");
+my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_ops");
+
+for my $i (0 .. $#operators)
+{
+	my $operator = $operators[$i];
+	my $opclass = $opclasses[$i];
+
+	# Add index
+	$node->safe_psql("postgres", "CREATE INDEX idx ON tst USING ivfflat (v $opclass);");
+
+	# Use concurrent inserts
+	$node->pgbench(
+		"--no-vacuum --client=10 --transactions=1000",
+		0,
+		[qr{actually processed}],
+		[qr{^$}],
+		"concurrent INSERTs",
+		{
+			"017_ivfflat_insert_recall_$opclass" => "INSERT INTO tst (v) SELECT ARRAY[random(), random(), random()] FROM generate_series(1, 10) i;"
+		}
+	);
+
+	# Get exact results
+	@expected = ();
+	foreach (@queries)
+	{
+		my $res = $node->safe_psql("postgres", qq(
+			SET enable_indexscan = off;
+			SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;
+		));
+		push(@expected, $res);
+	}
+
+	# Test approximate results
+	if ($operator ne "<#>")
+	{
+		# TODO Fix test (uniform random vectors all have similar inner product)
+		test_recall(1, 0.71, $operator);
+		test_recall(10, 0.95, $operator);
+	}
+	# Account for equal distances
+	test_recall(100, 0.9925, $operator);
+
+	$node->safe_psql("postgres", "DROP INDEX idx;");
+	$node->safe_psql("postgres", "TRUNCATE tst;");
+}
+
+done_testing();

test/t/018_ivfflat_deletes.pl

@@ -0,0 +1,43 @@
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More;
+
+my $dim = 3;
+
+my $array_sql = join(",", ('random()') x $dim);
+
+# Initialize node
+my $node = get_new_node('node');
+$node->init;
+$node->start;
+
+# Create table and index
+$node->safe_psql("postgres", "CREATE EXTENSION vector;");
+$node->safe_psql("postgres", "CREATE TABLE tst (i serial, v vector($dim));");
+$node->safe_psql("postgres",
+	"INSERT INTO tst (v) SELECT ARRAY[$array_sql] FROM generate_series(1, 10000) i;"
+);
+$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops);");
+
+# Delete data
+$node->safe_psql("postgres", "DELETE FROM tst WHERE i % 100 != 0;");
+
+my $exp = $node->safe_psql("postgres", qq(
+	SET enable_indexscan = off;
+	SELECT i FROM tst ORDER BY v <-> '[0,0,0]';
+));
+
+# Run twice to make sure correct tuples marked as dead
+for (1 .. 2)
+{
+	my $res = $node->safe_psql("postgres", qq(
+		SET enable_seqscan = off;
+		SET ivfflat.probes = 100;
+		SELECT i FROM tst ORDER BY v <-> '[0,0,0]';
+	));
+	is($res, $exp);
+}
+
+done_testing();

vector.control

@@ -1,4 +1,4 @@
-comment = 'vector data type and ivfflat access method'
+comment = 'vector data type and ivfflat and hnsw access methods'
 default_version = '0.5.0'
 module_pathname = '$libdir/vector'
 relocatable = true