Added strict ordering [skip ci]

CHANGELOG.md

@@ -1,5 +1,6 @@
 ## 0.8.0 (unreleased)
 
+- Added support for iterative index scans
 - Added casts for arrays to `sparsevec`
 - Improved cost estimation
 - Reduced memory usage for HNSW index scans

README.md

@@ -445,6 +445,63 @@ Use [partitioning](https://www.postgresql.org/docs/current/ddl-partitioning.html
 CREATE TABLE items (embedding vector(3), category_id int) PARTITION BY LIST(category_id);
 ```
 
+## Streaming Queries [unreleased]
+
+*Added in 0.8.0*
+
+With approximate indexes, you can end up with less results than expected due to filtering conditions in the query.
+
+Starting with 0.8.0, you can enable streaming queries. If too few results from the initial index scan match the query filters, it will resume scanning until enough results are found. This can significantly improve recall (at the cost of speed).
+
+```tsql
+SET hnsw.streaming = on;
+-- or
+SET ivfflat.streaming = on;
+```
+
+### Streaming Options
+
+Since scanning a large portion of the index is expensive, there are options to control when the scan ends.
+
+#### HNSW
+
+Specify the max number of additional tuples visited
+
+```sql
+SET hnsw.ef_stream = 10000;
+```
+
+The scan will also end if reaches `work_mem`, at which point a notice is shown
+
+```text
+NOTICE:  hnsw index scan exceeded work_mem after 50000 tuples
+HINT:  Increase work_mem to scan more tuples.
+```
+
+Adjust this with:
+
+```sql
+SET work_mem = '8MB';
+```
+
+#### IVFFlat
+
+Specify the max number of probes
+
+```sql
+SET ivfflat.max_probes = 100;
+```
+
+### Streaming Order
+
+With streaming queries, it’s possible for rows to be slightly out of order by distance. For strict ordering, use:
+
+```sql
+WITH approx_order AS MATERIALIZED (
+    SELECT *, embedding <-> '[1,2,3]' AS distance FROM items WHERE ... ORDER BY distance LIMIT 5
+) SELECT * FROM approx_order ORDER BY distance;
+```
+
 ## Half-Precision Vectors
 
 *Added in 0.7.0*

src/hnsw.c

@@ -19,6 +19,8 @@
 #endif
 
 int			hnsw_ef_search;
+int			hnsw_ef_stream;
+bool		hnsw_streaming;
 int			hnsw_lock_tranche_id;
 static relopt_kind hnsw_relopt_kind;
 
@@ -69,6 +71,17 @@ HnswInit(void)
 							"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);
 
+	/* TODO Figure out name */
+	DefineCustomBoolVariable("hnsw.streaming", "Use streaming mode",
+							 NULL, &hnsw_streaming,
+							 HNSW_DEFAULT_STREAMING, PGC_USERSET, 0, NULL, NULL, NULL);
+
+	/* TODO Figure out name */
+	/* TODO Use same value as ivfflat.max_probes for "all" */
+	DefineCustomIntVariable("hnsw.ef_stream", "Sets the max number of additional candidates to visit for streaming search",
+							"-1 means all", &hnsw_ef_stream,
+							HNSW_DEFAULT_EF_STREAM, HNSW_MIN_EF_STREAM, HNSW_MAX_EF_STREAM, PGC_USERSET, 0, NULL, NULL, NULL);
+
 	MarkGUCPrefixReserved("hnsw");
 }
 
@@ -89,6 +102,33 @@ hnswbuildphasename(int64 phasenum)
 	}
 }
 
+/*
+ * Estimate ef needed for iterative scans
+ */
+static int
+EstimateEf(PlannerInfo *root, IndexPath *path)
+{
+	double		selectivity = 1;
+	ListCell   *lc;
+
+	/* Cannot estimate without limit */
+	/* limit_tuples includes offset */
+	if (root->limit_tuples < 0)
+		return 0;
+
+	/* Get the selectivity of non-index conditions */
+	foreach(lc, path->indexinfo->indrestrictinfo)
+	{
+		RestrictInfo *rinfo = lfirst(lc);
+
+		/* Skip DEFAULT_INEQ_SEL since it may be a distance filter */
+		if (rinfo->norm_selec >= 0 && rinfo->norm_selec <= 1 && rinfo->norm_selec != (Selectivity) DEFAULT_INEQ_SEL)
+			selectivity *= rinfo->norm_selec;
+	}
+
+	return root->limit_tuples / Max(selectivity, 0.00001);
+}
+
 /*
  * Estimate the cost of an index scan
  */
@@ -100,6 +140,7 @@ hnswcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 {
 	GenericCosts costs;
 	int			m;
+	int			ef;
 	int			entryLevel;
 	int			layer0TuplesMax;
 	double		layer0Selectivity;
@@ -124,6 +165,8 @@ hnswcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 	HnswGetMetaPageInfo(index, &m, NULL);
 	index_close(index, NoLock);
 
+	ef = hnsw_streaming ? Max(hnsw_ef_search, EstimateEf(root, path)) : hnsw_ef_search;
+
 	/*
 	 * HNSW cost estimation follows a formula that accounts for the total
 	 * number of tuples indexed combined with the parameters that most
@@ -152,9 +195,9 @@ hnswcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 	 * "scalingFactor" (currently hardcoded).
 	 */
 	entryLevel = (int) (log(path->indexinfo->tuples + 1) * HnswGetMl(m));
-	layer0TuplesMax = HnswGetLayerM(m, 0) * hnsw_ef_search;
+	layer0TuplesMax = HnswGetLayerM(m, 0) * ef;
 	layer0Selectivity = (scalingFactor * log(path->indexinfo->tuples + 1)) /
-		(log(m) * (1 + log(hnsw_ef_search)));
+		(log(m) * (1 + log(ef)));
 
 	costs.numIndexTuples = (entryLevel * m) +
 		(layer0TuplesMax * layer0Selectivity);

src/hnsw.h

@@ -12,6 +12,10 @@
 #include "utils/sampling.h"
 #include "vector.h"
 
+#ifdef HNSW_BENCH
+#include "portability/instr_time.h"
+#endif
+
 #define HNSW_MAX_DIM 2000
 #define HNSW_MAX_NNZ 1000
 
@@ -42,6 +46,10 @@
 #define HNSW_DEFAULT_EF_SEARCH	40
 #define HNSW_MIN_EF_SEARCH		1
 #define HNSW_MAX_EF_SEARCH		1000
+#define HNSW_DEFAULT_STREAMING	false
+#define HNSW_DEFAULT_EF_STREAM	-1
+#define HNSW_MIN_EF_STREAM		-1
+#define HNSW_MAX_EF_STREAM		INT_MAX
 
 /* Tuple types */
 #define HNSW_ELEMENT_TUPLE_TYPE  1
@@ -68,6 +76,21 @@
 #define HnswPageGetOpaque(page)	((HnswPageOpaque) PageGetSpecialPointer(page))
 #define HnswPageGetMeta(page)	((HnswMetaPageData *) PageGetContents(page))
 
+#ifdef HNSW_BENCH
+#define HnswBench(name, code) \
+	do { \
+		instr_time	start; \
+		instr_time	duration; \
+		INSTR_TIME_SET_CURRENT(start); \
+		(code); \
+		INSTR_TIME_SET_CURRENT(duration); \
+		INSTR_TIME_SUBTRACT(duration, start); \
+		elog(INFO, "%s: %.3f ms", name, INSTR_TIME_GET_MILLISEC(duration)); \
+	} while (0)
+#else
+#define HnswBench(name, code) (code)
+#endif
+
 #if PG_VERSION_NUM >= 150000
 #define RandomDouble() pg_prng_double(&pg_global_prng_state)
 #define SeedRandom(seed) pg_prng_seed(&pg_global_prng_state, seed)
@@ -106,6 +129,8 @@
 
 /* Variables */
 extern int	hnsw_ef_search;
+extern int	hnsw_ef_stream;
+extern bool hnsw_streaming;
 extern int	hnsw_lock_tranche_id;
 
 typedef struct HnswElementData HnswElementData;
@@ -129,6 +154,7 @@ struct HnswElementData
 	uint8		heaptidsLength;
 	uint8		level;
 	uint8		deleted;
+	uint8		version;
 	uint32		hash;
 	HnswNeighborsPtr neighbors;
 	BlockNumber blkno;
@@ -163,6 +189,9 @@ typedef struct HnswSearchCandidate
 	float		distance;
 }			HnswSearchCandidate;
 
+#define HnswGetSearchCandidate(membername, ptr) pairingheap_container(HnswSearchCandidate, membername, ptr)
+#define HnswGetSearchCandidateConst(membername, ptr) pairingheap_const_container(HnswSearchCandidate, membername, ptr)
+
 /* HNSW index options */
 typedef struct HnswOptions
 {
@@ -306,10 +335,10 @@ typedef struct HnswElementTupleData
 	uint8		type;
 	uint8		level;
 	uint8		deleted;
-	uint8		unused;
+	uint8		version;
 	ItemPointerData heaptids[HNSW_HEAPTIDS];
 	ItemPointerData neighbortid;
-	uint16		unused2;
+	uint16		unused;
 	Vector		data;
 }			HnswElementTupleData;
 
@@ -318,18 +347,31 @@ typedef HnswElementTupleData * HnswElementTuple;
 typedef struct HnswNeighborTupleData
 {
 	uint8		type;
-	uint8		unused;
+	uint8		version;
 	uint16		count;
 	ItemPointerData indextids[FLEXIBLE_ARRAY_MEMBER];
 }			HnswNeighborTupleData;
 
 typedef HnswNeighborTupleData * HnswNeighborTuple;
 
+typedef union
+{
+	struct pointerhash_hash *pointers;
+	struct offsethash_hash *offsets;
+	struct tidhash_hash *tids;
+}			visited_hash;
+
 typedef struct HnswScanOpaqueData
 {
 	const		HnswTypeInfo *typeInfo;
 	bool		first;
 	List	   *w;
+	visited_hash v;
+	pairingheap *discarded;
+	Datum		q;
+	int			m;
+	int64		tuples;
+	double		previousDistance;
 	MemoryContext tmpCtx;
 
 	/* Support functions */
@@ -375,7 +417,7 @@ bool		HnswCheckNorm(FmgrInfo *procinfo, Oid collation, Datum value);
 Buffer		HnswNewBuffer(Relation index, ForkNumber forkNum);
 void		HnswInitPage(Buffer buf, Page page);
 void		HnswInit(void);
-List	   *HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement);
+List	   *HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement, visited_hash * v, pairingheap **discarded, bool initVisited, int64 *tuples);
 HnswElement HnswGetEntryPoint(Relation index);
 void		HnswGetMetaPageInfo(Relation index, int *m, HnswElement * entryPoint);
 void	   *HnswAlloc(HnswAllocator * allocator, Size size);

src/hnswinsert.c

@@ -36,7 +36,7 @@ GetInsertPage(Relation index)
  * Check for a free offset
  */
 static bool
-HnswFreeOffset(Relation index, Buffer buf, Page page, HnswElement element, Size etupSize, Size ntupSize, Buffer *nbuf, Page *npage, OffsetNumber *freeOffno, OffsetNumber *freeNeighborOffno, BlockNumber *newInsertPage)
+HnswFreeOffset(Relation index, Buffer buf, Page page, HnswElement element, Size etupSize, Size ntupSize, Buffer *nbuf, Page *npage, OffsetNumber *freeOffno, OffsetNumber *freeNeighborOffno, BlockNumber *newInsertPage, uint8 *tupleVersion)
 {
 	OffsetNumber offno;
 	OffsetNumber maxoffno = PageGetMaxOffsetNumber(page);
@@ -98,6 +98,7 @@ HnswFreeOffset(Relation index, Buffer buf, Page page, HnswElement element, Size
 			{
 				*freeOffno = offno;
 				*freeNeighborOffno = neighborOffno;
+				*tupleVersion = etup->version;
 				return true;
 			}
 			else if (*nbuf != buf)
@@ -153,6 +154,7 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B
 	OffsetNumber freeOffno = InvalidOffsetNumber;
 	OffsetNumber freeNeighborOffno = InvalidOffsetNumber;
 	BlockNumber newInsertPage = InvalidBlockNumber;
+	uint8		tupleVersion;
 	char	   *base = NULL;
 
 	/* Calculate sizes */
@@ -202,7 +204,7 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B
 		}
 
 		/* Next, try space from a deleted element */
-		if (HnswFreeOffset(index, buf, page, e, etupSize, ntupSize, &nbuf, &npage, &freeOffno, &freeNeighborOffno, &newInsertPage))
+		if (HnswFreeOffset(index, buf, page, e, etupSize, ntupSize, &nbuf, &npage, &freeOffno, &freeNeighborOffno, &newInsertPage, &tupleVersion))
 		{
 			if (nbuf != buf)
 			{
@@ -212,6 +214,10 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B
 					npage = GenericXLogRegisterBuffer(state, nbuf, 0);
 			}
 
+			/* Set tuple version */
+			etup->version = tupleVersion;
+			ntup->version = tupleVersion;
+
 			break;
 		}
 

src/hnswscan.c

@@ -1,5 +1,7 @@
 #include "postgres.h"
 
+#include <float.h>
+
 #include "access/relscan.h"
 #include "hnsw.h"
 #include "pgstat.h"
@@ -26,6 +28,9 @@ GetScanItems(IndexScanDesc scan, Datum q)
 	/* Get m and entry point */
 	HnswGetMetaPageInfo(index, &m, &entryPoint);
 
+	so->q = q;
+	so->m = m;
+
 	if (entryPoint == NULL)
 		return NIL;
 
@@ -33,11 +38,44 @@ GetScanItems(IndexScanDesc scan, Datum q)
 
 	for (int lc = entryPoint->level; lc >= 1; lc--)
 	{
-		w = HnswSearchLayer(base, q, ep, 1, lc, index, procinfo, collation, m, false, NULL);
+		w = HnswSearchLayer(base, q, ep, 1, lc, index, procinfo, collation, m, false, NULL, NULL, NULL, true, NULL);
 		ep = w;
 	}
 
-	return HnswSearchLayer(base, q, ep, hnsw_ef_search, 0, index, procinfo, collation, m, false, NULL);
+	return HnswSearchLayer(base, q, ep, hnsw_ef_search, 0, index, procinfo, collation, m, false, NULL, &so->v, hnsw_streaming ? &so->discarded : NULL, true, &so->tuples);
+}
+
+/*
+ * Resume scan at ground level with discarded candidates
+ */
+static List *
+ResumeScanItems(IndexScanDesc scan)
+{
+	HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
+	Relation	index = scan->indexRelation;
+	FmgrInfo   *procinfo = so->procinfo;
+	Oid			collation = so->collation;
+	List	   *ep = NIL;
+	char	   *base = NULL;
+	int			batch_size = hnsw_ef_search;
+
+	if (pairingheap_is_empty(so->discarded))
+		return NIL;
+
+	/* Get next batch of candidates */
+	for (int i = 0; i < batch_size; i++)
+	{
+		HnswSearchCandidate *hc;
+
+		if (pairingheap_is_empty(so->discarded))
+			break;
+
+		hc = HnswGetSearchCandidate(w_node, pairingheap_remove_first(so->discarded));
+
+		ep = lappend(ep, hc);
+	}
+
+	return HnswSearchLayer(base, so->q, ep, batch_size, 0, index, procinfo, collation, so->m, false, NULL, &so->v, &so->discarded, false, &so->tuples);
 }
 
 /*
@@ -81,6 +119,8 @@ hnswbeginscan(Relation index, int nkeys, int norderbys)
 	so = (HnswScanOpaque) palloc(sizeof(HnswScanOpaqueData));
 	so->typeInfo = HnswGetTypeInfo(index);
 	so->first = true;
+	so->v.tids = NULL;
+	so->discarded = NULL;
 	so->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
 									   "Hnsw scan temporary context",
 									   ALLOCSET_DEFAULT_SIZES);
@@ -103,7 +143,15 @@ hnswrescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys, int no
 {
 	HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
 
+	if (so->v.tids != NULL)
+		tidhash_reset(so->v.tids);
+
+	if (so->discarded != NULL)
+		pairingheap_reset(so->discarded);
+
 	so->first = true;
+	so->tuples = 0;
+	so->previousDistance = -INFINITY;
 	MemoryContextReset(so->tmpCtx);
 
 	if (keys && scan->numberOfKeys > 0)
@@ -153,7 +201,7 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
 		 */
 		LockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
 
-		so->w = GetScanItems(scan, value);
+		HnswBench("scan iteration", so->w = GetScanItems(scan, value));
 
 		/* Release shared lock */
 		UnlockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
@@ -165,22 +213,97 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
 #endif
 	}
 
-	while (list_length(so->w) > 0)
+	for (;;)
 	{
 		char	   *base = NULL;
-		HnswSearchCandidate *hc = llast(so->w);
-		HnswElement element = HnswPtrAccess(base, hc->element);
+		HnswSearchCandidate *hc;
+		HnswElement element;
 		ItemPointer heaptid;
 
+		if (list_length(so->w) == 0)
+		{
+			if (!hnsw_streaming)
+				break;
+
+			/* Empty index */
+			if (so->discarded == NULL)
+				break;
+
+			/* Reached max number of additional tuples */
+			if (hnsw_ef_stream != -1 && so->tuples >= hnsw_ef_search + hnsw_ef_stream)
+			{
+				if (pairingheap_is_empty(so->discarded))
+					break;
+
+				/* Return remaining tuples */
+				so->w = lappend(so->w, HnswGetSearchCandidate(w_node, pairingheap_remove_first(so->discarded)));
+			}
+			/* Prevent scans from consuming too much memory */
+			else if (MemoryContextMemAllocated(so->tmpCtx, false) > (Size) work_mem * 1024L)
+			{
+				if (pairingheap_is_empty(so->discarded))
+				{
+					ereport(NOTICE,
+							(errmsg("hnsw index scan exceeded work_mem after " INT64_FORMAT " tuples", so->tuples),
+							 errhint("Increase work_mem to scan more tuples.")));
+
+					break;
+				}
+
+				/* Return remaining tuples */
+				so->w = lappend(so->w, HnswGetSearchCandidate(w_node, pairingheap_remove_first(so->discarded)));
+			}
+			else
+			{
+				/*
+				 * Locking ensures when neighbors are read, the elements they
+				 * reference will not be deleted (and replaced) during the
+				 * iteration.
+				 *
+				 * Elements loaded into memory on previous iterations may have
+				 * been deleted (and replaced), so when reading neighbors, the
+				 * element version must be checked.
+				 */
+				LockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
+
+				HnswBench("scan iteration", so->w = ResumeScanItems(scan));
+
+				UnlockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
+
+#if defined(HNSW_MEMORY)
+				elog(INFO, "memory: %zu KB", MemoryContextMemAllocated(so->tmpCtx, false) / 1024);
+#endif
+			}
+
+			if (list_length(so->w) == 0)
+				break;
+		}
+
+		hc = llast(so->w);
+		element = HnswPtrAccess(base, hc->element);
+
 		/* Move to next element if no valid heap TIDs */
 		if (element->heaptidsLength == 0)
 		{
 			so->w = list_delete_last(so->w);
+
+			/* Mark memory as free for next iteration */
+			if (hnsw_streaming)
+			{
+				pfree(element);
+				pfree(hc);
+			}
+
 			continue;
 		}
 
 		heaptid = &element->heaptids[--element->heaptidsLength];
 
+		if (hc->distance < so->previousDistance)
+			continue;
+
+		so->previousDistance = hc->distance;
+
 		MemoryContextSwitchTo(oldCtx);
 
 		scan->xs_heaptid = *heaptid;

src/hnswutils.c

@@ -100,13 +100,6 @@ hash_offset(Size offset)
 #define SH_DEFINE
 #include "lib/simplehash.h"
 
-typedef union
-{
-	pointerhash_hash *pointers;
-	offsethash_hash *offsets;
-	tidhash_hash *tids;
-}			visited_hash;
-
 typedef union
 {
 	HnswElement element;
@@ -253,6 +246,8 @@ HnswInitElement(char *base, ItemPointer heaptid, int m, double ml, int maxLevel,
 
 	element->level = level;
 	element->deleted = 0;
+	/* Start at one to make it easier to find issues */
+	element->version = 1;
 
 	HnswInitNeighbors(base, element, m, allocator);
 
@@ -405,6 +400,7 @@ HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element)
 	etup->type = HNSW_ELEMENT_TUPLE_TYPE;
 	etup->level = element->level;
 	etup->deleted = 0;
+	etup->version = element->version;
 	for (int i = 0; i < HNSW_HEAPTIDS; i++)
 	{
 		if (i < element->heaptidsLength)
@@ -447,6 +443,7 @@ HnswSetNeighborTuple(char *base, HnswNeighborTuple ntup, HnswElement e, int m)
 	}
 
 	ntup->count = idx;
+	ntup->version = e->version;
 }
 
 /*
@@ -520,6 +517,7 @@ HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHe
 {
 	element->level = etup->level;
 	element->deleted = etup->deleted;
+	element->version = etup->version;
 	element->neighborPage = ItemPointerGetBlockNumber(&etup->neighbortid);
 	element->neighborOffno = ItemPointerGetOffsetNumber(&etup->neighbortid);
 	element->heaptidsLength = 0;
@@ -621,9 +619,6 @@ HnswEntryCandidate(char *base, HnswElement entryPoint, Datum q, Relation index,
 	return hc;
 }
 
-#define HnswGetSearchCandidate(membername, ptr) pairingheap_container(HnswSearchCandidate, membername, ptr)
-#define HnswGetSearchCandidateConst(membername, ptr) pairingheap_const_container(HnswSearchCandidate, membername, ptr)
-
 /*
  * Compare candidate distances
  */
@@ -639,6 +634,21 @@ CompareNearestCandidates(const pairingheap_node *a, const pairingheap_node *b, v
 	return 0;
 }
 
+/*
+ * Compare discarded candidate distances
+ */
+static int
+CompareNearestDiscardedCandidates(const pairingheap_node *a, const pairingheap_node *b, void *arg)
+{
+	if (HnswGetSearchCandidateConst(w_node, a)->distance < HnswGetSearchCandidateConst(w_node, b)->distance)
+		return 1;
+
+	if (HnswGetSearchCandidateConst(w_node, a)->distance > HnswGetSearchCandidateConst(w_node, b)->distance)
+		return -1;
+
+	return 0;
+}
+
 /*
  * Compare candidate distances
  */
@@ -754,14 +764,19 @@ HnswLoadUnvisitedFromDisk(HnswElement element, HnswUnvisited * unvisited, int *u
 	int			start;
 	ItemPointerData indextids[HNSW_MAX_M * 2];
 
+	*unvisitedLength = 0;
+
 	buf = ReadBuffer(index, element->neighborPage);
 	LockBuffer(buf, BUFFER_LOCK_SHARE);
 	page = BufferGetPage(buf);
 
 	ntup = (HnswNeighborTuple) PageGetItem(page, PageGetItemId(page, element->neighborOffno));
 
-	/* Ensure expected neighbors */
-	if (ntup->count != (element->level + 2) * m)
+	/*
+	 * Ensure the neighbor tuple has not been deleted or replaced between
+	 * index scan iterations
+	 */
+	if (ntup->version != element->version || ntup->count != (element->level + 2) * m)
 	{
 		UnlockReleaseBuffer(buf);
 		return;
@@ -773,8 +788,6 @@ HnswLoadUnvisitedFromDisk(HnswElement element, HnswUnvisited * unvisited, int *u
 
 	UnlockReleaseBuffer(buf);
 
-	*unvisitedLength = 0;
-
 	for (int i = 0; i < lm; i++)
 	{
 		ItemPointer indextid = &indextids[i];
@@ -794,13 +807,13 @@ HnswLoadUnvisitedFromDisk(HnswElement element, HnswUnvisited * unvisited, int *u
  * Algorithm 2 from paper
  */
 List *
-HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement)
+HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement, visited_hash * v, pairingheap **discarded, bool initVisited, int64 *tuples)
 {
 	List	   *w = NIL;
 	pairingheap *C = pairingheap_allocate(CompareNearestCandidates, NULL);
 	pairingheap *W = pairingheap_allocate(CompareFurthestCandidates, NULL);
 	int			wlen = 0;
-	visited_hash v;
+	visited_hash vh;
 	ListCell   *lc2;
 	HnswNeighborArray *localNeighborhood = NULL;
 	Size		neighborhoodSize = 0;
@@ -808,7 +821,19 @@ HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, F
 	HnswUnvisited *unvisited = palloc(lm * sizeof(HnswUnvisited));
 	int			unvisitedLength;
 
-	InitVisited(base, &v, index, ef, m);
+	if (v == NULL)
+	{
+		v = &vh;
+		initVisited = true;
+	}
+
+	if (initVisited)
+	{
+		InitVisited(base, v, index, ef, m);
+
+		if (discarded != NULL)
+			*discarded = pairingheap_allocate(CompareNearestDiscardedCandidates, NULL);
+	}
 
 	/* Create local memory for neighborhood if needed */
 	if (index == NULL)
@@ -823,7 +848,13 @@ HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, F
 		HnswSearchCandidate *hc = (HnswSearchCandidate *) lfirst(lc2);
 		bool		found;
 
-		AddToVisited(base, &v, hc->element, index, &found);
+		if (initVisited)
+		{
+			AddToVisited(base, v, hc->element, index, &found);
+
+			if (tuples != NULL)
+				(*tuples)++;
+		}
 
 		pairingheap_add(C, &hc->c_node);
 		pairingheap_add(W, &hc->w_node);
@@ -849,9 +880,12 @@ HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, F
 		cElement = HnswPtrAccess(base, c->element);
 
 		if (index == NULL)
-			HnswLoadUnvisitedFromMemory(base, cElement, unvisited, &unvisitedLength, &v, lc, localNeighborhood, neighborhoodSize);
+			HnswLoadUnvisitedFromMemory(base, cElement, unvisited, &unvisitedLength, v, lc, localNeighborhood, neighborhoodSize);
 		else
-			HnswLoadUnvisitedFromDisk(cElement, unvisited, &unvisitedLength, &v, index, m, lm, lc);
+			HnswLoadUnvisitedFromDisk(cElement, unvisited, &unvisitedLength, v, index, m, lm, lc);
+
+		if (tuples != NULL)
+			(*tuples) += unvisitedLength;
 
 		for (int i = 0; i < unvisitedLength; i++)
 		{
@@ -875,16 +909,22 @@ HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, F
 
 				/* Avoid any allocations if not adding */
 				eElement = NULL;
-				HnswLoadElementImpl(blkno, offno, &eDistance, &q, index, procinfo, collation, inserting, alwaysAdd ? NULL : &f->distance, &eElement);
-
-				if (eElement == NULL)
-					continue;
+				HnswLoadElementImpl(blkno, offno, &eDistance, &q, index, procinfo, collation, inserting, alwaysAdd || discarded != NULL ? NULL : &f->distance, &eElement);
 			}
 
-			if (!(eDistance < f->distance || alwaysAdd))
-				continue;
+			if (eElement == NULL || !(eDistance < f->distance || alwaysAdd))
+			{
+				if (discarded != NULL)
+				{
+					/* Create a new candidate */
+					e = palloc(sizeof(HnswSearchCandidate));
+					HnswPtrStore(base, e->element, eElement);
+					e->distance = eDistance;
+					pairingheap_add(*discarded, &e->w_node);
+				}
 
-			Assert(!eElement->deleted);
+				continue;
+			}
 
 			/* Make robust to issues */
 			if (eElement->level < lc)
@@ -908,7 +948,12 @@ HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, F
 
 				/* No need to decrement wlen */
 				if (wlen > ef)
-					pairingheap_remove_first(W);
+				{
+					HnswSearchCandidate *d = HnswGetSearchCandidate(w_node, pairingheap_remove_first(W));
+
+					if (discarded != NULL)
+						pairingheap_add(*discarded, &d->w_node);
+				}
 			}
 		}
 	}
@@ -1281,7 +1326,7 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
 	/* 1st phase: greedy search to insert level */
 	for (int lc = entryLevel; lc >= level + 1; lc--)
 	{
-		w = HnswSearchLayer(base, q, ep, 1, lc, index, procinfo, collation, m, true, skipElement);
+		w = HnswSearchLayer(base, q, ep, 1, lc, index, procinfo, collation, m, true, skipElement, NULL, NULL, true, NULL);
 		ep = w;
 	}
 
@@ -1300,7 +1345,7 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
 		List	   *lw = NIL;
 		ListCell   *lc2;
 
-		w = HnswSearchLayer(base, q, ep, efConstruction, lc, index, procinfo, collation, m, true, skipElement);
+		w = HnswSearchLayer(base, q, ep, efConstruction, lc, index, procinfo, collation, m, true, skipElement, NULL, NULL, true, NULL);
 
 		/* Convert search candidates to candidates */
 		foreach(lc2, w)

src/hnswvacuum.c

@@ -527,6 +527,14 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 			for (int i = 0; i < ntup->count; i++)
 				ItemPointerSetInvalid(&ntup->indextids[i]);
 
+			/* Increment version */
+			/* This is used to avoid incorrect reads for iterative scans */
+			/* Reserve some bits for future use */
+			etup->version++;
+			if (etup->version > 15)
+				etup->version = 1;
+			ntup->version = etup->version;
+
 			/*
 			 * We modified the tuples in place, no need to call
 			 * PageIndexTupleOverwrite

test/t/041_hnsw_streaming.pl

@@ -0,0 +1,66 @@
+use strict;
+use warnings FATAL => 'all';
+use PostgreSQL::Test::Cluster;
+use PostgreSQL::Test::Utils;
+use Test::More;
+
+my $dim = 3;
+my $array_sql = join(",", ('random()') x $dim);
+
+# Initialize node
+my $node = PostgreSQL::Test::Cluster->new('node');
+$node->init;
+$node->start;
+
+# Create table
+$node->safe_psql("postgres", "CREATE EXTENSION vector;");
+$node->safe_psql("postgres", "CREATE TABLE tst (i int4 PRIMARY KEY, v vector($dim));");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT i, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
+);
+$node->safe_psql("postgres", qq(
+	SET maintenance_work_mem = '128MB';
+	SET max_parallel_maintenance_workers = 2;
+	CREATE INDEX ON tst USING hnsw (v vector_l2_ops)
+));
+
+my $count = $node->safe_psql("postgres", qq(
+	SET enable_seqscan = off;
+	SET hnsw.streaming = on;
+	SET work_mem = '8MB';
+	SELECT COUNT(*) FROM (SELECT v FROM tst WHERE i % 10000 = 0 ORDER BY v <-> (SELECT v FROM tst LIMIT 1) LIMIT 11) t;
+));
+is($count, 10);
+
+foreach ((30000, 50000, 70000))
+{
+	my $ef_stream = $_;
+	my $expected = $ef_stream / 10000;
+	my $sum = 0;
+
+	for my $i (1 .. 20)
+	{
+		$count = $node->safe_psql("postgres", qq(
+			SET enable_seqscan = off;
+			SET hnsw.streaming = on;
+			SET hnsw.ef_stream = $ef_stream;
+			SET work_mem = '8MB';
+			SELECT COUNT(*) FROM (SELECT v FROM tst WHERE i % 10000 = 0 ORDER BY v <-> (SELECT v FROM tst WHERE i = $i) LIMIT 11) t;
+		));
+		$sum += $count;
+	}
+
+	my $avg = $sum / 20;
+	cmp_ok($avg, '>', $expected - 2);
+	cmp_ok($avg, '<', $expected + 2);
+}
+
+my ($ret, $stdout, $stderr) = $node->psql("postgres", qq(
+	SET enable_seqscan = off;
+	SET hnsw.streaming = on;
+	SET work_mem = '2MB';
+	SELECT COUNT(*) FROM (SELECT v FROM tst WHERE i % 10000 = 0 ORDER BY v <-> (SELECT v FROM tst LIMIT 1) LIMIT 11) t;
+));
+like($stderr, qr/hnsw index scan exceeded work_mem after \d+ tuples/);
+
+done_testing();

test/t/042_hnsw_streaming_recall.pl

@@ -0,0 +1,131 @@
+use strict;
+use warnings FATAL => 'all';
+use PostgreSQL::Test::Cluster;
+use PostgreSQL::Test::Utils;
+use Test::More;
+
+my $node;
+my @queries = ();
+my @expected;
+my $limit = 20;
+my $dim = 3;
+my $array_sql = join(",", ('random()') x $dim);
+my @cs = (100, 1000);
+
+sub test_recall
+{
+	my ($c, $ef_search, $min, $operator) = @_;
+	my $correct = 0;
+	my $total = 0;
+
+	my $explain = $node->safe_psql("postgres", qq(
+		SET enable_seqscan = off;
+		SET hnsw.ef_search = $ef_search;
+		SET hnsw.streaming = on;
+		EXPLAIN ANALYZE SELECT i FROM tst WHERE i % $c = 0 ORDER BY v $operator '$queries[0]' LIMIT $limit;
+	));
+	like($explain, qr/Index Scan using idx on tst/);
+
+	for my $i (0 .. $#queries)
+	{
+		my $actual = $node->safe_psql("postgres", qq(
+			SET enable_seqscan = off;
+			SET hnsw.ef_search = $ef_search;
+			SET hnsw.streaming = on;
+			SELECT i FROM tst WHERE i % $c = 0 ORDER BY v $operator '$queries[$i]' LIMIT $limit;
+		));
+		my @actual_ids = split("\n", $actual);
+
+		my @expected_ids = split("\n", $expected[$i]);
+		my %expected_set = map { $_ => 1 } @expected_ids;
+
+		foreach (@actual_ids)
+		{
+			if (exists($expected_set{$_}))
+			{
+				$correct++;
+			}
+		}
+
+		$total += $limit;
+	}
+
+	cmp_ok($correct / $total, ">=", $min, $operator);
+}
+
+# Initialize node
+$node = PostgreSQL::Test::Cluster->new('node');
+$node->init;
+$node->start;
+
+# Create table
+$node->safe_psql("postgres", "CREATE EXTENSION vector;");
+$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector($dim));");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT i, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
+);
+
+# Generate queries
+for (1 .. 20)
+{
+	my @r = ();
+	for (1 .. $dim)
+	{
+		push(@r, rand());
+	}
+	push(@queries, "[" . join(",", @r) . "]");
+}
+
+# Check each index type
+my @operators = ("<->", "<=>");
+my @opclasses = ("vector_l2_ops", "vector_cosine_ops");
+
+for my $i (0 .. $#operators)
+{
+	my $operator = $operators[$i];
+	my $opclass = $opclasses[$i];
+
+	$node->safe_psql("postgres", qq(
+		SET maintenance_work_mem = '128MB';
+		CREATE INDEX idx ON tst USING hnsw (v $opclass);
+	));
+
+	foreach (@cs)
+	{
+		my $c = $_;
+
+		# Get exact results
+		@expected = ();
+		foreach (@queries)
+		{
+			my $res = $node->safe_psql("postgres", qq(
+				SET enable_indexscan = off;
+				WITH top AS (
+					SELECT v $operator '$_' AS distance FROM tst WHERE i % $c = 0 ORDER BY distance LIMIT $limit
+				)
+				SELECT i FROM tst WHERE (v $operator '$_') <= (SELECT MAX(distance) FROM top)
+			));
+			push(@expected, $res);
+		}
+
+		if ($c == 100)
+		{
+			test_recall($c, 40, 0.99, $operator);
+		}
+		else
+		{
+			if ($operator eq "<->")
+			{
+				test_recall($c, 40, 0.99, $operator);
+			}
+			else
+			{
+				test_recall($c, 40, 0.99, $operator);
+			}
+		}
+	}
+
+	$node->safe_psql("postgres", "DROP INDEX idx;");
+}
+
+done_testing();