Added comment [skip ci]

This reverts commit dbef8d1ad1.

CHANGELOG.md

@@ -1,6 +1,6 @@
 ## 0.5.1 (unreleased)
 
-- Added `angular_distance` function
+- Improved performance of index scans for IVFFlat after updates and deletes
 
 ## 0.5.0 (2023-08-28)
 

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

sql/vector--0.5.0--0.5.1.sql

@@ -1,5 +0,0 @@
--- complain if script is sourced in psql, rather than via CREATE EXTENSION
-\echo Use "ALTER EXTENSION vector UPDATE TO '0.5.1'" to load this file. \quit
-
-CREATE FUNCTION angular_distance(vector, vector) RETURNS float8
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;

sql/vector.sql

@@ -87,9 +87,6 @@ CREATE FUNCTION vector_l2_squared_distance(vector, vector) RETURNS float8
 CREATE FUNCTION vector_negative_inner_product(vector, vector) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION angular_distance(vector, vector) RETURNS float8
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
 CREATE FUNCTION vector_spherical_distance(vector, vector) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 

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

@@ -266,8 +266,9 @@ 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);
@@ -284,7 +285,7 @@ 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);
+void		HnswLoadNeighbors(HnswElement element, Relation index, int m);
 
 /* Index access methods */
 IndexBuildResult *hnswbuild(Relation heap, Relation index, IndexInfo *indexInfo);

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
  */
@@ -130,20 +161,8 @@ 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 scan value */
+		value = GetScanValue(scan);
 
 		/*
 		 * Get a shared lock. This allows vacuum to ensure no in-flight scans
@@ -162,17 +181,17 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
 	while (list_length(so->w) > 0)
 	{
 		HnswCandidate *hc = llast(so->w);
-		ItemPointer tid;
+		ItemPointer heaptid;
 		BlockNumber indexblkno;
 
-		/* 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);
+		heaptid = llast(hc->element->heaptids);
 		indexblkno = hc->element->blkno;
 
 		hc->element->heaptids = list_delete_last(hc->element->heaptids);
@@ -180,9 +199,9 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
 		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 */

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

@@ -582,7 +582,6 @@ 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 +591,9 @@ 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/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;
@@ -182,6 +182,15 @@ ivfflatvalidate(Oid opclassoid)
 	return true;
 }
 
+/*
+ * Checks if index-only scan is supported
+ */
+static bool
+ivfflatcanreturn(Relation index, int attno)
+{
+	return attno == 1 && IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC) == NULL;
+}
+
 /*
  * Define index handler
  *
@@ -223,7 +232,7 @@ ivfflathandler(PG_FUNCTION_ARGS)
 	amroutine->aminsert = ivfflatinsert;
 	amroutine->ambulkdelete = ivfflatbulkdelete;
 	amroutine->amvacuumcleanup = ivfflatvacuumcleanup;
-	amroutine->amcanreturn = NULL;	/* tuple not included in heapsort */
+	amroutine->amcanreturn = ivfflatcanreturn;
 	amroutine->amcostestimate = ivfflatcostestimate;
 	amroutine->amoptions = ivfflatoptions;
 	amroutine->amproperty = NULL;	/* TODO AMPROP_DISTANCE_ORDERABLE */

src/ivfflat.h

@@ -244,8 +244,11 @@ typedef struct IvfflatScanList
 typedef struct IvfflatScanOpaqueData
 {
 	int			probes;
+	int			dimensions;
 	bool		first;
 	Buffer		buf;
+	ItemPointerData heaptid;
+	IndexTuple	itup;
 
 	/* Sorting */
 	Tuplesortstate *sortstate;
@@ -278,6 +281,7 @@ 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

@@ -31,36 +31,36 @@ 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 +75,8 @@ GetScanLists(IndexScanDesc scan, Datum value)
 			}
 			else if (distance < maxDistance)
 			{
+				IvfflatScanList *scanlist;
+
 				/* Remove */
 				scanlist = (IvfflatScanList *) pairingheap_remove_first(so->listQueue);
 
@@ -101,14 +103,6 @@ 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;
 
@@ -128,20 +122,35 @@ 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;
+				ItemPointerData indextid;
+				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);
+				ItemPointerSet(&indextid, searchPage, offno);
 
 				/*
 				 * Add virtual tuple
@@ -154,7 +163,7 @@ GetScanItems(IndexScanDesc scan, Datum value)
 				slot->tts_isnull[0] = false;
 				slot->tts_values[1] = PointerGetDatum(&itup->t_tid);
 				slot->tts_isnull[1] = false;
-				slot->tts_values[2] = Int32GetDatum((int) searchPage);
+				slot->tts_values[2] = PointerGetDatum(&indextid);
 				slot->tts_isnull[2] = false;
 				ExecStoreVirtualTuple(slot);
 
@@ -181,26 +190,80 @@ 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);
+}
+
+/*
+ * Set tuple for index-only scan
+ */
+static void
+SetIndexTuple(IndexScanDesc scan, ItemPointer indextid)
+{
+	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
+	Buffer		buf = so->buf;
+	Page		page;
+	OffsetNumber offno = ItemPointerGetOffsetNumber(indextid);
+	IndexTuple	itup;
+	Size		itupSize;
+
+	LockBuffer(buf, BUFFER_LOCK_SHARE);
+	page = BufferGetPage(buf);
+	itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, offno));
+	itupSize = IndexTupleSize(itup);
+
+	if (so->itup == NULL)
+		so->itup = palloc(BLCKSZ);
+
+	memcpy(so->itup, itup, itupSize);
+
+	scan->xs_itup = so->itup;
+
+	LockBuffer(buf, BUFFER_LOCK_UNLOCK);
 }
 
 /*
@@ -212,6 +275,7 @@ 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 +283,9 @@ 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 +293,10 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 	so = (IvfflatScanOpaque) palloc(offsetof(IvfflatScanOpaqueData, lists) + probes * sizeof(IvfflatScanList));
 	so->buf = InvalidBuffer;
 	so->first = true;
+	ItemPointerSetInvalid(&so->heaptid);
+	so->itup = NULL;
 	so->probes = probes;
+	so->dimensions = dimensions;
 
 	/* Set support functions */
 	so->procinfo = index_getprocinfo(index, 1, IVFFLAT_DISTANCE_PROC);
@@ -241,8 +310,8 @@ 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) 3, "indexblkno", INT4OID, -1, 0);
+	TupleDescInitEntry(so->tupdesc, (AttrNumber) 2, "heaptid", TIDOID, -1, 0);
+	TupleDescInitEntry(so->tupdesc, (AttrNumber) 3, "indextid", TIDOID, -1, 0);
 
 	/* Prep sort */
 	so->sortstate = tuplesort_begin_heap(so->tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, work_mem, NULL, false);
@@ -257,6 +326,8 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 
 	scan->opaque = so;
 
+	scan->xs_itupdesc = RelationGetDescr(index);
+
 	return scan;
 }
 
@@ -274,6 +345,7 @@ 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)
@@ -309,7 +381,7 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 			elog(ERROR, "cannot scan ivfflat index without order");
 
 		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 +403,27 @@ 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));
-		BlockNumber indexblkno = DatumGetInt32(slot_getattr(so->slot, 3, &so->isnull));
+		ItemPointer heaptid = (ItemPointer) DatumGetPointer(slot_getattr(so->slot, 2, &so->isnull));
+		ItemPointer indextid = (ItemPointer) DatumGetPointer(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);
@@ -353,7 +434,11 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 		 *
 		 * https://www.postgresql.org/docs/current/index-locking.html
 		 */
-		so->buf = ReadBuffer(scan->indexRelation, indexblkno);
+		so->buf = ReadBuffer(scan->indexRelation, ItemPointerGetBlockNumber(indextid));
+
+		/* Set tuple for index-only scan */
+		if (scan->xs_want_itup)
+			SetIndexTuple(scan, indextid);
 
 		scan->xs_recheckorderby = false;
 		return true;
@@ -374,6 +459,9 @@ ivfflatendscan(IndexScanDesc scan)
 	if (BufferIsValid(so->buf))
 		ReleaseBuffer(so->buf);
 
+	if (so->itup != NULL)
+		pfree(so->itup);
+
 	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/vector.c

@@ -684,49 +684,6 @@ cosine_distance(PG_FUNCTION_ARGS)
 	PG_RETURN_FLOAT8(1.0 - similarity);
 }
 
-/*
- * Get the angular distance between two vectors
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(angular_distance);
-Datum
-angular_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		distance = 0.0;
-	float		norma = 0.0;
-	float		normb = 0.0;
-	double		similarity;
-
-	CheckDims(a, b);
-
-	/* Auto-vectorized */
-	for (int i = 0; i < a->dim; i++)
-	{
-		distance += ax[i] * bx[i];
-		norma += ax[i] * ax[i];
-		normb += bx[i] * bx[i];
-	}
-
-	similarity = (double) distance / sqrt((double) norma * (double) normb);
-
-#ifdef _MSC_VER
-	/* /fp:fast may not propagate NaN */
-	if (isnan(similarity))
-		PG_RETURN_FLOAT8(NAN);
-#endif
-
-	/* Prevent NaN with acos with loss of precision */
-	if (similarity > 1)
-		similarity = 1;
-	else if (similarity < -1)
-		similarity = -1;
-
-	PG_RETURN_FLOAT8(acos(similarity) / M_PI);
-}
-
 /*
  * Get the distance for spherical k-means
  * Currently uses angular distance since needs to satisfy triangle inequality

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 
 -----------------
@@ -152,56 +158,6 @@ SELECT l1_distance('[3e38]', '[-3e38]');
     Infinity
 (1 row)
 
-SELECT angular_distance('[1,2]', '[2,4]');
- angular_distance 
-------------------
-                0
-(1 row)
-
-SELECT angular_distance('[1,2]', '[0,0]');
- angular_distance 
-------------------
-              NaN
-(1 row)
-
-SELECT angular_distance('[1,1]', '[1,1]');
- angular_distance 
-------------------
-                0
-(1 row)
-
-SELECT angular_distance('[1,0]', '[0,2]');
- angular_distance 
-------------------
-              0.5
-(1 row)
-
-SELECT angular_distance('[1,1]', '[-1,-1]');
- angular_distance 
-------------------
-                1
-(1 row)
-
-SELECT angular_distance('[1,2]', '[3]');
-ERROR:  different vector dimensions 2 and 1
-SELECT angular_distance('[1,1]', '[1.1,1.1]');
- angular_distance 
-------------------
-                0
-(1 row)
-
-SELECT angular_distance('[1,1]', '[-1.1,-1.1]');
- angular_distance 
-------------------
-                1
-(1 row)
-
-SELECT angular_distance('[3e38]', '[3e38]');
- angular_distance 
-------------------
-              NaN
-(1 row)
-
 SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]']) v;
     avg    
 -----------

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]');
@@ -36,16 +37,6 @@ SELECT l1_distance('[0,0]', '[0,1]');
 SELECT l1_distance('[1,2]', '[3]');
 SELECT l1_distance('[3e38]', '[-3e38]');
 
-SELECT angular_distance('[1,2]', '[2,4]');
-SELECT angular_distance('[1,2]', '[0,0]');
-SELECT angular_distance('[1,1]', '[1,1]');
-SELECT angular_distance('[1,0]', '[0,2]');
-SELECT angular_distance('[1,1]', '[-1,-1]');
-SELECT angular_distance('[1,2]', '[3]');
-SELECT angular_distance('[1,1]', '[1.1,1.1]');
-SELECT angular_distance('[1,1]', '[-1.1,-1.1]');
-SELECT angular_distance('[3e38]', '[3e38]');
-
 SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]']) v;
 SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]', NULL]) v;
 SELECT avg(v) FROM unnest(ARRAY[]::vector[]) v;

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