Use List for samples

This takes the approach from commit a713e2acaa further. Once we have
remove a candidate from the "closer" set, we still don't need to
recalculate everything that follows. Any candidates that were in the
closer set before still only need to be compared with any new
candidates that we have added.

CHANGELOG.md

@@ -1,6 +1,7 @@
-## 0.5.1 (unreleased)
+## 0.5.1 (2023-10-10)
 
-- Improved performance of index scans for IVFFlat after updates and deletes
+- Improved performance of HNSW index builds
+- Added check for MVCC-compliant snapshot for index scans
 
 ## 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": "0.5.0",
+	"version": "0.5.1",
 	"maintainer": [
 		"Andrew Kane <andrew@ankane.org>"
 	],
@@ -20,7 +20,7 @@
 		"vector": {
 			"file": "sql/vector.sql",
 			"docfile": "README.md",
-			"version": "0.5.0",
+			"version": "0.5.1",
 			"abstract": "Open-source vector similarity search for Postgres"
 		}
 	},

Makefile

@@ -1,5 +1,5 @@
 EXTENSION = vector
-EXTVERSION = 0.5.0
+EXTVERSION = 0.5.1
 
 MODULE_big = vector
 DATA = $(wildcard sql/*--*.sql)

Makefile.win

@@ -1,5 +1,5 @@
 EXTENSION = vector
-EXTVERSION = 0.5.0
+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\vector.obj
 HEADERS = src\vector.h

README.md

@@ -18,7 +18,7 @@ Compile and install the extension (supports Postgres 11+)
 
 ```sh
 cd /tmp
-git clone --branch v0.5.0 https://github.com/pgvector/pgvector.git
+git clone --branch v0.5.1 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 v0.5.0 https://github.com/pgvector/pgvector.git
+git clone --branch v0.5.1 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 v0.5.0 https://github.com/pgvector/pgvector.git
+git clone --branch v0.5.1 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

@@ -0,0 +1,2 @@
+-- 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

src/hnsw.h

@@ -57,6 +57,8 @@
 /* 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(_dim)	MAXALIGN(offsetof(HnswElementTupleData, vec) + VECTOR_SIZE(_dim))
 #define HNSW_NEIGHBOR_TUPLE_SIZE(level, m)	MAXALIGN(offsetof(HnswNeighborTupleData, indextids) + ((level) + 2) * (m) * sizeof(ItemPointerData))
 
@@ -110,11 +112,13 @@ typedef struct HnswCandidate
 {
 	HnswElement element;
 	float		distance;
+	bool		closer;
 }			HnswCandidate;
 
 typedef struct HnswNeighborArray
 {
 	int			length;
+	bool		closerSet;
 	HnswCandidate *items;
 }			HnswNeighborArray;
 

src/hnswbuild.c

@@ -117,12 +117,12 @@ CreateElementPages(HnswBuildState * buildstate)
 	ListCell   *lc;
 
 	/* Calculate sizes */
-	maxSize = BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData));
+	maxSize = HNSW_MAX_SIZE;
 	etupSize = HNSW_ELEMENT_TUPLE_SIZE(dimensions);
 
 	/* Allocate once */
 	etup = palloc0(etupSize);
-	ntup = palloc0(maxSize);
+	ntup = palloc0(BLCKSZ);
 
 	/* Prepare first page */
 	buf = HnswNewBuffer(index, forkNum);

src/hnswinsert.c

@@ -135,7 +135,7 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 	etupSize = HNSW_ELEMENT_TUPLE_SIZE(dimensions);
 	ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(e->level, m);
 	combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
-	maxSize = BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData));
+	maxSize = HNSW_MAX_SIZE;
 	minCombinedSize = etupSize + HNSW_NEIGHBOR_TUPLE_SIZE(0, m) + sizeof(ItemIdData);
 
 	/* Prepare element tuple */

src/hnswscan.c

@@ -160,6 +160,11 @@ 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);
 
@@ -201,15 +206,6 @@ 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;
 	}

src/hnswutils.c

@@ -139,6 +139,7 @@ HnswInitNeighbors(HnswElement element, int m)
 		a = &element->neighbors[lc];
 		a->length = 0;
 		a->items = palloc(sizeof(HnswCandidate) * lm);
+		a->closerSet = false;
 	}
 }
 
@@ -692,6 +693,34 @@ 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
  */
@@ -748,33 +777,77 @@ 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, HnswCandidate * *pruned)
+SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswElement e2, HnswCandidate * newCandidate, HnswCandidate * *pruned, bool sortCandidates)
 {
 	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);
 
-		closer = CheckElementCloser(e, r, lc, procinfo, collation);
+		/* 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);
 
-		if (closer)
+				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)
 			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);
@@ -828,28 +901,6 @@ 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
  */
@@ -903,13 +954,12 @@ HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int
 		{
 			List	   *c = NIL;
 
-			/* Add and sort candidates */
+			/* Add 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, &pruned);
+			SelectNeighbors(c, m, lc, procinfo, collation, hc->element, &hc2, &pruned, true);
 
 			/* Should not happen */
 			if (pruned == NULL)
@@ -1008,7 +1058,12 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
 		else
 			lw = w;
 
-		neighbors = SelectNeighbors(lw, lm, lc, procinfo, collation, NULL);
+		/*
+		 * 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);
 
 		AddConnections(element, neighbors, lm, lc);
 

src/ivfbuild.c

@@ -11,6 +11,7 @@
 #include "miscadmin.h"
 #include "storage/bufmgr.h"
 #include "tcop/tcopprot.h"
+#include "utils/datum.h"
 #include "utils/memutils.h"
 
 #if PG_VERSION_NUM >= 140000
@@ -65,11 +66,18 @@
 static void
 AddSample(Datum *values, IvfflatBuildState * buildstate)
 {
-	VectorArray samples = buildstate->samples;
-	int			targsamples = samples->maxlen;
+	MemoryContext oldCtx;
+	Datum		value;
+	int			targsamples = buildstate->targsamples;
+
+	/* Use memory context since detoast can allocate */
+	oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
 
 	/* Detoast once for all calls */
-	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
+	value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
+
+	/* Restore memory context */
+	MemoryContextSwitchTo(oldCtx);
 
 	/*
 	 * Normalize with KMEANS_NORM_PROC since spherical distance function
@@ -81,18 +89,23 @@ AddSample(Datum *values, IvfflatBuildState * buildstate)
 			return;
 	}
 
-	if (samples->length < targsamples)
-	{
-		VectorArraySet(samples, samples->length, DatumGetVector(value));
-		samples->length++;
-	}
+	/* Copy datum */
+	value = datumCopy(value, false, -1);
+
+	/* Reset memory context */
+	MemoryContextReset(buildstate->tmpCtx);
+
+	if (list_length(buildstate->samples) < targsamples)
+		buildstate->samples = lappend(buildstate->samples, DatumGetVector(value));
 	else
 	{
 		if (buildstate->rowstoskip < 0)
-			buildstate->rowstoskip = reservoir_get_next_S(&buildstate->rstate, samples->length, targsamples);
+			buildstate->rowstoskip = reservoir_get_next_S(&buildstate->rstate, list_length(buildstate->samples), targsamples);
 
 		if (buildstate->rowstoskip <= 0)
 		{
+			ListCell   *lc;
+
 #if PG_VERSION_NUM >= 150000
 			int			k = (int) (targsamples * sampler_random_fract(&buildstate->rstate.randstate));
 #else
@@ -100,7 +113,8 @@ AddSample(Datum *values, IvfflatBuildState * buildstate)
 #endif
 
 			Assert(k >= 0 && k < targsamples);
-			VectorArraySet(samples, k, DatumGetVector(value));
+			lc = list_nth_cell(buildstate->samples, k);
+			lfirst(lc) = DatumGetVector(value);
 		}
 
 		buildstate->rowstoskip -= 1;
@@ -115,21 +129,13 @@ SampleCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 			   bool *isnull, bool tupleIsAlive, void *state)
 {
 	IvfflatBuildState *buildstate = (IvfflatBuildState *) state;
-	MemoryContext oldCtx;
 
 	/* Skip nulls */
 	if (isnull[0])
 		return;
 
-	/* Use memory context since detoast can allocate */
-	oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
-
 	/* Add sample */
-	AddSample(values, state);
-
-	/* Reset memory context */
-	MemoryContextSwitchTo(oldCtx);
-	MemoryContextReset(buildstate->tmpCtx);
+	AddSample(values, buildstate);
 }
 
 /*
@@ -138,7 +144,7 @@ SampleCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 static void
 SampleRows(IvfflatBuildState * buildstate)
 {
-	int			targsamples = buildstate->samples->maxlen;
+	int			targsamples = buildstate->targsamples;
 	BlockNumber totalblocks = RelationGetNumberOfBlocks(buildstate->heap);
 
 	buildstate->rowstoskip = -1;
@@ -449,12 +455,13 @@ ComputeCenters(IvfflatBuildState * buildstate)
 
 	/* Sample rows */
 	/* TODO Ensure within maintenance_work_mem */
-	buildstate->samples = VectorArrayInit(numSamples, buildstate->dimensions);
+	buildstate->samples = NIL;
+	buildstate->targsamples = numSamples;
 	if (buildstate->heap != NULL)
 	{
 		SampleRows(buildstate);
 
-		if (buildstate->samples->length < buildstate->lists)
+		if (list_length(buildstate->samples) < buildstate->lists)
 		{
 			ereport(NOTICE,
 					(errmsg("ivfflat index created with little data"),
@@ -467,7 +474,7 @@ ComputeCenters(IvfflatBuildState * buildstate)
 	IvfflatBench("k-means", IvfflatKmeans(buildstate->index, buildstate->samples, buildstate->centers));
 
 	/* Free samples before we allocate more memory */
-	VectorArrayFree(buildstate->samples);
+	list_free_deep(buildstate->samples);
 }
 
 /*

src/ivfflat.h

@@ -80,6 +80,10 @@
 #define RandomInt() random()
 #endif
 
+#if PG_VERSION_NUM < 130000
+#define list_sort(list, cmp) list_qsort(list, cmp)
+#endif
+
 /* Variables */
 extern int	ivfflat_probes;
 
@@ -178,7 +182,8 @@ typedef struct IvfflatBuildState
 	Oid			collation;
 
 	/* Variables */
-	VectorArray samples;
+	List	   *samples;
+	int			targsamples;
 	VectorArray centers;
 	ListInfo   *listInfo;
 	Vector	   *normvec;
@@ -246,8 +251,6 @@ typedef struct IvfflatScanOpaqueData
 	int			probes;
 	int			dimensions;
 	bool		first;
-	Buffer		buf;
-	ItemPointerData heaptid;
 
 	/* Sorting */
 	Tuplesortstate *sortstate;
@@ -276,7 +279,7 @@ typedef IvfflatScanOpaqueData * IvfflatScanOpaque;
 VectorArray VectorArrayInit(int maxlen, int dimensions);
 void		VectorArrayFree(VectorArray arr);
 void		PrintVectorArray(char *msg, VectorArray arr);
-void		IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers);
+void		IvfflatKmeans(Relation index, List *samples, VectorArray centers);
 FmgrInfo   *IvfflatOptionalProcInfo(Relation index, uint16 procnum);
 bool		IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
 int			IvfflatGetLists(Relation index);

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)));
+	Assert(itemsz <= BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(IvfflatPageOpaqueData)) - sizeof(ItemIdData));
 
 	/* Find a page to insert the item */
 	for (;;)

src/ivfkmeans.c

@@ -12,20 +12,20 @@
  * https://theory.stanford.edu/~sergei/papers/kMeansPP-soda.pdf
  */
 static void
-InitCenters(Relation index, VectorArray samples, VectorArray centers, float *lowerBound)
+InitCenters(Relation index, List *samples, VectorArray centers, float *lowerBound)
 {
 	FmgrInfo   *procinfo;
 	Oid			collation;
 	int64		j;
-	float	   *weight = palloc(samples->length * sizeof(float));
+	float	   *weight = palloc(list_length(samples) * sizeof(float));
 	int			numCenters = centers->maxlen;
-	int			numSamples = samples->length;
+	int			numSamples = list_length(samples);
 
 	procinfo = index_getprocinfo(index, 1, IVFFLAT_KMEANS_DISTANCE_PROC);
 	collation = index->rd_indcollation[0];
 
 	/* Choose an initial center uniformly at random */
-	VectorArraySet(centers, 0, VectorArrayGet(samples, RandomInt() % samples->length));
+	VectorArraySet(centers, 0, list_nth(samples, RandomInt() % list_length(samples)));
 	centers->length++;
 
 	for (j = 0; j < numSamples; j++)
@@ -42,7 +42,7 @@ InitCenters(Relation index, VectorArray samples, VectorArray centers, float *low
 
 		for (j = 0; j < numSamples; j++)
 		{
-			Vector	   *vec = VectorArrayGet(samples, j);
+			Vector	   *vec = list_nth(samples, j);
 			double		distance;
 
 			/* Only need to compute distance for new center */
@@ -74,7 +74,7 @@ InitCenters(Relation index, VectorArray samples, VectorArray centers, float *low
 				break;
 		}
 
-		VectorArraySet(centers, i + 1, VectorArrayGet(samples, j));
+		VectorArraySet(centers, i + 1, list_nth(samples, j));
 		centers->length++;
 	}
 
@@ -106,25 +106,41 @@ CompareVectors(const void *a, const void *b)
 	return vector_cmp_internal((Vector *) a, (Vector *) b);
 }
 
+/*
+ * Compare list vectors
+ */
+static int
+#if PG_VERSION_NUM >= 130000
+CompareListVectors(const ListCell *a, const ListCell *b)
+#else
+CompareListVectors(const void *a, const void *b)
+#endif
+{
+	Vector	   *va = lfirst((ListCell *) a);
+	Vector	   *vb = lfirst((ListCell *) b);
+
+	return CompareVectors(va, vb);
+}
+
 /*
  * Quick approach if we have little data
  */
 static void
-QuickCenters(Relation index, VectorArray samples, VectorArray centers)
+QuickCenters(Relation index, List *samples, VectorArray centers)
 {
 	int			dimensions = centers->dim;
 	Oid			collation = index->rd_indcollation[0];
 	FmgrInfo   *normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_KMEANS_NORM_PROC);
 
 	/* Copy existing vectors while avoiding duplicates */
-	if (samples->length > 0)
+	if (list_length(samples) > 0)
 	{
-		qsort(samples->items, samples->length, VECTOR_SIZE(samples->dim), CompareVectors);
-		for (int i = 0; i < samples->length; i++)
+		list_sort(samples, CompareListVectors);
+		for (int i = 0; i < list_length(samples); i++)
 		{
-			Vector	   *vec = VectorArrayGet(samples, i);
+			Vector	   *vec = list_nth(samples, i);
 
-			if (i == 0 || CompareVectors(vec, VectorArrayGet(samples, i - 1)) != 0)
+			if (i == 0 || CompareVectors(vec, list_nth(samples, i - 1)) != 0)
 			{
 				VectorArraySet(centers, centers->length, vec);
 				centers->length++;
@@ -160,7 +176,7 @@ QuickCenters(Relation index, VectorArray samples, VectorArray centers)
  * https://www.aaai.org/Papers/ICML/2003/ICML03-022.pdf
  */
 static void
-ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
+ElkanKmeans(Relation index, List *samples, VectorArray centers)
 {
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
@@ -171,7 +187,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 	int64		k;
 	int			dimensions = centers->dim;
 	int			numCenters = centers->maxlen;
-	int			numSamples = samples->length;
+	int			numSamples = list_length(samples);
 	VectorArray newCenters;
 	int		   *centerCounts;
 	int		   *closestCenters;
@@ -182,7 +198,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 	float	   *newcdist;
 
 	/* Calculate allocation sizes */
-	Size		samplesSize = VECTOR_ARRAY_SIZE(samples->maxlen, samples->dim);
+	Size		samplesSize = 0;
 	Size		centersSize = VECTOR_ARRAY_SIZE(centers->maxlen, centers->dim);
 	Size		newCentersSize = VECTOR_ARRAY_SIZE(numCenters, dimensions);
 	Size		centerCountsSize = sizeof(int) * numCenters;
@@ -326,7 +342,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 				if (upperBound[j] <= halfcdist[closestCenters[j] * numCenters + k])
 					continue;
 
-				vec = VectorArrayGet(samples, j);
+				vec = list_nth(samples, j);
 
 				/* Step 3a */
 				if (rj)
@@ -377,7 +393,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 		{
 			int			closestCenter;
 
-			vec = VectorArrayGet(samples, j);
+			vec = list_nth(samples, j);
 			closestCenter = closestCenters[j];
 
 			/* Increment sum and count of closest center */
@@ -514,9 +530,9 @@ CheckCenters(Relation index, VectorArray centers)
  * We use spherical k-means for inner product and cosine
  */
 void
-IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers)
+IvfflatKmeans(Relation index, List *samples, VectorArray centers)
 {
-	if (samples->length <= centers->maxlen)
+	if (list_length(samples) <= centers->maxlen)
 		QuickCenters(index, samples, centers);
 	else
 		ElkanKmeans(index, samples, centers);

src/ivfscan.c

@@ -143,10 +143,6 @@ 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);
 
@@ -161,8 +157,6 @@ 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);
@@ -187,55 +181,6 @@ 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
  */
@@ -261,9 +206,7 @@ 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;
 
@@ -274,13 +217,12 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 
 	/* Create tuple description for sorting */
 #if PG_VERSION_NUM >= 120000
-	so->tupdesc = CreateTemplateTupleDesc(3);
+	so->tupdesc = CreateTemplateTupleDesc(2);
 #else
-	so->tupdesc = CreateTemplateTupleDesc(3, false);
+	so->tupdesc = CreateTemplateTupleDesc(2, 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);
@@ -312,7 +254,6 @@ 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)
@@ -347,6 +288,11 @@ 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
@@ -370,17 +316,10 @@ 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;
@@ -388,21 +327,6 @@ 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;
 	}
@@ -418,10 +342,6 @@ 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);
 

vector.control

@@ -1,5 +1,4 @@
 comment = 'vector data type and ivfflat and hnsw access methods'
-default_version = '0.5.0'
+default_version = '0.5.1'
 module_pathname = '$libdir/vector'
 relocatable = true
-trusted = true