Include vectors with zero norm in indexes for cosine distance

The HnswInitElement() allocate an element with not initialize value
filed, which may has garbage that lead HnswFreeElement() free an
invalid pointer.

CHANGELOG.md

@@ -1,6 +1,6 @@
-## 0.6.0 (unreleased)
+## 0.5.2 (unreleased)
 
-- Added support for sparse vectors
+- Added support for on-disk parallel index builds for HNSW
 
 ## 0.5.1 (2023-10-10)
 

Makefile

@@ -3,8 +3,8 @@ EXTVERSION = 0.5.1
 
 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/svector.o src/vector.o
+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
-HEADERS = src/svector.h src/vector.h
+HEADERS = src/vector.h
 
 TESTS = $(wildcard test/sql/*.sql)
 REGRESS = $(patsubst test/sql/%.sql,%,$(TESTS))

Makefile.win

@@ -1,8 +1,8 @@
 EXTENSION = vector
 EXTVERSION = 0.5.1
 
-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\svector.obj src\vector.obj
+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
-HEADERS = src\svector.h src\vector.h
+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=$(EXTENSION)

README.md

@@ -269,6 +269,8 @@ Specify HNSW parameters
 CREATE INDEX ON items USING hnsw (embedding vector_l2_ops) WITH (m = 16, ef_construction = 64);
 ```
 
+A higher value of `ef_construction` provides better recall at the cost of index build time / insert speed.
+
 ### Query Options
 
 Specify the size of the dynamic candidate list for search (40 by default)
@@ -369,26 +371,6 @@ To speed up queries with an IVFFlat index, increase the number of inverted lists
 CREATE INDEX ON items USING ivfflat (embedding vector_l2_ops) WITH (lists = 1000);
 ```
 
-## Sparse Vectors
-
-Create a sparse vector column with 10 dimensions
-
-```sql
-CREATE TABLE items (id bigserial PRIMARY KEY, embedding svector(10));
-```
-
-Insert vectors
-
-```sql
-INSERT INTO items (embedding) VALUES ('(0,1),(1,2),(2,3)|10|'), ('(0,4),(1,5),(4,6)|10|');
-```
-
-Get the nearest neighbors by L2 distance
-
-```sql
-SELECT * FROM items ORDER BY embedding <-> '(0,3),(1,1),(2,2)|10|' LIMIT 5;
-```
-
 ## Languages
 
 Use pgvector from any language with a Postgres client. You can even generate and store vectors in one language and query them in another.

sql/vector--0.5.1--0.6.0.sql

@@ -1,79 +0,0 @@
--- complain if script is sourced in psql, rather than via CREATE EXTENSION
-\echo Use "ALTER EXTENSION vector UPDATE TO '0.6.0'" to load this file. \quit
-
-CREATE TYPE svector;
-
-CREATE FUNCTION svector_in(cstring, oid, integer) RETURNS svector
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION svector_out(svector) RETURNS cstring
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION svector_typmod_in(cstring[]) RETURNS integer
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION svector_recv(internal, oid, integer) RETURNS svector
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION svector_send(svector) RETURNS bytea
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE TYPE svector (
-	INPUT     = svector_in,
-	OUTPUT    = svector_out,
-	TYPMOD_IN = svector_typmod_in,
-	RECEIVE   = svector_recv,
-	SEND      = svector_send,
-	STORAGE   = external
-);
-
-CREATE FUNCTION l2_distance(svector, svector) RETURNS float8
-	AS 'MODULE_PATHNAME', 'svector_l2_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION inner_product(svector, svector) RETURNS float8
-	AS 'MODULE_PATHNAME', 'svector_inner_product' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION cosine_distance(svector, svector) RETURNS float8
-	AS 'MODULE_PATHNAME', 'svector_cosine_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION jaccard_distance(svector, svector) RETURNS float8
-	AS 'MODULE_PATHNAME', 'svector_jaccard_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION svector_l2_squared_distance(svector, svector) RETURNS float8
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION svector_negative_inner_product(svector, svector) RETURNS float8
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION svector(svector, integer, boolean) RETURNS svector
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION vector_to_svector(vector, integer, boolean) RETURNS svector
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION svector_to_vector(svector, integer, boolean) RETURNS vector
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE CAST (svector AS svector)
-	WITH FUNCTION svector(svector, integer, boolean) AS IMPLICIT;
-
-CREATE CAST (svector AS vector)
-	WITH FUNCTION svector_to_vector(svector, integer, boolean) AS IMPLICIT;
-
-CREATE CAST (vector AS svector)
-	WITH FUNCTION vector_to_svector(vector, integer, boolean) AS IMPLICIT;
-
-CREATE OPERATOR <-> (
-	LEFTARG = svector, RIGHTARG = svector, PROCEDURE = l2_distance,
-	COMMUTATOR = '<->'
-);
-
-CREATE OPERATOR <#> (
-	LEFTARG = svector, RIGHTARG = svector, PROCEDURE = svector_negative_inner_product,
-	COMMUTATOR = '<#>'
-);
-
-CREATE OPERATOR <=> (
-	LEFTARG = svector, RIGHTARG = svector, PROCEDURE = cosine_distance,
-	COMMUTATOR = '<=>'
-);

sql/vector.sql

@@ -290,92 +290,3 @@ CREATE OPERATOR CLASS vector_cosine_ops
 	OPERATOR 1 <=> (vector, vector) FOR ORDER BY float_ops,
 	FUNCTION 1 vector_negative_inner_product(vector, vector),
 	FUNCTION 2 vector_norm(vector);
-
---- svector type
-
-CREATE TYPE svector;
-
-CREATE FUNCTION svector_in(cstring, oid, integer) RETURNS svector
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION svector_out(svector) RETURNS cstring
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION svector_typmod_in(cstring[]) RETURNS integer
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION svector_recv(internal, oid, integer) RETURNS svector
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION svector_send(svector) RETURNS bytea
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE TYPE svector (
-	INPUT     = svector_in,
-	OUTPUT    = svector_out,
-	TYPMOD_IN = svector_typmod_in,
-	RECEIVE   = svector_recv,
-	SEND      = svector_send,
-	STORAGE   = external
-);
-
--- svector functions
-
-CREATE FUNCTION l2_distance(svector, svector) RETURNS float8
-	AS 'MODULE_PATHNAME', 'svector_l2_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION inner_product(svector, svector) RETURNS float8
-	AS 'MODULE_PATHNAME', 'svector_inner_product' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION cosine_distance(svector, svector) RETURNS float8
-	AS 'MODULE_PATHNAME', 'svector_cosine_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION jaccard_distance(svector, svector) RETURNS float8
-	AS 'MODULE_PATHNAME', 'svector_jaccard_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
--- svector private functions
-
-CREATE FUNCTION svector_l2_squared_distance(svector, svector) RETURNS float8
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION svector_negative_inner_product(svector, svector) RETURNS float8
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
--- svector cast functions
-
-CREATE FUNCTION svector(svector, integer, boolean) RETURNS svector
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION vector_to_svector(vector, integer, boolean) RETURNS svector
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION svector_to_vector(svector, integer, boolean) RETURNS vector
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
--- svector casts
-
-CREATE CAST (svector AS svector)
-	WITH FUNCTION svector(svector, integer, boolean) AS IMPLICIT;
-
-CREATE CAST (svector AS vector)
-	WITH FUNCTION svector_to_vector(svector, integer, boolean) AS IMPLICIT;
-
-CREATE CAST (vector AS svector)
-	WITH FUNCTION vector_to_svector(vector, integer, boolean) AS IMPLICIT;
-
--- svector operators
-
-CREATE OPERATOR <-> (
-	LEFTARG = svector, RIGHTARG = svector, PROCEDURE = l2_distance,
-	COMMUTATOR = '<->'
-);
-
-CREATE OPERATOR <#> (
-	LEFTARG = svector, RIGHTARG = svector, PROCEDURE = svector_negative_inner_product,
-	COMMUTATOR = '<#>'
-);
-
-CREATE OPERATOR <=> (
-	LEFTARG = svector, RIGHTARG = svector, PROCEDURE = cosine_distance,
-	COMMUTATOR = '<=>'
-);

src/hnsw.c

@@ -14,6 +14,7 @@
 #endif
 
 int			hnsw_ef_search;
+bool		hnsw_enable_parallel_build;
 static relopt_kind hnsw_relopt_kind;
 
 /*
@@ -39,6 +40,11 @@ HnswInit(void)
 	DefineCustomIntVariable("hnsw.ef_search", "Sets the size of the dynamic candidate list for search",
 							"Valid range is 1..1000.", &hnsw_ef_search,
 							HNSW_DEFAULT_EF_SEARCH, HNSW_MIN_EF_SEARCH, HNSW_MAX_EF_SEARCH, PGC_USERSET, 0, NULL, NULL, NULL);
+
+	/* Behind a variable for now since can be slower than building in memory */
+	DefineCustomBoolVariable("hnsw.enable_parallel_build", "Enables or disables building indexes in parallel",
+							 NULL, &hnsw_enable_parallel_build,
+							 false, PGC_USERSET, 0, NULL, NULL, NULL);
 }
 
 /*

src/hnsw.h

@@ -4,6 +4,7 @@
 #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() */
@@ -14,6 +15,10 @@
 #error "Requires PostgreSQL 11+"
 #endif
 
+#if PG_VERSION_NUM < 120000
+#include "access/relscan.h"
+#endif
+
 #define HNSW_MAX_DIM 2000
 
 /* Support functions */
@@ -59,7 +64,7 @@
 
 #define HNSW_MAX_SIZE (BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - sizeof(ItemIdData))
 
-#define HNSW_ELEMENT_TUPLE_SIZE(_dim)	MAXALIGN(offsetof(HnswElementTupleData, vec) + VECTOR_SIZE(_dim))
+#define HNSW_ELEMENT_TUPLE_SIZE(size)	MAXALIGN(offsetof(HnswElementTupleData, data) + (size))
 #define HNSW_NEIGHBOR_TUPLE_SIZE(level, m)	MAXALIGN(offsetof(HnswNeighborTupleData, indextids) + ((level) + 2) * (m) * sizeof(ItemPointerData))
 
 #define HnswPageGetOpaque(page)	((HnswPageOpaque) PageGetSpecialPointer(page))
@@ -90,6 +95,7 @@
 
 /* Variables */
 extern int	hnsw_ef_search;
+extern bool hnsw_enable_parallel_build;
 
 typedef struct HnswNeighborArray HnswNeighborArray;
 
@@ -103,7 +109,7 @@ typedef struct HnswElementData
 	OffsetNumber offno;
 	OffsetNumber neighborOffno;
 	BlockNumber neighborPage;
-	Vector	   *vec;
+	Datum		value;
 }			HnswElementData;
 
 typedef HnswElementData * HnswElement;
@@ -136,6 +142,49 @@ 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 */
@@ -163,12 +212,16 @@ typedef struct HnswBuildState
 	HnswElement entryPoint;
 	double		ml;
 	int			maxLevel;
-	double		maxInMemoryElements;
+	long		memoryLeft;
 	bool		flushed;
 	Vector	   *normvec;
 
 	/* Memory */
 	MemoryContext tmpCtx;
+
+	/* Parallel builds */
+	HnswLeader *hnswleader;
+	HnswShared *hnswshared;
 }			HnswBuildState;
 
 typedef struct HnswMetaPageData
@@ -204,7 +257,7 @@ typedef struct HnswElementTupleData
 	ItemPointerData heaptids[HNSW_HEAPTIDS];
 	ItemPointerData neighbortid;
 	uint16		unused2;
-	Vector		vec;
+	Vector		data;
 }			HnswElementTupleData;
 
 typedef HnswElementTupleData * HnswElementTuple;
@@ -263,7 +316,7 @@ typedef struct HnswVacuumState
 int			HnswGetM(Relation index);
 int			HnswGetEfConstruction(Relation index);
 FmgrInfo   *HnswOptionalProcInfo(Relation index, uint16 procnum);
-bool		HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
+void		HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
 void		HnswCommitBuffer(Buffer buf, GenericXLogState *state);
 Buffer		HnswNewBuffer(Relation index, ForkNumber forkNum);
 void		HnswInitPage(Buffer buf, Page page);
@@ -289,6 +342,7 @@ void		HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation i
 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, int m);
+PGDLLEXPORT void HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc);
 
 /* Index access methods */
 IndexBuildResult *hnswbuild(Relation heap, Relation index, IndexInfo *indexInfo);

src/hnswbuild.c

@@ -2,12 +2,16 @@
 
 #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/datum.h"
 #include "utils/memutils.h"
 
 #if PG_VERSION_NUM >= 140000
@@ -35,6 +39,23 @@
 #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
  */
@@ -105,8 +126,7 @@ CreateElementPages(HnswBuildState * buildstate)
 {
 	Relation	index = buildstate->index;
 	ForkNumber	forkNum = buildstate->forkNum;
-	int			dimensions = buildstate->dimensions;
+	Size		etupAllocSize;
-	Size		etupSize;
 	Size		maxSize;
 	HnswElementTuple etup;
 	HnswNeighborTuple ntup;
@@ -117,11 +137,11 @@ CreateElementPages(HnswBuildState * buildstate)
 	ListCell   *lc;
 
 	/* Calculate sizes */
+	etupAllocSize = BLCKSZ;
 	maxSize = HNSW_MAX_SIZE;
-	etupSize = HNSW_ELEMENT_TUPLE_SIZE(dimensions);
 
 	/* Allocate once */
-	etup = palloc0(etupSize);
+	etup = palloc0(etupAllocSize);
 	ntup = palloc0(BLCKSZ);
 
 	/* Prepare first page */
@@ -133,15 +153,24 @@ CreateElementPages(HnswBuildState * buildstate)
 	foreach(lc, buildstate->elements)
 	{
 		HnswElement element = lfirst(lc);
+		Size		etupSize;
 		Size		ntupSize;
 		Size		combinedSize;
 
-		HnswSetElementTuple(etup, element);
+		/* Zero memory for each element */
+		MemSet(etup, 0, etupAllocSize);
 
 		/* Calculate sizes */
+		etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(DatumGetPointer(element->value)));
 		ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(element->level, buildstate->m);
 		combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
 
+		/* Initial size check */
+		if (etupSize > etupAllocSize)
+			elog(ERROR, "index tuple too large");
+
+		HnswSetElementTuple(etup, element);
+
 		/* Keep element and neighbors on the same page if possible */
 		if (PageGetFreeSpace(page) < etupSize || (combinedSize <= maxSize && PageGetFreeSpace(page) < combinedSize))
 			HnswBuildAppendPage(index, &buf, &page, &state, forkNum);
@@ -264,26 +293,26 @@ FlushPages(HnswBuildState * buildstate)
  * Insert tuple
  */
 static bool
-InsertTuple(Relation index, Datum *values, HnswElement element, HnswBuildState * buildstate, HnswElement * dup)
+InsertTuple(Relation index, Datum *values, HnswElement element, HnswBuildState * buildstate, HnswElement * dup, MemoryContext outerCtx)
 {
 	FmgrInfo   *procinfo = buildstate->procinfo;
 	Oid			collation = buildstate->collation;
 	HnswElement entryPoint = buildstate->entryPoint;
 	int			efConstruction = buildstate->efConstruction;
 	int			m = buildstate->m;
+	MemoryContext oldCtx;
 
 	/* Detoast once for all calls */
 	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
 
 	/* Normalize if needed */
 	if (buildstate->normprocinfo != NULL)
-	{
+		HnswNormValue(buildstate->normprocinfo, collation, &value, buildstate->normvec);
-		if (!HnswNormValue(buildstate->normprocinfo, collation, &value, buildstate->normvec))
-			return false;
-	}
 
 	/* Copy value to element so accessible outside of memory context */
-	memcpy(element->vec, DatumGetVector(value), VECTOR_SIZE(buildstate->dimensions));
+	oldCtx = MemoryContextSwitchTo(outerCtx);
+	element->value = datumCopy(value, false, -1);
+	MemoryContextSwitchTo(oldCtx);
 
 	/* Insert element in graph */
 	HnswInsertElement(element, entryPoint, NULL, procinfo, collation, m, efConstruction, false);
@@ -313,6 +342,21 @@ InsertTuple(Relation index, Datum *values, HnswElement element, HnswBuildState *
 	return *dup == NULL;
 }
 
+/*
+ * Get the memory used by an element
+ */
+static long
+HnswElementMemory(HnswElement e, int m)
+{
+	long		elementSize = sizeof(HnswElementData);
+
+	elementSize += sizeof(HnswNeighborArray) * (e->level + 1);
+	elementSize += sizeof(HnswCandidate) * (m * (e->level + 2));
+	elementSize += sizeof(ItemPointerData);
+	elementSize += VARSIZE_ANY(DatumGetPointer(e->value));
+	return elementSize;
+}
+
 /*
  * Callback for table_index_build_scan
  */
@@ -334,7 +378,7 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 	if (isnull[0])
 		return;
 
-	if (buildstate->indtuples >= buildstate->maxInMemoryElements)
+	if (buildstate->memoryLeft <= 0)
 	{
 		if (!buildstate->flushed)
 		{
@@ -349,7 +393,18 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 		oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
 
 		if (HnswInsertTuple(buildstate->index, values, isnull, tid, buildstate->heap))
-			UpdateProgress(PROGRESS_CREATEIDX_TUPLES_DONE, ++buildstate->indtuples);
+		{
+			if (buildstate->hnswshared)
+			{
+				HnswShared *hnswshared = buildstate->hnswshared;
+
+				SpinLockAcquire(&hnswshared->mutex);
+				UpdateProgress(PROGRESS_CREATEIDX_TUPLES_DONE, ++hnswshared->indtuples);
+				SpinLockRelease(&hnswshared->mutex);
+			}
+			else
+				UpdateProgress(PROGRESS_CREATEIDX_TUPLES_DONE, ++buildstate->indtuples);
+		}
 
 		/* Reset memory context */
 		MemoryContextSwitchTo(oldCtx);
@@ -360,13 +415,12 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 
 	/* Allocate necessary memory outside of memory context */
 	element = HnswInitElement(tid, buildstate->m, buildstate->ml, buildstate->maxLevel);
-	element->vec = palloc(VECTOR_SIZE(buildstate->dimensions));
 
 	/* Use memory context since detoast can allocate */
 	oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
 
 	/* Insert tuple */
-	inserted = InsertTuple(index, values, element, buildstate, &dup);
+	inserted = InsertTuple(index, values, element, buildstate, &dup, oldCtx);
 
 	/* Reset memory context */
 	MemoryContextSwitchTo(oldCtx);
@@ -374,31 +428,21 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 
 	/* Add outside memory context */
 	if (dup != NULL)
+	{
 		HnswAddHeapTid(dup, tid);
+		buildstate->memoryLeft -= sizeof(ItemPointerData);
+	}
 
 	/* Add to buildstate or free */
 	if (inserted)
+	{
 		buildstate->elements = lappend(buildstate->elements, element);
+		buildstate->memoryLeft -= HnswElementMemory(element, buildstate->m);
+	}
 	else
 		HnswFreeElement(element);
 }
 
-/*
- * Get the max number of elements that fit into maintenance_work_mem
- */
-static double
-HnswGetMaxInMemoryElements(int m, double ml, int dimensions)
-{
-	Size		elementSize = sizeof(HnswElementData);
-	double		avgLevel = -log(0.5) * ml;
-
-	elementSize += sizeof(HnswNeighborArray) * (avgLevel + 1);
-	elementSize += sizeof(HnswCandidate) * (m * (avgLevel + 2));
-	elementSize += sizeof(ItemPointerData);
-	elementSize += VECTOR_SIZE(dimensions);
-	return (maintenance_work_mem * 1024L) / elementSize;
-}
-
 /*
  * Initialize the build state
  */
@@ -436,7 +480,7 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 	buildstate->entryPoint = NULL;
 	buildstate->ml = HnswGetMl(buildstate->m);
 	buildstate->maxLevel = HnswGetMaxLevel(buildstate->m);
-	buildstate->maxInMemoryElements = HnswGetMaxInMemoryElements(buildstate->m, buildstate->ml, buildstate->dimensions);
+	buildstate->memoryLeft = maintenance_work_mem * 1024L;
 	buildstate->flushed = false;
 
 	/* Reuse for each tuple */
@@ -445,6 +489,9 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 	buildstate->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
 											   "Hnsw build temporary context",
 											   ALLOCSET_DEFAULT_SIZES);
+
+	buildstate->hnswleader = NULL;
+	buildstate->hnswshared = NULL;
 }
 
 /*
@@ -457,21 +504,374 @@ 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.memoryLeft = 0;
+	buildstate.flushed = true;
+	buildstate.hnswshared = hnswshared;
+#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;
+	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 */
+	if (hnsw_enable_parallel_build)
+		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,
+		buildstate->reltuples = table_index_build_scan(buildstate->heap, buildstate->index, buildstate->indexInfo,
-												   true, true, BuildCallback, (void *) buildstate, NULL);
+													   true, true, BuildCallback, (void *) buildstate, NULL);
 #else
-	buildstate->reltuples = IndexBuildHeapScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
+		buildstate->reltuples = IndexBuildHeapScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
-											   true, BuildCallback, (void *) buildstate, NULL);
+												   true, BuildCallback, (void *) buildstate, NULL);
 #endif
+	}
+
+	/* End parallel build */
+	if (buildstate->hnswleader)
+		HnswEndParallel(buildstate->hnswleader);
 }
 
 /*

src/hnswinsert.c

@@ -123,7 +123,6 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 	Size		minCombinedSize;
 	HnswElementTuple etup;
 	BlockNumber currentPage = insertPage;
-	int			dimensions = e->vec->dim;
 	HnswNeighborTuple ntup;
 	Buffer		nbuf;
 	Page		npage;
@@ -132,7 +131,7 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 	BlockNumber newInsertPage = InvalidBlockNumber;
 
 	/* Calculate sizes */
-	etupSize = HNSW_ELEMENT_TUPLE_SIZE(dimensions);
+	etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(DatumGetPointer(e->value)));
 	ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(e->level, m);
 	combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
 	maxSize = HNSW_MAX_SIZE;
@@ -405,8 +404,9 @@ HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup)
 	Buffer		buf;
 	Page		page;
 	GenericXLogState *state;
-	Size		etupSize = HNSW_ELEMENT_TUPLE_SIZE(dup->vec->dim);
+	ItemId		itemid;
 	HnswElementTuple etup;
+	Size		etupSize;
 	int			i;
 
 	/* Read page */
@@ -416,7 +416,9 @@ HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup)
 	page = GenericXLogRegisterBuffer(state, buf, 0);
 
 	/* Find space */
-	etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, dup->offno));
+	itemid = PageGetItemId(page, dup->offno);
+	etup = (HnswElementTuple) PageGetItem(page, itemid);
+	etupSize = ItemIdGetLength(itemid);
 	for (i = 0; i < HNSW_HEAPTIDS; i++)
 	{
 		if (!ItemPointerIsValid(&etup->heaptids[i]))
@@ -498,10 +500,7 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 	/* Normalize if needed */
 	normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
 	if (normprocinfo != NULL)
-	{
+		HnswNormValue(normprocinfo, collation, &value, NULL);
-		if (!HnswNormValue(normprocinfo, collation, &value, NULL))
-			return false;
-	}
 
 	/*
 	 * Get a shared lock. This allows vacuum to ensure no in-flight inserts
@@ -515,7 +514,7 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 
 	/* Create an element */
 	element = HnswInitElement(heap_tid, m, HnswGetMl(m), HnswGetMaxLevel(m));
-	element->vec = DatumGetVector(value);
+	element->value = value;
 
 	/* Prevent concurrent inserts when likely updating entry point */
 	if (entryPoint == NULL || element->level > entryPoint->level)

src/hnswutils.c

@@ -4,6 +4,7 @@
 
 #include "hnsw.h"
 #include "storage/bufmgr.h"
+#include "utils/datum.h"
 #include "vector.h"
 
 /*
@@ -54,7 +55,7 @@ HnswOptionalProcInfo(Relation index, uint16 procnum)
  * The caller needs to free the pointer stored in value
  * if it's different than the original value
  */
-bool
+void
 HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result)
 {
 	double		norm = DatumGetFloat8(FunctionCall1Coll(procinfo, collation, *value));
@@ -70,11 +71,7 @@ HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result)
 			result->x[i] = v->x[i] / norm;
 
 		*value = PointerGetDatum(result);
-
-		return true;
 	}
-
-	return false;
 }
 
 /*
@@ -176,6 +173,8 @@ HnswInitElement(ItemPointer heaptid, int m, double ml, int maxLevel)
 
 	HnswInitNeighbors(element, m);
 
+	element->value = PointerGetDatum(NULL);
+
 	return element;
 }
 
@@ -187,7 +186,8 @@ HnswFreeElement(HnswElement element)
 {
 	HnswFreeNeighbors(element);
 	list_free_deep(element->heaptids);
-	pfree(element->vec);
+	if (DatumGetPointer(element->value))
+		pfree(DatumGetPointer(element->value));
 	pfree(element);
 }
 
@@ -214,7 +214,7 @@ HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno)
 	element->blkno = blkno;
 	element->offno = offno;
 	element->neighbors = NULL;
-	element->vec = NULL;
+	element->value = PointerGetDatum(NULL);
 	return element;
 }
 
@@ -324,7 +324,7 @@ HnswSetElementTuple(HnswElementTuple etup, HnswElement element)
 		else
 			ItemPointerSetInvalid(&etup->heaptids[i]);
 	}
-	memcpy(&etup->vec, element->vec, VECTOR_SIZE(element->vec->dim));
+	memcpy(&etup->data, DatumGetPointer(element->value), VARSIZE_ANY(DatumGetPointer(element->value)));
 }
 
 /*
@@ -446,10 +446,7 @@ HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHe
 	}
 
 	if (loadVec)
-	{
+		element->value = datumCopy(PointerGetDatum(&etup->data), false, -1);
-		element->vec = palloc(VECTOR_SIZE(etup->vec.dim));
-		memcpy(element->vec, &etup->vec, VECTOR_SIZE(etup->vec.dim));
-	}
 }
 
 /*
@@ -476,7 +473,7 @@ HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index,
 
 	/* Calculate distance */
 	if (distance != NULL)
-		*distance = (float) DatumGetFloat8(FunctionCall2Coll(procinfo, collation, *q, PointerGetDatum(&etup->vec)));
+		*distance = (float) DatumGetFloat8(FunctionCall2Coll(procinfo, collation, *q, PointerGetDatum(&etup->data)));
 
 	UnlockReleaseBuffer(buf);
 }
@@ -487,7 +484,7 @@ HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index,
 static float
 GetCandidateDistance(HnswCandidate * hc, Datum q, FmgrInfo *procinfo, Oid collation)
 {
-	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, q, PointerGetDatum(hc->element->vec)));
+	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, q, hc->element->value));
 }
 
 /*
@@ -750,7 +747,7 @@ HnswGetDistance(HnswElement a, HnswElement b, int lc, FmgrInfo *procinfo, Oid co
 		}
 	}
 
-	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(a->vec), PointerGetDatum(b->vec)));
+	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, a->value, b->value));
 }
 
 /*
@@ -877,7 +874,7 @@ HnswFindDuplicate(HnswElement e)
 		HnswCandidate *neighbor = &neighbors->items[i];
 
 		/* Exit early since ordered by distance */
-		if (vector_cmp_internal(e->vec, neighbor->element->vec) != 0)
+		if (!datumIsEqual(e->value, neighbor->element->value, false, -1))
 			break;
 
 		/* Check for space */
@@ -930,13 +927,13 @@ HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int
 		/* Load elements on insert */
 		if (index != NULL)
 		{
-			Datum		q = PointerGetDatum(hc->element->vec);
+			Datum		q = hc->element->value;
 
 			for (int i = 0; i < currentNeighbors->length; i++)
 			{
 				HnswCandidate *hc3 = &currentNeighbors->items[i];
 
-				if (hc3->element->vec == NULL)
+				if (DatumGetPointer(hc3->element->value) == NULL)
 					HnswLoadElement(hc3->element, &hc3->distance, &q, index, procinfo, collation, true);
 				else
 					hc3->distance = GetCandidateDistance(hc3, q, procinfo, collation);
@@ -1017,7 +1014,7 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
 	List	   *w;
 	int			level = element->level;
 	int			entryLevel;
-	Datum		q = PointerGetDatum(element->vec);
+	Datum		q = element->value;
 	HnswElement skipElement = existing ? element : NULL;
 
 	/* No neighbors if no entry point */

src/hnswvacuum.c

@@ -62,7 +62,8 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 		/* Iterate over nodes */
 		for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
 		{
-			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, offno));
+			ItemId		itemid = PageGetItemId(page, offno);
+			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, itemid);
 			int			idx = 0;
 			bool		itemUpdated = false;
 
@@ -93,7 +94,7 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 
 				if (itemUpdated)
 				{
-					Size		etupSize = HNSW_ELEMENT_TUPLE_SIZE(etup->vec.dim);
+					Size		etupSize = ItemIdGetLength(itemid);
 
 					/* Mark rest as invalid */
 					for (int i = idx; i < HNSW_HEAPTIDS; i++)
@@ -477,7 +478,8 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 		/* Update element and neighbors together */
 		for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
 		{
-			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, offno));
+			ItemId		itemid = PageGetItemId(page, offno);
+			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, itemid);
 			HnswNeighborTuple ntup;
 			Size		etupSize;
 			Size		ntupSize;
@@ -505,7 +507,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 				continue;
 
 			/* Calculate sizes */
-			etupSize = HNSW_ELEMENT_TUPLE_SIZE(etup->vec.dim);
+			etupSize = ItemIdGetLength(itemid);
 			ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(etup->level, vacuumstate->m);
 
 			/* Get neighbor page */
@@ -528,7 +530,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 
 			/* Overwrite element */
 			etup->deleted = 1;
-			MemSet(&etup->vec.x, 0, etup->vec.dim * sizeof(float));
+			MemSet(&etup->data, 0, VARSIZE_ANY(&etup->data));
 
 			/* Overwrite neighbors */
 			for (int i = 0; i < ntup->count; i++)

src/ivfbuild.c

@@ -177,10 +177,7 @@ AddTupleToSort(Relation index, ItemPointer tid, Datum *values, IvfflatBuildState
 
 	/* Normalize if needed */
 	if (buildstate->normprocinfo != NULL)
-	{
+		IvfflatNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->normvec);
-		if (!IvfflatNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->normvec))
-			return;
-	}
 
 	/* Find the list that minimizes the distance */
 	for (int i = 0; i < centers->length; i++)

src/ivfinsert.c

@@ -83,10 +83,7 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
 	/* Normalize if needed */
 	normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
 	if (normprocinfo != NULL)
-	{
+		IvfflatNormValue(normprocinfo, index->rd_indcollation[0], &value, NULL);
-		if (!IvfflatNormValue(normprocinfo, index->rd_indcollation[0], &value, NULL))
-			return;
-	}
 
 	/* Find the insert page - sets the page and list info */
 	FindInsertPage(index, values, &insertPage, &listInfo);

src/svector.c

@@ -1,705 +0,0 @@
-#include "postgres.h"
-
-#include <math.h>
-
-#include "fmgr.h"
-#include "libpq/pqformat.h"
-#include "svector.h"
-#include "utils/array.h"
-#include "vector.h"
-
-#if PG_VERSION_NUM >= 120000
-#include "common/shortest_dec.h"
-#include "utils/float.h"
-#else
-#include <float.h>
-#include "utils/builtins.h"
-#endif
-
-/*
- * Ensure same dimensions
- */
-static inline void
-CheckDims(SVector * a, SVector * b)
-{
-	if (a->dim != b->dim)
-		ereport(ERROR,
-				(errcode(ERRCODE_DATA_EXCEPTION),
-				 errmsg("different svector dimensions %d and %d", a->dim, b->dim)));
-}
-
-/*
- * Ensure expected dimensions
- */
-static inline void
-CheckExpectedDim(int32 typmod, int dim)
-{
-	if (typmod != -1 && typmod != dim)
-		ereport(ERROR,
-				(errcode(ERRCODE_DATA_EXCEPTION),
-				 errmsg("expected %d dimensions, not %d", typmod, dim)));
-}
-
-/*
- * Ensure valid dimensions
- */
-static inline void
-CheckDim(int dim)
-{
-	if (dim < 1)
-		ereport(ERROR,
-				(errcode(ERRCODE_DATA_EXCEPTION),
-				 errmsg("svector must have at least 1 dimension")));
-
-	if (dim > SVECTOR_MAX_DIM)
-		ereport(ERROR,
-				(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
-				 errmsg("svector cannot have more than %d dimensions", SVECTOR_MAX_DIM)));
-}
-
-/*
- * Ensure valid nnz
- */
-static inline void
-CheckNnz(int nnz, int dim)
-{
-	if (nnz < 0)
-		ereport(ERROR,
-				(errcode(ERRCODE_DATA_EXCEPTION),
-				 errmsg("svector must have at least one element")));
-
-	if (nnz > dim)
-		ereport(ERROR,
-				(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
-				 errmsg("svector cannot have more elements than dimensions")));
-}
-
-/*
- * Ensure valid index
- */
-static inline void
-CheckIndex(int32 *indices, int i, int dim)
-{
-	int32		index = indices[i];
-
-	if (index < 0)
-		ereport(ERROR,
-				(errcode(ERRCODE_DATA_EXCEPTION),
-				 errmsg("index must not be negative")));
-
-	if (index >= dim)
-		ereport(ERROR,
-				(errcode(ERRCODE_DATA_EXCEPTION),
-				 errmsg("index must be less than dimensions")));
-
-	if (i > 0)
-	{
-		if (index < indices[i - 1])
-			ereport(ERROR,
-					(errcode(ERRCODE_DATA_EXCEPTION),
-					 errmsg("indexes must be in ascending order")));
-
-		if (index == indices[i - 1])
-			ereport(ERROR,
-					(errcode(ERRCODE_DATA_EXCEPTION),
-					 errmsg("indexes must not contain duplicates")));
-	}
-}
-
-/*
- * Ensure finite element
- */
-static inline void
-CheckElement(float value)
-{
-	if (isnan(value))
-		ereport(ERROR,
-				(errcode(ERRCODE_DATA_EXCEPTION),
-				 errmsg("NaN not allowed in svector")));
-
-	if (isinf(value))
-		ereport(ERROR,
-				(errcode(ERRCODE_DATA_EXCEPTION),
-				 errmsg("infinite value not allowed in svector")));
-}
-
-/*
- * Allocate and initialize a new sparse vector
- */
-SVector *
-InitSVector(int dim, int nnz)
-{
-	SVector    *result;
-	int			size;
-
-	size = SVECTOR_SIZE(nnz);
-	result = (SVector *) palloc0(size);
-	SET_VARSIZE(result, size);
-	result->dim = dim;
-	result->nnz = nnz;
-
-	return result;
-}
-
-/*
- * Convert textual representation to internal representation
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(svector_in);
-Datum
-svector_in(PG_FUNCTION_ARGS)
-{
-	char	   *str = PG_GETARG_CSTRING(0);
-	int32		typmod = PG_GETARG_INT32(2);
-	int			dim;
-	char	   *pt;
-	SVector    *result;
-	float	   *rvalues;
-	char	   *lit = pstrdup(str);
-	int			n;
-	int32	   *indices;
-	float	   *values;
-	int			index;
-	float		value;
-	int			maxNnz;
-	int			nnz = 0;
-
-	/* TODO Improve code and checks after deciding on format */
-
-	maxNnz = 1;
-	pt = str;
-	while (*pt != '\0')
-	{
-		if (*pt == ',')
-			maxNnz++;
-
-		pt++;
-	}
-	maxNnz /= 2;
-
-	indices = palloc(maxNnz * sizeof(int32));
-	values = palloc(maxNnz * sizeof(float));
-
-	while (sscanf(str, "(%d,%f)%n", &index, &value, &n) == 2)
-	{
-		/* TODO Better error */
-		if (nnz == maxNnz)
-			ereport(ERROR,
-					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-					 errmsg("ran out of buffer: \"%s\"", lit)));
-
-		/* TODO Decide whether to store zero values */
-		indices[nnz] = index;
-		values[nnz] = value;
-		nnz++;
-
-		str += n;
-
-		if (*str == ',')
-			str++;
-		else if (*str == '|')
-			break;
-		else
-			ereport(ERROR,
-					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-					 errmsg("malformed svector literal: \"%s\"", lit)));
-	}
-
-	if (sscanf(str, "|%d|%n", &dim, &n) != 1)
-		ereport(ERROR,
-				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-				 errmsg("malformed svector literal: \"%s\"", lit)));
-
-	str += n;
-
-	if (*str != '\0')
-		ereport(ERROR,
-				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-				 errmsg("malformed svector literal: \"%s\"", lit),
-				 errdetail("Junk after closing pipe.")));
-
-	pfree(lit);
-
-	CheckDim(dim);
-	CheckExpectedDim(typmod, dim);
-
-	result = InitSVector(dim, nnz);
-	rvalues = SVECTOR_VALUES(result);
-	for (int i = 0; i < nnz; i++)
-	{
-		result->indices[i] = indices[i];
-		rvalues[i] = values[i];
-
-		CheckIndex(result->indices, i, dim);
-		CheckElement(rvalues[i]);
-	}
-
-	PG_RETURN_POINTER(result);
-}
-
-/*
- * Convert internal representation to textual representation
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(svector_out);
-Datum
-svector_out(PG_FUNCTION_ARGS)
-{
-	SVector    *svector = PG_GETARG_SVECTOR_P(0);
-	float	   *values = SVECTOR_VALUES(svector);
-	char	   *buf;
-	char	   *ptr;
-	int			n;
-
-	/* TODO Improve code after deciding on format */
-
-#if PG_VERSION_NUM < 120000
-	int			ndig = FLT_DIG + extra_float_digits;
-
-	if (ndig < 1)
-		ndig = 1;
-
-#define FLOAT_SHORTEST_DECIMAL_LEN (ndig + 10)
-#endif
-
-	/* TODO Move */
-#define APPEND_CHAR(ptr, ch) (*(ptr)++ = (ch))
-
-	/* TODO Improve */
-	buf = (char *) palloc((FLOAT_SHORTEST_DECIMAL_LEN + 20) * svector->nnz + 20);
-	ptr = buf;
-
-	for (int i = 0; i < svector->nnz; i++)
-	{
-		if (i > 0)
-			APPEND_CHAR(ptr, ',');
-
-		n = sprintf(ptr, "(%d,", svector->indices[i]);
-		ptr += n;
-
-#if PG_VERSION_NUM >= 120000
-		n = float_to_shortest_decimal_bufn(values[i], ptr);
-#else
-		n = sprintf(ptr, "%.*g", ndig, values[i]);
-#endif
-		ptr += n;
-
-		APPEND_CHAR(ptr, ')');
-	}
-
-	n = sprintf(ptr, "|%d|", svector->dim);
-	ptr += n;
-
-	APPEND_CHAR(ptr, '\0');
-
-	PG_FREE_IF_COPY(svector, 0);
-	PG_RETURN_CSTRING(buf);
-}
-
-/*
- * Convert type modifier
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(svector_typmod_in);
-Datum
-svector_typmod_in(PG_FUNCTION_ARGS)
-{
-	ArrayType  *ta = PG_GETARG_ARRAYTYPE_P(0);
-	int32	   *tl;
-	int			n;
-
-	tl = ArrayGetIntegerTypmods(ta, &n);
-
-	if (n != 1)
-		ereport(ERROR,
-				(errcode(ERRCODE_INVALID_PARAMETER_VALUE),
-				 errmsg("invalid type modifier")));
-
-	if (*tl < 1)
-		ereport(ERROR,
-				(errcode(ERRCODE_INVALID_PARAMETER_VALUE),
-				 errmsg("dimensions for type svector must be at least 1")));
-
-	if (*tl > SVECTOR_MAX_DIM)
-		ereport(ERROR,
-				(errcode(ERRCODE_INVALID_PARAMETER_VALUE),
-				 errmsg("dimensions for type svector cannot exceed %d", SVECTOR_MAX_DIM)));
-
-	PG_RETURN_INT32(*tl);
-}
-
-/*
- * Convert external binary representation to internal representation
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(svector_recv);
-Datum
-svector_recv(PG_FUNCTION_ARGS)
-{
-	StringInfo	buf = (StringInfo) PG_GETARG_POINTER(0);
-	int32		typmod = PG_GETARG_INT32(2);
-	SVector    *result;
-	int32		dim;
-	int32		nnz;
-	int32		unused;
-	float	   *values;
-
-	dim = pq_getmsgint(buf, sizeof(int32));
-	nnz = pq_getmsgint(buf, sizeof(int32));
-	unused = pq_getmsgint(buf, sizeof(int32));
-
-	CheckDim(dim);
-	CheckNnz(nnz, dim);
-	CheckExpectedDim(typmod, dim);
-
-	if (unused != 0)
-		ereport(ERROR,
-				(errcode(ERRCODE_DATA_EXCEPTION),
-				 errmsg("expected unused to be 0, not %d", unused)));
-
-	result = InitSVector(dim, nnz);
-	values = SVECTOR_VALUES(result);
-
-	for (int i = 0; i < nnz; i++)
-	{
-		result->indices[i] = pq_getmsgint(buf, sizeof(int32));
-		CheckIndex(result->indices, i, dim);
-	}
-
-	for (int i = 0; i < nnz; i++)
-	{
-		values[i] = pq_getmsgfloat4(buf);
-		CheckElement(values[i]);
-	}
-
-	PG_RETURN_POINTER(result);
-}
-
-/*
- * Convert internal representation to the external binary representation
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(svector_send);
-Datum
-svector_send(PG_FUNCTION_ARGS)
-{
-	SVector    *svec = PG_GETARG_SVECTOR_P(0);
-	float	   *values = SVECTOR_VALUES(svec);
-	StringInfoData buf;
-
-	pq_begintypsend(&buf);
-	pq_sendint(&buf, svec->dim, sizeof(int32));
-	pq_sendint(&buf, svec->nnz, sizeof(int32));
-	pq_sendint(&buf, svec->unused, sizeof(int32));
-	for (int i = 0; i < svec->nnz; i++)
-		pq_sendint(&buf, svec->indices[i], sizeof(int32));
-	for (int i = 0; i < svec->nnz; i++)
-		pq_sendfloat4(&buf, values[i]);
-
-	PG_RETURN_BYTEA_P(pq_endtypsend(&buf));
-}
-
-/*
- * Convert sparse vector to sparse vector
- * This is needed to check the type modifier
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(svector);
-Datum
-svector(PG_FUNCTION_ARGS)
-{
-	SVector    *svec = PG_GETARG_SVECTOR_P(0);
-	int32		typmod = PG_GETARG_INT32(1);
-
-	CheckExpectedDim(typmod, svec->dim);
-
-	PG_RETURN_POINTER(svec);
-}
-
-/*
- * Convert dense vector to sparse vector
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_to_svector);
-Datum
-vector_to_svector(PG_FUNCTION_ARGS)
-{
-	Vector	   *vec = PG_GETARG_VECTOR_P(0);
-	int32		typmod = PG_GETARG_INT32(1);
-	SVector    *result;
-	int			dim = vec->dim;
-	int			nnz = 0;
-	float	   *values;
-	int			j = 0;
-
-	CheckDim(dim);
-	CheckExpectedDim(typmod, dim);
-
-	for (int i = 0; i < dim; i++)
-	{
-		if (vec->x[i] != 0)
-			nnz++;
-	}
-
-	result = InitSVector(dim, nnz);
-	values = SVECTOR_VALUES(result);
-	for (int i = 0; i < dim; i++)
-	{
-		if (vec->x[i] != 0)
-		{
-			/* Safety check */
-			if (j == nnz)
-				elog(ERROR, "safety check failed");
-
-			result->indices[j] = i;
-			values[j] = vec->x[i];
-			j++;
-		}
-	}
-
-	PG_RETURN_POINTER(result);
-}
-
-/*
- *  Get the L2 squared distance between sparse vectors
- */
-static double
-l2_distance_squared_internal(SVector * a, SVector * b)
-{
-	float	   *ax = SVECTOR_VALUES(a);
-	float	   *bx = SVECTOR_VALUES(b);
-	double		distance = 0.0;
-	int			bpos = 0;
-
-	for (int i = 0; i < a->nnz; i++)
-	{
-		int			ai = a->indices[i];
-		int			bi = -1;
-
-		for (int j = bpos; j < b->nnz; j++)
-		{
-			bi = b->indices[j];
-
-			if (ai == bi)
-			{
-				double		diff = ax[i] - bx[j];
-
-				distance += diff * diff;
-			}
-			else if (ai > bi)
-				distance += bx[j] * bx[j];
-
-			/* Update start for next iteration */
-			if (ai >= bi)
-				bpos = j + 1;
-
-			/* Found or passed it */
-			if (bi >= ai)
-				break;
-		}
-
-		if (ai != bi)
-			distance += ax[i] * ax[i];
-	}
-
-	for (int j = bpos; j < b->nnz; j++)
-		distance += bx[j] * bx[j];
-
-	return distance;
-}
-
-/*
- * Get the L2 distance between sparse vectors
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(svector_l2_distance);
-Datum
-svector_l2_distance(PG_FUNCTION_ARGS)
-{
-	SVector    *a = PG_GETARG_SVECTOR_P(0);
-	SVector    *b = PG_GETARG_SVECTOR_P(1);
-
-	CheckDims(a, b);
-
-	PG_RETURN_FLOAT8(sqrt(l2_distance_squared_internal(a, b)));
-}
-
-/*
- * Get the L2 squared distance between sparse vectors
- * This saves a sqrt calculation
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(svector_l2_squared_distance);
-Datum
-svector_l2_squared_distance(PG_FUNCTION_ARGS)
-{
-	SVector    *a = PG_GETARG_SVECTOR_P(0);
-	SVector    *b = PG_GETARG_SVECTOR_P(1);
-
-	CheckDims(a, b);
-
-	PG_RETURN_FLOAT8(l2_distance_squared_internal(a, b));
-}
-
-/*
- * Get the inner product of two sparse vectors
- */
-static double
-inner_product_internal(SVector * a, SVector * b)
-{
-	float	   *ax = SVECTOR_VALUES(a);
-	float	   *bx = SVECTOR_VALUES(b);
-	double		distance = 0.0;
-	int			bpos = 0;
-
-	for (int i = 0; i < a->nnz; i++)
-	{
-		int			ai = a->indices[i];
-
-		for (int j = bpos; j < b->nnz; j++)
-		{
-			int			bi = b->indices[j];
-
-			/* Only update when the same index */
-			if (ai == bi)
-				distance += ax[i] * bx[j];
-
-			/* Update start for next iteration */
-			if (ai >= bi)
-				bpos = j + 1;
-
-			/* Found or passed it */
-			if (bi >= ai)
-				break;
-		}
-	}
-
-	return distance;
-}
-
-/*
- * Get the inner product of two sparse vectors
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(svector_inner_product);
-Datum
-svector_inner_product(PG_FUNCTION_ARGS)
-{
-	SVector    *a = PG_GETARG_SVECTOR_P(0);
-	SVector    *b = PG_GETARG_SVECTOR_P(1);
-
-	CheckDims(a, b);
-
-	PG_RETURN_FLOAT8(inner_product_internal(a, b));
-}
-
-/*
- * Get the negative inner product of two sparse vectors
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(svector_negative_inner_product);
-Datum
-svector_negative_inner_product(PG_FUNCTION_ARGS)
-{
-	SVector    *a = PG_GETARG_SVECTOR_P(0);
-	SVector    *b = PG_GETARG_SVECTOR_P(1);
-
-	CheckDims(a, b);
-
-	PG_RETURN_FLOAT8(-inner_product_internal(a, b));
-}
-
-/*
- * Get the cosine distance between two sparse vectors
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(svector_cosine_distance);
-Datum
-svector_cosine_distance(PG_FUNCTION_ARGS)
-{
-	SVector    *a = PG_GETARG_SVECTOR_P(0);
-	SVector    *b = PG_GETARG_SVECTOR_P(1);
-	float	   *ax = SVECTOR_VALUES(a);
-	float	   *bx = SVECTOR_VALUES(b);
-	float		norma = 0.0;
-	float		normb = 0.0;
-	double		similarity;
-
-	CheckDims(a, b);
-
-	similarity = inner_product_internal(a, b);
-
-	/* Auto-vectorized */
-	for (int i = 0; i < a->nnz; i++)
-		norma += ax[i] * ax[i];
-
-	/* Auto-vectorized */
-	for (int i = 0; i < b->nnz; i++)
-		normb += bx[i] * bx[i];
-
-	/* Use sqrt(a * b) over sqrt(a) * sqrt(b) */
-	similarity /= sqrt((double) norma * (double) normb);
-
-#ifdef _MSC_VER
-	/* /fp:fast may not propagate NaN */
-	if (isnan(similarity))
-		PG_RETURN_FLOAT8(NAN);
-#endif
-
-	/* Keep in range */
-	if (similarity > 1)
-		similarity = 1.0;
-	else if (similarity < -1)
-		similarity = -1.0;
-
-	PG_RETURN_FLOAT8(1.0 - similarity);
-}
-
-/*
- * Get the weighted Jaccard distance between two sparse vectors
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(svector_jaccard_distance);
-Datum
-svector_jaccard_distance(PG_FUNCTION_ARGS)
-{
-	SVector    *a = PG_GETARG_SVECTOR_P(0);
-	SVector    *b = PG_GETARG_SVECTOR_P(1);
-	float	   *ax = SVECTOR_VALUES(a);
-	float	   *bx = SVECTOR_VALUES(b);
-	double		num = 0.0;
-	double		denom = 0.0;
-	int			bpos = 0;
-
-	CheckDims(a, b);
-
-	/*
-	 * Weighted Jaccard distance is not defined for vectors with negative
-	 * values. Could check and return NaN if minimal impact on performance.
-	 */
-
-	for (int i = 0; i < a->nnz; i++)
-	{
-		int			ai = a->indices[i];
-		int			bi = -1;
-
-		for (int j = bpos; j < b->nnz; j++)
-		{
-			bi = b->indices[j];
-
-			if (ai == bi)
-			{
-				num += ax[i] < bx[j] ? ax[i] : bx[j];
-				denom += ax[i] > bx[j] ? ax[i] : bx[j];
-			}
-			else if (ai > bi)
-				denom += bx[j];
-
-			/* Update start for next iteration */
-			if (ai >= bi)
-				bpos = j + 1;
-
-			/* Found or passed it */
-			if (bi >= ai)
-				break;
-		}
-
-		if (ai != bi)
-			denom += ax[i];
-	}
-
-	for (int j = bpos; j < b->nnz; j++)
-		denom += bx[j];
-
-	if (denom > 0)
-		PG_RETURN_FLOAT8(1.0 - (num / denom));
-	else
-		PG_RETURN_FLOAT8(NAN);
-}

src/svector.h

@@ -1,23 +0,0 @@
-#ifndef SVECTOR_H
-#define SVECTOR_H
-
-#define SVECTOR_MAX_DIM 100000
-
-#define SVECTOR_SIZE(_nnz)		(offsetof(SVector, indices) + (_nnz) * sizeof(int32) + (_nnz * sizeof(float)))
-#define SVECTOR_VALUES(x)		((float *) (((char *) (x)) + offsetof(SVector, indices) + (x)->nnz * sizeof(int32)))
-#define DatumGetSVector(x)		((SVector *) PG_DETOAST_DATUM(x))
-#define PG_GETARG_SVECTOR_P(x)	DatumGetSVector(PG_GETARG_DATUM(x))
-#define PG_RETURN_SVECTOR_P(x)	PG_RETURN_POINTER(x)
-
-typedef struct SVector
-{
-	int32		vl_len_;		/* varlena header (do not touch directly!) */
-	int32		dim;			/* number of dimensions */
-	int32		nnz;
-	int32		unused;
-	int32		indices[FLEXIBLE_ARRAY_MEMBER];
-}			SVector;
-
-SVector    *InitSVector(int dim, int nnz);
-
-#endif

src/vector.c

@@ -9,7 +9,6 @@
 #include "lib/stringinfo.h"
 #include "libpq/pqformat.h"
 #include "port.h"				/* for strtof() */
-#include "svector.h"
 #include "utils/array.h"
 #include "utils/builtins.h"
 #include "utils/lsyscache.h"
@@ -1152,26 +1151,3 @@ vector_avg(PG_FUNCTION_ARGS)
 
 	PG_RETURN_POINTER(result);
 }
-
-/*
- * Convert sparse vector to dense vector
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(svector_to_vector);
-Datum
-svector_to_vector(PG_FUNCTION_ARGS)
-{
-	SVector    *svec = PG_GETARG_SVECTOR_P(0);
-	int32		typmod = PG_GETARG_INT32(1);
-	Vector	   *result;
-	int			dim = svec->dim;
-	float	   *values = SVECTOR_VALUES(svec);
-
-	CheckDim(dim);
-	CheckExpectedDim(typmod, dim);
-
-	result = InitVector(dim);
-	for (int i = 0; i < svec->nnz; i++)
-		result->x[svec->indices[i]] = values[i];
-
-	PG_RETURN_POINTER(result);
-}

test/expected/functions.out

@@ -54,85 +54,85 @@ SELECT vector_norm('[3e37,4e37]')::real;
        5e+37
 (1 row)
 
-SELECT l2_distance('[0,0]'::vector, '[3,4]');
+SELECT l2_distance('[0,0]', '[3,4]');
  l2_distance 
 -------------
            5
 (1 row)
 
-SELECT l2_distance('[0,0]'::vector, '[0,1]');
+SELECT l2_distance('[0,0]', '[0,1]');
  l2_distance 
 -------------
            1
 (1 row)
 
-SELECT l2_distance('[1,2]'::vector, '[3]');
+SELECT l2_distance('[1,2]', '[3]');
 ERROR:  different vector dimensions 2 and 1
-SELECT l2_distance('[3e38]'::vector, '[-3e38]');
+SELECT l2_distance('[3e38]', '[-3e38]');
  l2_distance 
 -------------
     Infinity
 (1 row)
 
-SELECT inner_product('[1,2]'::vector, '[3,4]');
+SELECT inner_product('[1,2]', '[3,4]');
  inner_product 
 ---------------
             11
 (1 row)
 
-SELECT inner_product('[1,2]'::vector, '[3]');
+SELECT inner_product('[1,2]', '[3]');
 ERROR:  different vector dimensions 2 and 1
-SELECT inner_product('[3e38]'::vector, '[3e38]');
+SELECT inner_product('[3e38]', '[3e38]');
  inner_product 
 ---------------
       Infinity
 (1 row)
 
-SELECT cosine_distance('[1,2]'::vector, '[2,4]');
+SELECT cosine_distance('[1,2]', '[2,4]');
  cosine_distance 
 -----------------
                0
 (1 row)
 
-SELECT cosine_distance('[1,2]'::vector, '[0,0]');
+SELECT cosine_distance('[1,2]', '[0,0]');
  cosine_distance 
 -----------------
              NaN
 (1 row)
 
-SELECT cosine_distance('[1,1]'::vector, '[1,1]');
+SELECT cosine_distance('[1,1]', '[1,1]');
  cosine_distance 
 -----------------
                0
 (1 row)
 
-SELECT cosine_distance('[1,0]'::vector, '[0,2]');
+SELECT cosine_distance('[1,0]', '[0,2]');
  cosine_distance 
 -----------------
                1
 (1 row)
 
-SELECT cosine_distance('[1,1]'::vector, '[-1,-1]');
+SELECT cosine_distance('[1,1]', '[-1,-1]');
  cosine_distance 
 -----------------
                2
 (1 row)
 
-SELECT cosine_distance('[1,2]'::vector, '[3]');
+SELECT cosine_distance('[1,2]', '[3]');
 ERROR:  different vector dimensions 2 and 1
-SELECT cosine_distance('[1,1]'::vector, '[1.1,1.1]');
+SELECT cosine_distance('[1,1]', '[1.1,1.1]');
  cosine_distance 
 -----------------
                0
 (1 row)
 
-SELECT cosine_distance('[1,1]'::vector, '[-1.1,-1.1]');
+SELECT cosine_distance('[1,1]', '[-1.1,-1.1]');
  cosine_distance 
 -----------------
                2
 (1 row)
 
-SELECT cosine_distance('[3e38]'::vector, '[3e38]');
+SELECT cosine_distance('[3e38]', '[3e38]');
  cosine_distance 
 -----------------
              NaN

test/expected/hnsw_cosine.out

@@ -9,18 +9,19 @@ SELECT * FROM t ORDER BY val <=> '[3,3,3]';
  [1,1,1]
  [1,2,3]
  [1,2,4]
-(3 rows)
+ [0,0,0]
+(4 rows)
 
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
  count 
 -------
-     3
+     4
 (1 row)
 
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector)) t2;
  count 
 -------
-     3
+     4
 (1 row)
 
 DROP TABLE t;

test/expected/ivfflat_cosine.out

@@ -9,18 +9,19 @@ SELECT * FROM t ORDER BY val <=> '[3,3,3]';
  [1,1,1]
  [1,2,3]
  [1,2,4]
-(3 rows)
+ [0,0,0]
+(4 rows)
 
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
  count 
 -------
-     3
+     4
 (1 row)
 
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector)) t2;
  count 
 -------
-     3
+     4
 (1 row)
 
 DROP TABLE t;

test/expected/svector.out

@@ -1,140 +0,0 @@
-SELECT '(0,1.5),(2,3.5)|5|'::svector;
-      svector       
---------------------
- (0,1.5),(2,3.5)|5|
-(1 row)
-
-SELECT '(0,1.5),(2,3.5)|5|'::svector::vector;
-     vector      
------------------
- [1.5,0,3.5,0,0]
-(1 row)
-
-SELECT '(0,1.5),(2,3.5)|5|'::svector::vector(5);
-     vector      
------------------
- [1.5,0,3.5,0,0]
-(1 row)
-
-SELECT '(0,1.5),(2,3.5)|5|'::svector::vector(4);
-ERROR:  expected 4 dimensions, not 5
-SELECT '[0,1.5,0,3.5,0]'::vector::svector;
-      svector       
---------------------
- (1,1.5),(3,3.5)|5|
-(1 row)
-
-SELECT '(0,0),(1,1),(2,0)|3|'::svector;
-       svector        
-----------------------
- (0,0),(1,1),(2,0)|3|
-(1 row)
-
-SELECT '|5|'::svector;
- svector 
----------
- |5|
-(1 row)
-
-SELECT '|-1|'::svector;
-ERROR:  svector must have at least 1 dimension
-LINE 1: SELECT '|-1|'::svector;
-               ^
-SELECT '|100001|'::svector;
-ERROR:  svector cannot have more than 100000 dimensions
-LINE 1: SELECT '|100001|'::svector;
-               ^
-SELECT '|16001|'::svector::vector;
-ERROR:  vector cannot have more than 16000 dimensions
-SELECT '(-1,1)|1|'::svector;
-ERROR:  index must not be negative
-LINE 1: SELECT '(-1,1)|1|'::svector;
-               ^
-SELECT '(1,1)|1|'::svector;
-ERROR:  index must be less than dimensions
-LINE 1: SELECT '(1,1)|1|'::svector;
-               ^
-SELECT '|1|'::svector(2);
-ERROR:  expected 2 dimensions, not 1
-SELECT l2_distance('|2|'::svector, '(0,3),(1,4)|2|');
- l2_distance 
--------------
-           5
-(1 row)
-
-SELECT l2_distance('|2|'::svector, '(1,1)|2|');
- l2_distance 
--------------
-           1
-(1 row)
-
-SELECT '|2|'::svector <-> '(0,3),(1,4)|2|';
- ?column? 
-----------
-        5
-(1 row)
-
-SELECT inner_product('(0,1),(1,2)|2|'::svector, '(0,2),(1,4)|2|');
- inner_product 
----------------
-            10
-(1 row)
-
-SELECT svector_negative_inner_product('(0,1),(1,2)|2|', '(0,2),(1,4)|2|');
- svector_negative_inner_product 
---------------------------------
-                            -10
-(1 row)
-
-SELECT cosine_distance('(0,1),(1,2)|2|'::svector, '(0,2),(1,4)|2|');
- cosine_distance 
------------------
-               0
-(1 row)
-
-SELECT cosine_distance('(0,1),(1,2)|2|'::svector, '|2|');
- cosine_distance 
------------------
-             NaN
-(1 row)
-
-SELECT cosine_distance('(0,1),(1,1)|2|'::svector, '(0,-1),(1,-1)|2|');
- cosine_distance 
------------------
-               2
-(1 row)
-
-SELECT cosine_distance('(0,1)|2|'::svector, '(1,2)|2|');
- cosine_distance 
------------------
-               1
-(1 row)
-
-SELECT cosine_distance('|1|'::svector, '|1|');
- cosine_distance 
------------------
-             NaN
-(1 row)
-
-SELECT cosine_distance('(0,1)|2|'::svector, '(0,1)|3|');
-ERROR:  different svector dimensions 2 and 3
-SELECT jaccard_distance('(0,1)|2|', '(0,1)|2|');
- jaccard_distance 
-------------------
-                0
-(1 row)
-
-SELECT jaccard_distance('(0,1)|2|', '(1,1)|2|');
- jaccard_distance 
-------------------
-                1
-(1 row)
-
-SELECT jaccard_distance('|1|', '|1|');
- jaccard_distance 
-------------------
-              NaN
-(1 row)
-
-SELECT jaccard_distance('(0,1)|2|', '(0,1)|3|');
-ERROR:  different svector dimensions 2 and 3

test/sql/functions.sql

@@ -13,24 +13,24 @@ SELECT vector_norm('[3,4]');
 SELECT vector_norm('[0,1]');
 SELECT vector_norm('[3e37,4e37]')::real;
 
-SELECT l2_distance('[0,0]'::vector, '[3,4]');
+SELECT l2_distance('[0,0]', '[3,4]');
-SELECT l2_distance('[0,0]'::vector, '[0,1]');
+SELECT l2_distance('[0,0]', '[0,1]');
-SELECT l2_distance('[1,2]'::vector, '[3]');
+SELECT l2_distance('[1,2]', '[3]');
-SELECT l2_distance('[3e38]'::vector, '[-3e38]');
+SELECT l2_distance('[3e38]', '[-3e38]');
 
-SELECT inner_product('[1,2]'::vector, '[3,4]');
+SELECT inner_product('[1,2]', '[3,4]');
-SELECT inner_product('[1,2]'::vector, '[3]');
+SELECT inner_product('[1,2]', '[3]');
-SELECT inner_product('[3e38]'::vector, '[3e38]');
+SELECT inner_product('[3e38]', '[3e38]');
 
-SELECT cosine_distance('[1,2]'::vector, '[2,4]');
+SELECT cosine_distance('[1,2]', '[2,4]');
-SELECT cosine_distance('[1,2]'::vector, '[0,0]');
+SELECT cosine_distance('[1,2]', '[0,0]');
-SELECT cosine_distance('[1,1]'::vector, '[1,1]');
+SELECT cosine_distance('[1,1]', '[1,1]');
-SELECT cosine_distance('[1,0]'::vector, '[0,2]');
+SELECT cosine_distance('[1,0]', '[0,2]');
-SELECT cosine_distance('[1,1]'::vector, '[-1,-1]');
+SELECT cosine_distance('[1,1]', '[-1,-1]');
-SELECT cosine_distance('[1,2]'::vector, '[3]');
+SELECT cosine_distance('[1,2]', '[3]');
-SELECT cosine_distance('[1,1]'::vector, '[1.1,1.1]');
+SELECT cosine_distance('[1,1]', '[1.1,1.1]');
-SELECT cosine_distance('[1,1]'::vector, '[-1.1,-1.1]');
+SELECT cosine_distance('[1,1]', '[-1.1,-1.1]');
-SELECT cosine_distance('[3e38]'::vector, '[3e38]');
+SELECT cosine_distance('[3e38]', '[3e38]');
 
 SELECT l1_distance('[0,0]', '[3,4]');
 SELECT l1_distance('[0,0]', '[0,1]');

test/sql/svector.sql

@@ -1,36 +0,0 @@
-SELECT '(0,1.5),(2,3.5)|5|'::svector;
-SELECT '(0,1.5),(2,3.5)|5|'::svector::vector;
-SELECT '(0,1.5),(2,3.5)|5|'::svector::vector(5);
-SELECT '(0,1.5),(2,3.5)|5|'::svector::vector(4);
-SELECT '[0,1.5,0,3.5,0]'::vector::svector;
-
-SELECT '(0,0),(1,1),(2,0)|3|'::svector;
-
-SELECT '|5|'::svector;
-SELECT '|-1|'::svector;
-SELECT '|100001|'::svector;
-SELECT '|16001|'::svector::vector;
-
-SELECT '(-1,1)|1|'::svector;
-SELECT '(1,1)|1|'::svector;
-
-SELECT '|1|'::svector(2);
-
-SELECT l2_distance('|2|'::svector, '(0,3),(1,4)|2|');
-SELECT l2_distance('|2|'::svector, '(1,1)|2|');
-SELECT '|2|'::svector <-> '(0,3),(1,4)|2|';
-
-SELECT inner_product('(0,1),(1,2)|2|'::svector, '(0,2),(1,4)|2|');
-SELECT svector_negative_inner_product('(0,1),(1,2)|2|', '(0,2),(1,4)|2|');
-
-SELECT cosine_distance('(0,1),(1,2)|2|'::svector, '(0,2),(1,4)|2|');
-SELECT cosine_distance('(0,1),(1,2)|2|'::svector, '|2|');
-SELECT cosine_distance('(0,1),(1,1)|2|'::svector, '(0,-1),(1,-1)|2|');
-SELECT cosine_distance('(0,1)|2|'::svector, '(1,2)|2|');
-SELECT cosine_distance('|1|'::svector, '|1|');
-SELECT cosine_distance('(0,1)|2|'::svector, '(0,1)|3|');
-
-SELECT jaccard_distance('(0,1)|2|', '(0,1)|2|');
-SELECT jaccard_distance('(0,1)|2|', '(1,1)|2|');
-SELECT jaccard_distance('|1|', '|1|');
-SELECT jaccard_distance('(0,1)|2|', '(0,1)|3|');

test/t/012_hnsw_build_recall.pl

@@ -83,11 +83,32 @@ for my $i (0 .. $#operators)
 		push(@expected, $res);
 	}
 
-	# Add index
+	# Build index serially
-	$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v $opclass);");
+	$node->safe_psql("postgres", qq(
+		SET max_parallel_maintenance_workers = 0;
+		CREATE INDEX idx 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;
+		SET hnsw.enable_parallel_build = on;
+		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();