Updated function name

CHANGELOG.md

@@ -1,7 +1,7 @@
-## 0.5.1 (2023-10-10)
+## 0.5.1 (unreleased)
 
-- Improved performance of HNSW index builds
-- Added check for MVCC-compliant snapshot for index scans
+- Improved performance of index scans for IVFFlat after updates and deletes
+- Fixed locking for index scans for HNSW
 
 ## 0.5.0 (2023-08-28)
 

META.json

@@ -2,7 +2,7 @@
 	"name": "vector",
 	"abstract": "Open-source vector similarity search for Postgres",
 	"description": "Supports L2 distance, inner product, and cosine distance",
-	"version": "0.5.1",
+	"version": "0.5.0",
 	"maintainer": [
 		"Andrew Kane <andrew@ankane.org>"
 	],
@@ -20,7 +20,7 @@
 		"vector": {
 			"file": "sql/vector.sql",
 			"docfile": "README.md",
-			"version": "0.5.1",
+			"version": "0.5.0",
 			"abstract": "Open-source vector similarity search for Postgres"
 		}
 	},

Makefile

@@ -1,9 +1,9 @@
 EXTENSION = vector
-EXTVERSION = 0.5.1
+EXTVERSION = 0.5.0
 
 MODULE_big = vector
 DATA = $(wildcard sql/*--*.sql)
-OBJS = src/hnsw.o src/hnswbuild.o src/hnswinsert.o src/hnswscan.o src/hnswutils.o src/hnswvacuum.o src/ivfbuild.o src/ivfflat.o src/ivfinsert.o src/ivfkmeans.o src/ivfscan.o src/ivfutils.o src/ivfvacuum.o src/vector.o
+OBJS = src/float4.o src/hnsw.o src/hnswbuild.o src/hnswinsert.o src/hnswscan.o src/hnswutils.o src/hnswvacuum.o src/ivfbuild.o src/ivfflat.o src/ivfinsert.o src/ivfkmeans.o src/ivfscan.o src/ivfutils.o src/ivfvacuum.o src/vector.o
 HEADERS = src/vector.h
 
 TESTS = $(wildcard test/sql/*.sql)

Makefile.win

@@ -1,7 +1,7 @@
 EXTENSION = vector
-EXTVERSION = 0.5.1
+EXTVERSION = 0.5.0
 
-OBJS = src\hnsw.obj src\hnswbuild.obj src\hnswinsert.obj src\hnswscan.obj src\hnswutils.obj src\hnswvacuum.obj src\ivfbuild.obj src\ivfflat.obj src\ivfinsert.obj src\ivfkmeans.obj src\ivfscan.obj src\ivfutils.obj src\ivfvacuum.obj src\vector.obj
+OBJS = src\float4.obj src\hnsw.obj src\hnswbuild.obj src\hnswinsert.obj src\hnswscan.obj src\hnswutils.obj src\hnswvacuum.obj src\ivfbuild.obj src\ivfflat.obj src\ivfinsert.obj src\ivfkmeans.obj src\ivfscan.obj src\ivfutils.obj src\ivfvacuum.obj src\vector.obj
 HEADERS = src\vector.h
 
 REGRESS = btree cast copy functions input ivfflat_cosine ivfflat_ip ivfflat_l2 ivfflat_options ivfflat_unlogged

README.md

@@ -18,7 +18,7 @@ Compile and install the extension (supports Postgres 11+)
 
 ```sh
 cd /tmp
-git clone --branch v0.5.1 https://github.com/pgvector/pgvector.git
+git clone --branch v0.5.0 https://github.com/pgvector/pgvector.git
 cd pgvector
 make
 make install # may need sudo
@@ -509,7 +509,7 @@ Then use `nmake` to build:
 
 ```cmd
 set "PGROOT=C:\Program Files\PostgreSQL\15"
-git clone --branch v0.5.1 https://github.com/pgvector/pgvector.git
+git clone --branch v0.5.0 https://github.com/pgvector/pgvector.git
 cd pgvector
 nmake /F Makefile.win
 nmake /F Makefile.win install
@@ -530,7 +530,7 @@ This adds pgvector to the [Postgres image](https://hub.docker.com/_/postgres) (r
 You can also build the image manually:
 
 ```sh
-git clone --branch v0.5.1 https://github.com/pgvector/pgvector.git
+git clone --branch v0.5.0 https://github.com/pgvector/pgvector.git
 cd pgvector
 docker build --build-arg PG_MAJOR=15 -t myuser/pgvector .
 ```

sql/vector--0.5.0--0.5.1.sql

@@ -1,2 +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

sql/vector.sql

@@ -34,6 +34,9 @@ CREATE TYPE vector (
 CREATE FUNCTION l2_distance(vector, vector) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
+CREATE FUNCTION l2_distance(float4[], float4[]) RETURNS float8
+	AS 'MODULE_PATHNAME', 'float4_l2_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
 CREATE FUNCTION inner_product(vector, vector) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
@@ -84,6 +87,9 @@ CREATE FUNCTION vector_cmp(vector, vector) RETURNS int4
 CREATE FUNCTION vector_l2_squared_distance(vector, vector) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
+CREATE FUNCTION float4_l2_squared_distance(float4[], float4[]) RETURNS float8
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
 CREATE FUNCTION vector_negative_inner_product(vector, vector) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
@@ -164,6 +170,11 @@ CREATE OPERATOR <-> (
 	COMMUTATOR = '<->'
 );
 
+CREATE OPERATOR <-> (
+	LEFTARG = float4[], RIGHTARG = float4[], PROCEDURE = l2_distance,
+	COMMUTATOR = '<->'
+);
+
 CREATE OPERATOR <#> (
 	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_negative_inner_product,
 	COMMUTATOR = '<#>'
@@ -280,6 +291,11 @@ CREATE OPERATOR CLASS vector_l2_ops
 	OPERATOR 1 <-> (vector, vector) FOR ORDER BY float_ops,
 	FUNCTION 1 vector_l2_squared_distance(vector, vector);
 
+CREATE OPERATOR CLASS float4_l2_ops
+	FOR TYPE float4[] USING hnsw AS
+	OPERATOR 1 <-> (float4[], float4[]) FOR ORDER BY float_ops,
+	FUNCTION 1 float4_l2_squared_distance(float4[], float4[]);
+
 CREATE OPERATOR CLASS vector_ip_ops
 	FOR TYPE vector USING hnsw AS
 	OPERATOR 1 <#> (vector, vector) FOR ORDER BY float_ops,

src/float4.c

@@ -0,0 +1,60 @@
+#include "postgres.h"
+
+#include <math.h>
+
+#include "utils/array.h"
+
+/*
+ * Get the L2 distance between vectors
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(float4_l2_distance);
+Datum
+float4_l2_distance(PG_FUNCTION_ARGS)
+{
+	ArrayType  *a = PG_GETARG_ARRAYTYPE_P(0);
+	ArrayType  *b = PG_GETARG_ARRAYTYPE_P(1);
+	float	   *ax = (float *) ARR_DATA_PTR(a);
+	float	   *bx = (float *) ARR_DATA_PTR(b);
+	float		distance = 0.0;
+	float		diff;
+
+	/* TODO Check rank, dimensions, and nulls */
+	int			dim = ARR_DIMS(a)[0];
+
+	/* Auto-vectorized */
+	for (int i = 0; i < dim; i++)
+	{
+		diff = ax[i] - bx[i];
+		distance += diff * diff;
+	}
+
+	PG_RETURN_FLOAT8(sqrt((double) distance));
+}
+
+/*
+ * Get the L2 squared distance between vectors
+ * This saves a sqrt calculation
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(float4_l2_squared_distance);
+Datum
+float4_l2_squared_distance(PG_FUNCTION_ARGS)
+{
+	ArrayType  *a = PG_GETARG_ARRAYTYPE_P(0);
+	ArrayType  *b = PG_GETARG_ARRAYTYPE_P(1);
+	float	   *ax = (float *) ARR_DATA_PTR(a);
+	float	   *bx = (float *) ARR_DATA_PTR(b);
+	float		distance = 0.0;
+	float		diff;
+
+	/* TODO Check rank, dimensions, and nulls */
+	int			dim = ARR_DIMS(a)[0];
+
+	/* Auto-vectorized */
+	for (int i = 0; i < dim; i++)
+	{
+		diff = ax[i] - bx[i];
+		distance += diff * diff;
+	}
+
+	PG_RETURN_FLOAT8((double) distance);
+}

src/hnsw.h

@@ -60,8 +60,6 @@
 /* PROGRESS_CREATEIDX_SUBPHASE_INITIALIZE is 1 */
 #define PROGRESS_HNSW_PHASE_LOAD		2
 
-#define HNSW_MAX_SIZE (BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - sizeof(ItemIdData))
-
 #define HNSW_ELEMENT_TUPLE_SIZE(_datum)	MAXALIGN(offsetof(HnswElementTupleData, value) + VARSIZE_ANY(_datum))
 #define HNSW_NEIGHBOR_TUPLE_SIZE(level, m)	MAXALIGN(offsetof(HnswNeighborTupleData, indextids) + ((level) + 2) * (m) * sizeof(ItemPointerData))
 
@@ -101,6 +99,7 @@ typedef struct HnswElementData
 	List	   *heaptids;
 	uint8		level;
 	uint8		deleted;
+	bool		loaded;
 	HnswNeighborArray *neighbors;
 	BlockNumber blkno;
 	OffsetNumber offno;
@@ -115,13 +114,11 @@ typedef struct HnswCandidate
 {
 	HnswElement element;
 	float		distance;
-	bool		closer;
 }			HnswCandidate;
 
 typedef struct HnswNeighborArray
 {
 	int			length;
-	bool		closerSet;
 	HnswCandidate *items;
 }			HnswNeighborArray;
 

src/hnswbuild.c

@@ -82,6 +82,7 @@ HnswBuildAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **
 	HnswPageGetOpaque(*page)->nextblkno = BufferGetBlockNumber(newbuf);
 
 	/* Commit */
+	MarkBufferDirty(*buf);
 	GenericXLogFinish(*state);
 	UnlockReleaseBuffer(*buf);
 
@@ -116,11 +117,11 @@ CreateElementPages(HnswBuildState * buildstate)
 	ListCell   *lc;
 
 	/* Calculate sizes */
-	maxSize = HNSW_MAX_SIZE;
+	maxSize = BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData));
 
 	/* Allocate once */
-	etup = palloc0(BLCKSZ);
-	ntup = palloc0(BLCKSZ);
+	etup = palloc0(maxSize);
+	ntup = palloc0(maxSize);
 
 	/* Prepare first page */
 	buf = HnswNewBuffer(index, forkNum);
@@ -178,6 +179,7 @@ CreateElementPages(HnswBuildState * buildstate)
 	insertPage = BufferGetBlockNumber(buf);
 
 	/* Commit */
+	MarkBufferDirty(buf);
 	GenericXLogFinish(state);
 	UnlockReleaseBuffer(buf);
 
@@ -225,6 +227,7 @@ CreateNeighborPages(HnswBuildState * buildstate)
 			elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
 
 		/* Commit */
+		MarkBufferDirty(buf);
 		GenericXLogFinish(state);
 		UnlockReleaseBuffer(buf);
 	}
@@ -375,13 +378,11 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 	if (inserted)
 	{
 		element->value = datumCopy(element->value, false, -1);
+		element->loaded = true;
 		buildstate->elements = lappend(buildstate->elements, element);
 	}
 	else
-	{
-		element->value = PointerGetDatum(NULL);
 		HnswFreeElement(element);
-	}
 
 	/* Reset memory context */
 	MemoryContextReset(buildstate->tmpCtx);

src/hnswinsert.c

@@ -134,7 +134,7 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 	etupSize = HNSW_ELEMENT_TUPLE_SIZE(e->value);
 	ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(e->level, m);
 	combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
-	maxSize = HNSW_MAX_SIZE;
+	maxSize = BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData));
 	minCombinedSize = etupSize + HNSW_NEIGHBOR_TUPLE_SIZE(0, m) + sizeof(ItemIdData);
 
 	/* Prepare element tuple */
@@ -201,6 +201,8 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 			HnswInsertAppendPage(index, &newbuf, &newpage, state, page);
 
 			/* Commit */
+			MarkBufferDirty(newbuf);
+			MarkBufferDirty(buf);
 			GenericXLogFinish(state);
 
 			/* Unlock previous buffer */
@@ -267,6 +269,9 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 	}
 
 	/* Commit */
+	MarkBufferDirty(buf);
+	if (nbuf != buf)
+		MarkBufferDirty(nbuf);
 	GenericXLogFinish(state);
 	UnlockReleaseBuffer(buf);
 	if (nbuf != buf)
@@ -385,6 +390,7 @@ HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswE
 					elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
 
 				/* Commit */
+				MarkBufferDirty(buf);
 				GenericXLogFinish(state);
 			}
 			else
@@ -438,6 +444,7 @@ HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup)
 		elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
 
 	/* Commit */
+	MarkBufferDirty(buf);
 	GenericXLogFinish(state);
 	UnlockReleaseBuffer(buf);
 

src/hnswscan.c

@@ -113,6 +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;
 }
 
@@ -160,25 +166,11 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
 		if (scan->orderByData == NULL)
 			elog(ERROR, "cannot scan hnsw index without order");
 
-		/* Requires MVCC-compliant snapshot as not able to maintain a pin */
-		/* https://www.postgresql.org/docs/current/index-locking.html */
-		if (!IsMVCCSnapshot(scan->xs_snapshot))
-			elog(ERROR, "non-MVCC snapshots are not supported with hnsw");
-
 		/* Get scan value */
 		value = GetScanValue(scan);
 
-		/*
-		 * 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);
-
 		so->w = GetScanItems(scan, value);
 
-		/* Release shared lock */
-		UnlockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
-
 		so->first = false;
 	}
 
@@ -206,6 +198,15 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
 		scan->xs_ctup.t_self = *heaptid;
 #endif
 
+		/*
+		 * 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
+		 */
+
 		scan->xs_recheckorderby = false;
 		return true;
 	}
@@ -222,6 +223,9 @@ hnswendscan(IndexScanDesc scan)
 {
 	HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
 
+	/* Release shared lock */
+	UnlockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
+
 	MemoryContextDelete(so->tmpCtx);
 
 	pfree(so);

src/hnswutils.c

@@ -132,6 +132,7 @@ HnswInitRegisterPage(Relation index, Buffer *buf, Page *page, GenericXLogState *
 void
 HnswCommitBuffer(Buffer buf, GenericXLogState *state)
 {
+	MarkBufferDirty(buf);
 	GenericXLogFinish(state);
 	UnlockReleaseBuffer(buf);
 }
@@ -154,7 +155,6 @@ HnswInitNeighbors(HnswElement element, int m)
 		a = &element->neighbors[lc];
 		a->length = 0;
 		a->items = palloc(sizeof(HnswCandidate) * lm);
-		a->closerSet = false;
 	}
 }
 
@@ -202,7 +202,7 @@ HnswFreeElement(HnswElement element)
 {
 	HnswFreeNeighbors(element);
 	list_free_deep(element->heaptids);
-	if (DatumGetPointer(element->value))
+	if (element->loaded)
 		pfree(DatumGetPointer(element->value));
 	pfree(element);
 }
@@ -230,7 +230,7 @@ HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno)
 	element->blkno = blkno;
 	element->offno = offno;
 	element->neighbors = NULL;
-	element->value = PointerGetDatum(NULL);
+	element->loaded = false;
 	return element;
 }
 
@@ -466,6 +466,7 @@ HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHe
 		Datum		value = PointerGetDatum(&etup->value);
 
 		element->value = datumCopy(value, false, -1);
+		element->loaded = true;
 	}
 }
 
@@ -710,34 +711,6 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
 	return w;
 }
 
-/*
- * Compare candidate distances
- */
-static int
-#if PG_VERSION_NUM >= 130000
-CompareCandidateDistances(const ListCell *a, const ListCell *b)
-#else
-CompareCandidateDistances(const void *a, const void *b)
-#endif
-{
-	HnswCandidate *hca = lfirst((ListCell *) a);
-	HnswCandidate *hcb = lfirst((ListCell *) b);
-
-	if (hca->distance < hcb->distance)
-		return 1;
-
-	if (hca->distance > hcb->distance)
-		return -1;
-
-	if (hca->element < hcb->element)
-		return 1;
-
-	if (hca->element > hcb->element)
-		return -1;
-
-	return 0;
-}
-
 /*
  * Calculate the distance between elements
  */
@@ -794,77 +767,33 @@ CheckElementCloser(HnswCandidate * e, List *r, int lc, FmgrInfo *procinfo, Oid c
  * Algorithm 4 from paper
  */
 static List *
-SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswElement e2, HnswCandidate * newCandidate, HnswCandidate * *pruned, bool sortCandidates)
+SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswCandidate * *pruned)
 {
 	List	   *r = NIL;
 	List	   *w = list_copy(c);
 	pairingheap *wd;
-	bool		mustCalculate = !e2->neighbors[lc].closerSet;
-	List	   *added = NIL;
-	bool		removedAny = false;
 
 	if (list_length(w) <= m)
 		return w;
 
 	wd = pairingheap_allocate(CompareNearestCandidates, NULL);
 
-	/* Ensure order of candidates is deterministic for closer caching */
-	if (sortCandidates)
-		list_sort(w, CompareCandidateDistances);
-
 	while (list_length(w) > 0 && list_length(r) < m)
 	{
 		/* Assumes w is already ordered desc */
 		HnswCandidate *e = llast(w);
+		bool		closer;
 
 		w = list_delete_last(w);
 
-		/* Use previous state of r and wd to skip work when possible */
-		if (mustCalculate)
-			e->closer = CheckElementCloser(e, r, lc, procinfo, collation);
-		else if (list_length(added) > 0)
-		{
-			/*
-			 * If the current candidate was closer, we only need to compare it
-			 * with the other candidates that we have added.
-			 */
-			if (e->closer)
-			{
-				e->closer = CheckElementCloser(e, added, lc, procinfo, collation);
+		closer = CheckElementCloser(e, r, lc, procinfo, collation);
 
-				if (!e->closer)
-					removedAny = true;
-			}
-			else
-			{
-				/*
-				 * If we have removed any candidates from closer, a candidate
-				 * that was not closer earlier might now be.
-				 */
-				if (removedAny)
-				{
-					e->closer = CheckElementCloser(e, r, lc, procinfo, collation);
-					if (e->closer)
-						added = lappend(added, e);
-				}
-			}
-		}
-		else if (e == newCandidate)
-		{
-			e->closer = CheckElementCloser(e, r, lc, procinfo, collation);
-			if (e->closer)
-				added = lappend(added, e);
-		}
-
-		if (e->closer)
+		if (closer)
 			r = lappend(r, e);
 		else
 			pairingheap_add(wd, &(CreatePairingHeapNode(e)->ph_node));
 	}
 
-	/* Cached value can only be used in future if sorted deterministically */
-	e2->neighbors[lc].closerSet = sortCandidates;
-
 	/* Keep pruned connections */
 	while (!pairingheap_is_empty(wd) && list_length(r) < m)
 		r = lappend(r, ((HnswPairingHeapNode *) pairingheap_remove_first(wd))->inner);
@@ -918,6 +847,28 @@ AddConnections(HnswElement element, List *neighbors, int m, int lc)
 		a->items[a->length++] = *((HnswCandidate *) lfirst(lc2));
 }
 
+/*
+ * Compare candidate distances
+ */
+static int
+#if PG_VERSION_NUM >= 130000
+CompareCandidateDistances(const ListCell *a, const ListCell *b)
+#else
+CompareCandidateDistances(const void *a, const void *b)
+#endif
+{
+	HnswCandidate *hca = lfirst((ListCell *) a);
+	HnswCandidate *hcb = lfirst((ListCell *) b);
+
+	if (hca->distance < hcb->distance)
+		return 1;
+
+	if (hca->distance > hcb->distance)
+		return -1;
+
+	return 0;
+}
+
 /*
  * Update connections
  */
@@ -953,7 +904,7 @@ HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int
 			{
 				HnswCandidate *hc3 = &currentNeighbors->items[i];
 
-				if (!DatumGetPointer(hc3->element->value))
+				if (!hc3->element->loaded)
 					HnswLoadElement(hc3->element, &hc3->distance, &q, index, procinfo, collation, true);
 				else
 					hc3->distance = GetCandidateDistance(hc3, q, procinfo, collation);
@@ -971,12 +922,13 @@ HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int
 		{
 			List	   *c = NIL;
 
-			/* Add candidates */
+			/* Add and sort candidates */
 			for (int i = 0; i < currentNeighbors->length; i++)
 				c = lappend(c, &currentNeighbors->items[i]);
 			c = lappend(c, &hc2);
+			list_sort(c, CompareCandidateDistances);
 
-			SelectNeighbors(c, m, lc, procinfo, collation, hc->element, &hc2, &pruned, true);
+			SelectNeighbors(c, m, lc, procinfo, collation, &pruned);
 
 			/* Should not happen */
 			if (pruned == NULL)
@@ -1075,12 +1027,7 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
 		else
 			lw = w;
 
-		/*
-		 * Candidates are sorted, but not deterministically. Could set
-		 * sortCandidates to true for in-memory builds to enable closer
-		 * caching, but there does not seem to be a difference in performance.
-		 */
-		neighbors = SelectNeighbors(lw, lm, lc, procinfo, collation, element, NULL, NULL, false);
+		neighbors = SelectNeighbors(lw, lm, lc, procinfo, collation, NULL);
 
 		AddConnections(element, neighbors, lm, lc);
 

src/hnswvacuum.c

@@ -128,7 +128,10 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 		blkno = HnswPageGetOpaque(page)->nextblkno;
 
 		if (updated)
+		{
+			MarkBufferDirty(buf);
 			GenericXLogFinish(state);
+		}
 		else
 			GenericXLogAbort(state);
 
@@ -226,6 +229,7 @@ RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswEleme
 		elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
 
 	/* Commit */
+	MarkBufferDirty(buf);
 	GenericXLogFinish(state);
 	UnlockReleaseBuffer(buf);
 
@@ -547,6 +551,9 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 				elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
 
 			/* Commit */
+			MarkBufferDirty(buf);
+			if (nbuf != buf)
+				MarkBufferDirty(nbuf);
 			GenericXLogFinish(state);
 			if (nbuf != buf)
 				UnlockReleaseBuffer(nbuf);

src/ivfbuild.c

@@ -506,11 +506,9 @@ CreateListPages(Relation index, VectorArray centers, int dimensions,
 	Buffer		buf;
 	Page		page;
 	GenericXLogState *state;
-	Size		listSize;
 	IvfflatList list;
 
-	listSize = MAXALIGN(IVFFLAT_LIST_SIZE(dimensions));
-	list = palloc(listSize);
+	list = palloc0(BLCKSZ);
 
 	buf = IvfflatNewBuffer(index, forkNum);
 	IvfflatInitRegisterPage(index, &buf, &page, &state);
@@ -518,11 +516,13 @@ CreateListPages(Relation index, VectorArray centers, int dimensions,
 	for (int i = 0; i < lists; i++)
 	{
 		OffsetNumber offno;
+		Datum		center = PointerGetDatum(VectorArrayGet(centers, i));
+		Size		listSize = MAXALIGN(IVFFLAT_LIST_SIZE(center));
 
 		/* Load list */
 		list->startPage = InvalidBlockNumber;
 		list->insertPage = InvalidBlockNumber;
-		memcpy(&list->center, VectorArrayGet(centers, i), VECTOR_SIZE(dimensions));
+		memcpy(&list->center, DatumGetPointer(center), VARSIZE_ANY(center));
 
 		/* Ensure free space */
 		if (PageGetFreeSpace(page) < listSize)

src/ivfflat.h

@@ -52,7 +52,7 @@
 #define PROGRESS_IVFFLAT_PHASE_ASSIGN	3
 #define PROGRESS_IVFFLAT_PHASE_LOAD		4
 
-#define IVFFLAT_LIST_SIZE(_dim)	(offsetof(IvfflatListData, center) + VECTOR_SIZE(_dim))
+#define IVFFLAT_LIST_SIZE(_datum)	(offsetof(IvfflatListData, center) + VARSIZE_ANY(_datum))
 
 #define IvfflatPageGetOpaque(page)	((IvfflatPageOpaque) PageGetSpecialPointer(page))
 #define IvfflatPageGetMeta(page)	((IvfflatMetaPageData *) PageGetContents(page))
@@ -229,7 +229,7 @@ typedef struct IvfflatListData
 {
 	BlockNumber startPage;
 	BlockNumber insertPage;
-	Vector		center;
+	char		center[FLEXIBLE_ARRAY_MEMBER];
 }			IvfflatListData;
 
 typedef IvfflatListData * IvfflatList;
@@ -246,6 +246,8 @@ typedef struct IvfflatScanOpaqueData
 	int			probes;
 	int			dimensions;
 	bool		first;
+	Buffer		buf;
+	ItemPointerData heaptid;
 
 	/* Sorting */
 	Tuplesortstate *sortstate;

src/ivfinsert.c

@@ -99,7 +99,7 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
 
 	/* Get tuple size */
 	itemsz = MAXALIGN(IndexTupleSize(itup));
-	Assert(itemsz <= BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(IvfflatPageOpaqueData)) - sizeof(ItemIdData));
+	Assert(itemsz <= BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(IvfflatPageOpaqueData)));
 
 	/* Find a page to insert the item */
 	for (;;)
@@ -142,6 +142,8 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
 			IvfflatPageGetOpaque(page)->nextblkno = insertPage;
 
 			/* Commit */
+			MarkBufferDirty(newbuf);
+			MarkBufferDirty(buf);
 			GenericXLogFinish(state);
 
 			/* Unlock previous buffer */

src/ivfscan.c

@@ -143,6 +143,10 @@ GetScanItems(IndexScanDesc scan, Datum value)
 				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);
 
@@ -157,6 +161,8 @@ 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_isnull[2] = false;
 				ExecStoreVirtualTuple(slot);
 
 				tuplesort_puttupleslot(so->sortstate, slot);
@@ -181,6 +187,55 @@ GetScanItems(IndexScanDesc scan, Datum value)
 	tuplesort_performsort(so->sortstate);
 }
 
+/*
+ * Mark prior tuple as dead
+ */
+static void
+MarkPriorTupleDead(IndexScanDesc scan)
+{
+	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
+	Buffer		buf = so->buf;
+	Page		page;
+	OffsetNumber maxoffno;
+
+	/* 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);
+	maxoffno = PageGetMaxOffsetNumber(page);
+
+	for (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
+	{
+		ItemId		itemid = PageGetItemId(page, offno);
+		IndexTuple	itup = (IndexTuple) PageGetItem(page, itemid);
+
+		/*
+		 * 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);
+			}
+
+			break;
+		}
+	}
+
+	/* Unlock buffer */
+	LockBuffer(buf, BUFFER_LOCK_UNLOCK);
+}
+
 /*
  * Prepare for an index scan
  */
@@ -206,7 +261,9 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 		probes = lists;
 
 	so = (IvfflatScanOpaque) palloc(offsetof(IvfflatScanOpaqueData, lists) + probes * sizeof(IvfflatScanList));
+	so->buf = InvalidBuffer;
 	so->first = true;
+	ItemPointerSetInvalid(&so->heaptid);
 	so->probes = probes;
 	so->dimensions = dimensions;
 
@@ -217,12 +274,13 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 
 	/* Create tuple description for sorting */
 #if PG_VERSION_NUM >= 120000
-	so->tupdesc = CreateTemplateTupleDesc(2);
+	so->tupdesc = CreateTemplateTupleDesc(3);
 #else
-	so->tupdesc = CreateTemplateTupleDesc(2, false);
+	so->tupdesc = CreateTemplateTupleDesc(3, false);
 #endif
 	TupleDescInitEntry(so->tupdesc, (AttrNumber) 1, "distance", FLOAT8OID, -1, 0);
 	TupleDescInitEntry(so->tupdesc, (AttrNumber) 2, "heaptid", TIDOID, -1, 0);
+	TupleDescInitEntry(so->tupdesc, (AttrNumber) 3, "indexblkno", INT4OID, -1, 0);
 
 	/* Prep sort */
 	so->sortstate = tuplesort_begin_heap(so->tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, work_mem, NULL, false);
@@ -254,6 +312,7 @@ 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)
@@ -288,11 +347,6 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 		if (scan->orderByData == NULL)
 			elog(ERROR, "cannot scan ivfflat index without order");
 
-		/* Requires MVCC-compliant snapshot as not able to pin during sorting */
-		/* https://www.postgresql.org/docs/current/index-locking.html */
-		if (!IsMVCCSnapshot(scan->xs_snapshot))
-			elog(ERROR, "non-MVCC snapshots are not supported with ivfflat");
-
 		if (scan->orderByData->sk_flags & SK_ISNULL)
 			value = PointerGetDatum(InitVector(so->dimensions));
 		else
@@ -316,10 +370,17 @@ 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 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 = *heaptid;
@@ -327,6 +388,21 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 		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);
+
+		/*
+		 * An index scan must maintain a pin on the index page holding the
+		 * item last returned by amgettuple
+		 *
+		 * https://www.postgresql.org/docs/current/index-locking.html
+		 */
+		so->buf = ReadBuffer(scan->indexRelation, indexblkno);
+
 		scan->xs_recheckorderby = false;
 		return true;
 	}
@@ -342,6 +418,10 @@ ivfflatendscan(IndexScanDesc scan)
 {
 	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
 
+	/* Release pin */
+	if (BufferIsValid(so->buf))
+		ReleaseBuffer(so->buf);
+
 	pairingheap_free(so->listQueue);
 	tuplesort_end(so->sortstate);
 

src/ivfutils.c

@@ -136,6 +136,7 @@ IvfflatInitRegisterPage(Relation index, Buffer *buf, Page *page, GenericXLogStat
 void
 IvfflatCommitBuffer(Buffer buf, GenericXLogState *state)
 {
+	MarkBufferDirty(buf);
 	GenericXLogFinish(state);
 	UnlockReleaseBuffer(buf);
 }
@@ -159,6 +160,8 @@ IvfflatAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **st
 	IvfflatInitPage(newbuf, newpage);
 
 	/* Commit */
+	MarkBufferDirty(*buf);
+	MarkBufferDirty(newbuf);
 	GenericXLogFinish(*state);
 
 	/* Unlock */

src/ivfvacuum.c

@@ -107,6 +107,7 @@ ivfflatbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 				{
 					/* Delete tuples */
 					PageIndexMultiDelete(page, deletable, ndeletable);
+					MarkBufferDirty(buf);
 					GenericXLogFinish(state);
 				}
 				else

test/expected/functions.out

@@ -54,21 +54,21 @@ SELECT vector_norm('[3e37,4e37]')::real;
        5e+37
 (1 row)
 
-SELECT l2_distance('[0,0]', '[3,4]');
+SELECT l2_distance('[0,0]'::vector, '[3,4]');
  l2_distance 
 -------------
            5
 (1 row)
 
-SELECT l2_distance('[0,0]', '[0,1]');
+SELECT l2_distance('[0,0]'::vector, '[0,1]');
  l2_distance 
 -------------
            1
 (1 row)
 
-SELECT l2_distance('[1,2]', '[3]');
+SELECT l2_distance('[1,2]'::vector, '[3]');
 ERROR:  different vector dimensions 2 and 1
-SELECT l2_distance('[3e38]', '[-3e38]');
+SELECT l2_distance('[3e38]'::vector, '[-3e38]');
  l2_distance 
 -------------
     Infinity

test/sql/functions.sql

@@ -13,10 +13,10 @@ SELECT vector_norm('[3,4]');
 SELECT vector_norm('[0,1]');
 SELECT vector_norm('[3e37,4e37]')::real;
 
-SELECT l2_distance('[0,0]', '[3,4]');
-SELECT l2_distance('[0,0]', '[0,1]');
-SELECT l2_distance('[1,2]', '[3]');
-SELECT l2_distance('[3e38]', '[-3e38]');
+SELECT l2_distance('[0,0]'::vector, '[3,4]');
+SELECT l2_distance('[0,0]'::vector, '[0,1]');
+SELECT l2_distance('[1,2]'::vector, '[3]');
+SELECT l2_distance('[3e38]'::vector, '[-3e38]');
 
 SELECT inner_product('[1,2]', '[3,4]');
 SELECT inner_product('[1,2]', '[3]');

test/t/019_hnsw_array.pl

@@ -57,26 +57,6 @@ $node->safe_psql("postgres",
 	"INSERT INTO tst SELECT i, ARRAY[random(), random(), random()] FROM generate_series(1, 10000) i;"
 );
 
-$node->safe_psql("postgres", qq(
-	CREATE FUNCTION float4_l2_distance(float4[], float4[]) RETURNS float8
-		AS 'BEGIN RETURN l2_distance(\$1::vector, \$2::vector); END;'
-		LANGUAGE plpgsql IMMUTABLE STRICT PARALLEL SAFE;
-
-	CREATE FUNCTION float4_l2_squared_distance(float4[], float4[]) RETURNS float8
-		AS 'BEGIN RETURN vector_l2_squared_distance(\$1::vector, \$2::vector); END;'
-		LANGUAGE plpgsql IMMUTABLE STRICT PARALLEL SAFE;
-
-	CREATE OPERATOR <-> (
-		LEFTARG = float4[], RIGHTARG = float4[], PROCEDURE = float4_l2_distance,
-		COMMUTATOR = '<->'
-	);
-
-	CREATE OPERATOR CLASS float4_l2_ops
-		FOR TYPE float4[] USING hnsw AS
-		OPERATOR 1 <-> (float4[], float4[]) FOR ORDER BY float_ops,
-		FUNCTION 1 float4_l2_squared_distance(float4[], float4[]);
-));
-
 # Generate queries
 for (1 .. 20)
 {

vector.control

@@ -1,4 +1,4 @@
 comment = 'vector data type and ivfflat and hnsw access methods'
-default_version = '0.5.1'
+default_version = '0.5.0'
 module_pathname = '$libdir/vector'
 relocatable = true