Added search parameters to EXPLAIN [skip ci]

Updated comment [skip ci]
Updated readme [skip ci]
2026-07-22 20:15:46 +08:00 · 2024-10-28 01:20:59 -07:00 · 2024-10-28 00:56:10 -07:00 · 2024-10-28 00:55:25 -07:00 · 2024-10-28 00:45:26 -07:00 · 2024-10-28 00:43:39 -07:00
28 changed files with 1240 additions and 736 deletions
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -8,8 +8,8 @@ jobs:
      fail-fast: false
      matrix:
        include:
-          # - postgres: 18
-          #   os: ubuntu-24.04
+          - postgres: 18
+            os: ubuntu-24.04
          - postgres: 17
            os: ubuntu-24.04
          - postgres: 16
@@ -49,7 +49,7 @@ jobs:
          - postgres: 16
            os: macos-14
          - postgres: 14
-            os: macos-12
+            os: macos-13
    steps:
      - uses: actions/checkout@v4
      - uses: ankane/setup-postgres@v1
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,6 +1,6 @@
 ## 0.8.0 (unreleased)

- Added support for inline filtering with HNSW
+- Added support for iterative index scans
 - Added casts for arrays to `sparsevec`
 - Improved cost estimation
 - Improved performance of HNSW inserts and on-disk index builds
--- a/README.md
+++ b/README.md
@@ -439,12 +439,6 @@ Create an index on one [or more](https://www.postgresql.org/docs/current/indexes
 CREATE INDEX ON items (category_id);
 ```

-Or a composite HNSW index for approximate search (added in 0.8.0)
-
-```sql
-CREATE INDEX ON items USING hnsw (embedding vector_l2_ops, category_id);
-```
-
 Or a [partial index](https://www.postgresql.org/docs/current/indexes-partial.html) on the vector column for approximate search

 ```sql
@@ -457,6 +451,89 @@ 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);
 ```

+## Iterative Index Scans
+
+*Unreleased*
+
+With approximate indexes, queries with filtering can return less results (due to post-filtering). Starting with 0.8.0, you can enable iterative index scans. If too few results from the initial scan match the filters, the scan will resume until enough results are found (or it reaches `hnsw.max_scan_tuples` or `ivfflat.max_probes`). This can significantly improve recall.
+
+There are two modes for iterative scans: strict and relaxed.
+
+Strict ensures results are in the exact order by distance
+
+```sql
+SET hnsw.iterative_scan = strict_order;
+```
+
+Relaxed allows results to be slightly out of order by distance, but provides better recall
+
+```sql
+SET hnsw.iterative_scan = relaxed_order;
+# or
+SET ivfflat.iterative_scan = relaxed_order;
+```
+
+With relaxed ordering, you can use a [materialized CTE](https://www.postgresql.org/docs/current/queries-with.html#QUERIES-WITH-CTE-MATERIALIZATION) to get strict ordering
+
+```sql
+WITH relaxed_results AS MATERIALIZED (
+    SELECT id, embedding <-> '[1,2,3]' AS distance FROM items WHERE category_id = 123 ORDER BY distance LIMIT 5
+) SELECT * FROM relaxed_results ORDER BY distance;
+```
+
+For queries that filter by distance, use a materialized CTE and place the distance filter outside of it for best performance (due to the [current behavior](https://www.postgresql.org/message-id/flat/CAOdR5yGUoMQ6j7M5hNUXrySzaqZVGf_Ne%2B8fwZMRKTFxU1nbJg%40mail.gmail.com) of the Postgres executor)
+
+```sql
+WITH nearest_results AS MATERIALIZED (
+    SELECT id, embedding <-> '[1,2,3]' AS distance FROM items ORDER BY distance LIMIT 5
+) SELECT * FROM nearest_results WHERE distance < 5 ORDER BY distance;
+```
+
+Note: Place any other filters inside the CTE
+
+### Iterative Scan Options
+
+Since scanning a large portion of an approximate index is expensive, there are options to control when a scan ends
+
+#### HNSW
+
+Specify the max number of tuples to visit (20,000 by default)
+
+```sql
+SET hnsw.max_scan_tuples = 20000;
+```
+
+Note: This is approximate and does not affect the initial scan
+
+Specify the max amount of memory to use, as a multiple of `work_mem` (1 by default)
+
+```sql
+SET hnsw.scan_mem_multiplier = 2;
+```
+
+You can see when increasing this is needed by enabling debug messages
+
+```sql
+SET client_min_messages = debug1;
+```
+
+which will show when a scan reaches the memory limit
+
+```text
+DEBUG:  hnsw index scan reached memory limit after 20000 tuples
+HINT:  Increase hnsw.scan_mem_multiplier to scan more tuples.
+```
+
+#### IVFFlat
+
+Specify the max number of probes
+
+```sql
+SET ivfflat.max_probes = 100;
+```
+
+Note: If this is lower than `ivfflat.probes`, `ivfflat.probes` will be used
+
 ## Half-Precision Vectors

 *Added in 0.7.0*
@@ -1195,7 +1272,6 @@ Thanks to:
 - [k-means++: The Advantage of Careful Seeding](https://theory.stanford.edu/~sergei/papers/kMeansPP-soda.pdf)
 - [Concept Decompositions for Large Sparse Text Data using Clustering](https://www.cs.utexas.edu/users/inderjit/public_papers/concept_mlj.pdf)
 - [Efficient and Robust Approximate Nearest Neighbor Search using Hierarchical Navigable Small World Graphs](https://arxiv.org/ftp/arxiv/papers/1603/1603.09320.pdf)
- [HQANN: Efficient and Robust Similarity Search for Hybrid Queries with Structured and Unstructured Constraints](https://arxiv.org/pdf/2207.07940.pdf)

 ## History

--- a/sql/vector--0.7.4--0.8.0.sql
+++ b/sql/vector--0.7.4--0.8.0.sql
@@ -24,11 +24,3 @@ CREATE CAST (double precision[] AS sparsevec)

 CREATE CAST (numeric[] AS sparsevec)
 	WITH FUNCTION array_to_sparsevec(numeric[], integer, boolean) AS ASSIGNMENT;
-
-CREATE FUNCTION hnsw_attribute_distance(integer, integer) RETURNS float8
-	AS 'MODULE_PATHNAME', 'hnsw_int4_attribute_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE OPERATOR CLASS vector_integer_ops
-	DEFAULT FOR TYPE integer USING hnsw AS
-	OPERATOR 2 = (integer, integer),
-	FUNCTION 4 hnsw_attribute_distance(integer, integer);
--- a/sql/vector.sql
+++ b/sql/vector.sql
@@ -916,13 +916,3 @@ CREATE OPERATOR CLASS sparsevec_l1_ops
 	OPERATOR 1 <+> (sparsevec, sparsevec) FOR ORDER BY float_ops,
 	FUNCTION 1 l1_distance(sparsevec, sparsevec),
 	FUNCTION 3 hnsw_sparsevec_support(internal);
-
-- hnsw attributes
-
-CREATE FUNCTION hnsw_attribute_distance(integer, integer) RETURNS float8
-	AS 'MODULE_PATHNAME', 'hnsw_int4_attribute_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE OPERATOR CLASS vector_integer_ops
-	DEFAULT FOR TYPE integer USING hnsw AS
-	OPERATOR 2 = (integer, integer),
-	FUNCTION 4 hnsw_attribute_distance(integer, integer);
--- a/src/hnsw.c
+++ b/src/hnsw.c
@@ -18,7 +18,17 @@
 #define MarkGUCPrefixReserved(x) EmitWarningsOnPlaceholders(x)
 #endif

+static const struct config_enum_entry hnsw_iterative_scan_options[] = {
+	{"off", HNSW_ITERATIVE_SCAN_OFF, false},
+	{"relaxed_order", HNSW_ITERATIVE_SCAN_RELAXED, false},
+	{"strict_order", HNSW_ITERATIVE_SCAN_STRICT, false},
+	{NULL, 0, false}
+};
+
 int			hnsw_ef_search;
+int			hnsw_iterative_scan;
+int			hnsw_max_scan_tuples;
+double		hnsw_scan_mem_multiplier;
 int			hnsw_lock_tranche_id;
 static relopt_kind hnsw_relopt_kind;

@@ -67,7 +77,21 @@ 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);
+							HNSW_DEFAULT_EF_SEARCH, HNSW_MIN_EF_SEARCH, HNSW_MAX_EF_SEARCH, PGC_USERSET, GUC_EXPLAIN, NULL, NULL, NULL);
+
+	DefineCustomEnumVariable("hnsw.iterative_scan", "Sets the mode for iterative scans",
+							 NULL, &hnsw_iterative_scan,
+							 HNSW_ITERATIVE_SCAN_OFF, hnsw_iterative_scan_options, PGC_USERSET, GUC_EXPLAIN, NULL, NULL, NULL);
+
+	/* This is approximate and does not affect the initial scan */
+	DefineCustomIntVariable("hnsw.max_scan_tuples", "Sets the max number of tuples to visit for iterative scans",
+							NULL, &hnsw_max_scan_tuples,
+							20000, 1, INT_MAX, PGC_USERSET, GUC_EXPLAIN, NULL, NULL, NULL);
+
+	/* Same range as hash_mem_multiplier */
+	DefineCustomRealVariable("hnsw.scan_mem_multiplier", "Sets the multiple of work_mem to use for iterative scans",
+							 NULL, &hnsw_scan_mem_multiplier,
+							 1, 1, 1000, PGC_USERSET, GUC_EXPLAIN, NULL, NULL, NULL);

 	MarkGUCPrefixReserved("hnsw");
 }
@@ -113,6 +137,10 @@ hnswcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 		*indexSelectivity = 0;
 		*indexCorrelation = 0;
 		*indexPages = 0;
+#if PG_VERSION_NUM >= 180000
+		/* See "On disable_cost" thread on pgsql-hackers */
+		path->path.disabled_nodes = 2;
+#endif
 		return;
 	}

@@ -227,13 +255,13 @@ hnswhandler(PG_FUNCTION_ARGS)
 	IndexAmRoutine *amroutine = makeNode(IndexAmRoutine);

 	amroutine->amstrategies = 0;
-	amroutine->amsupport = 4;
+	amroutine->amsupport = 3;
 	amroutine->amoptsprocnum = 0;
 	amroutine->amcanorder = false;
 	amroutine->amcanorderbyop = true;
 	amroutine->amcanbackward = false;	/* can change direction mid-scan */
 	amroutine->amcanunique = false;
-	amroutine->amcanmulticol = true;
+	amroutine->amcanmulticol = false;
 	amroutine->amoptionalkey = true;
 	amroutine->amsearcharray = false;
 	amroutine->amsearchnulls = false;
@@ -285,17 +313,3 @@ hnswhandler(PG_FUNCTION_ARGS)

 	PG_RETURN_POINTER(amroutine);
 }
-
-/*
- * Get the distance between two int4 attributes
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(hnsw_int4_attribute_distance);
-Datum
-hnsw_int4_attribute_distance(PG_FUNCTION_ARGS)
-{
-	int32		a = PG_GETARG_INT32(0);
-	int32		b = PG_GETARG_INT32(1);
-	double		distance = ((double) a) - ((double) b);
-
-	PG_RETURN_FLOAT8(distance);
-}
--- a/src/hnsw.h
+++ b/src/hnsw.h
@@ -19,7 +19,6 @@
 #define HNSW_DISTANCE_PROC 1
 #define HNSW_NORM_PROC 2
 #define HNSW_TYPE_INFO_PROC 3
-#define HNSW_ATTRIBUTE_DISTANCE_PROC 4

 #define HNSW_VERSION	1
 #define HNSW_MAGIC_NUMBER 0xA953A953
@@ -89,6 +88,9 @@
 /* Ensure fits on page and in uint8 */
 #define HnswGetMaxLevel(m) Min(((BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - offsetof(HnswNeighborTupleData, indextids) - sizeof(ItemIdData)) / (sizeof(ItemPointerData)) / (m)) - 2, 255)

+#define HnswGetSearchCandidate(membername, ptr) pairingheap_container(HnswSearchCandidate, membername, ptr)
+#define HnswGetSearchCandidateConst(membername, ptr) pairingheap_const_container(HnswSearchCandidate, membername, ptr)
+
 #define HnswGetValue(base, element) PointerGetDatum(HnswPtrAccess(base, (element)->value))

 #if PG_VERSION_NUM < 140005
@@ -105,12 +107,20 @@
 #define HnswPtrPointer(hp) (hp).ptr
 #define HnswPtrOffset(hp) relptr_offset((hp).relptr)

-#define HnswUseIndexTuple(index) (IndexRelationGetNumberOfAttributes(index) > 1)
-
 /* Variables */
 extern int	hnsw_ef_search;
+extern int	hnsw_iterative_scan;
+extern int	hnsw_max_scan_tuples;
+extern double hnsw_scan_mem_multiplier;
 extern int	hnsw_lock_tranche_id;

+typedef enum HnswIterativeScanMode
+{
+	HNSW_ITERATIVE_SCAN_OFF,
+	HNSW_ITERATIVE_SCAN_RELAXED,
+	HNSW_ITERATIVE_SCAN_STRICT
+}			HnswIterativeScanMode;
+
 typedef struct HnswElementData HnswElementData;
 typedef struct HnswNeighborArray HnswNeighborArray;

@@ -124,7 +134,6 @@ HnswPtrDeclare(HnswElementData, HnswElementRelptr, HnswElementPtr);
 HnswPtrDeclare(HnswNeighborArray, HnswNeighborArrayRelptr, HnswNeighborArrayPtr);
 HnswPtrDeclare(HnswNeighborArrayPtr, HnswNeighborsRelptr, HnswNeighborsPtr);
 HnswPtrDeclare(char, DatumRelptr, DatumPtr);
-HnswPtrDeclare(IndexTupleData, IndexTupleRelptr, IndexTuplePtr);

 struct HnswElementData
 {
@@ -133,6 +142,7 @@ struct HnswElementData
 	uint8		heaptidsLength;
 	uint8		level;
 	uint8		deleted;
+	uint8		version;
 	uint32		hash;
 	HnswNeighborsPtr neighbors;
 	BlockNumber blkno;
@@ -140,7 +150,6 @@ struct HnswElementData
 	OffsetNumber neighborOffno;
 	BlockNumber neighborPage;
 	DatumPtr	value;
-	IndexTuplePtr itup;
 	LWLock		lock;
 };

@@ -166,7 +175,6 @@ typedef struct HnswSearchCandidate
 	pairingheap_node w_node;
 	HnswElementPtr element;
 	double		distance;
-	bool		matches;
 }			HnswSearchCandidate;

 /* HNSW index options */
@@ -243,6 +251,18 @@ typedef struct HnswTypeInfo
 	void		(*checkValue) (Pointer v);
 }			HnswTypeInfo;

+typedef struct HnswSupport
+{
+	FmgrInfo   *procinfo;
+	FmgrInfo   *normprocinfo;
+	Oid			collation;
+}			HnswSupport;
+
+typedef struct HnswQuery
+{
+	Datum		value;
+}			HnswQuery;
+
 typedef struct HnswBuildState
 {
 	/* Info */
@@ -262,17 +282,13 @@ typedef struct HnswBuildState
 	double		reltuples;

 	/* Support functions */
-	FmgrInfo   *procinfo[2];
-	FmgrInfo   *normprocinfo;
-	Oid		   *collation;
+	HnswSupport support;

 	/* Variables */
 	HnswGraph	graphData;
 	HnswGraph  *graph;
 	double		ml;
 	int			maxLevel;
-	bool		useIndexTuple;
-	TupleDesc	tupdesc;

 	/* Memory */
 	MemoryContext graphCtx;
@@ -314,10 +330,10 @@ 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;

@@ -326,24 +342,42 @@ 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 union
+{
+	HnswElement element;
+	ItemPointerData indextid;
+}			HnswUnvisited;
+
 typedef struct HnswScanOpaqueData
 {
 	const		HnswTypeInfo *typeInfo;
 	bool		first;
 	List	   *w;
+	visited_hash v;
+	pairingheap *discarded;
+	HnswQuery	q;
+	int			m;
+	int64		tuples;
+	double		previousDistance;
+	Size		maxMemory;
 	MemoryContext tmpCtx;

 	/* Support functions */
-	FmgrInfo   *procinfo[2];
-	FmgrInfo   *normprocinfo;
-	Oid		   *collation;
+	HnswSupport support;
 }			HnswScanOpaqueData;

 typedef HnswScanOpaqueData * HnswScanOpaque;
@@ -361,8 +395,7 @@ typedef struct HnswVacuumState
 	int			efConstruction;

 	/* Support functions */
-	FmgrInfo   *procinfo[2];
-	Oid		   *collation;
+	HnswSupport support;

 	/* Variables */
 	struct tidhash_hash *deleted;
@@ -378,37 +411,36 @@ typedef struct HnswVacuumState
 int			HnswGetM(Relation index);
 int			HnswGetEfConstruction(Relation index);
 FmgrInfo   *HnswOptionalProcInfo(Relation index, uint16 procnum);
+void		HnswInitSupport(HnswSupport * support, Relation index);
 Datum		HnswNormValue(const HnswTypeInfo * typeInfo, Oid collation, Datum value);
-bool		HnswCheckNorm(FmgrInfo *procinfo, Oid collation, Datum value);
+bool		HnswCheckNorm(HnswSupport * support, Datum value);
 Buffer		HnswNewBuffer(Relation index, ForkNumber forkNum);
 void		HnswInitPage(Buffer buf, Page page);
 void		HnswInit(void);
-List	   *HnswSearchLayer(char *base, Datum q, IndexTuple qtup, ScanKeyData *keyData, List *ep, int ef, int lc, Relation index, FmgrInfo **procinfo, Oid *collation, int m, bool inserting, HnswElement skipElement, bool inMemory);
+List	   *HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation index, HnswSupport * support, 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);
 HnswElement HnswInitElement(char *base, ItemPointer tid, int m, double ml, int maxLevel, HnswAllocator * alloc);
 HnswElement HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno);
-void		HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo **procinfo, Oid *collation, int m, int efConstruction, bool existing, bool inMemory);
-HnswSearchCandidate *HnswEntryCandidate(char *base, HnswElement em, Datum q, IndexTuple qtup, ScanKeyData *keyData, Relation rel, FmgrInfo **procinfo, Oid *collation, bool loadVec, bool inMemory);
+void		HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint, Relation index, HnswSupport * support, int m, int efConstruction, bool existing);
+HnswSearchCandidate *HnswEntryCandidate(char *base, HnswElement em, HnswQuery * q, Relation rel, HnswSupport * support, bool loadVec);
 void		HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, BlockNumber insertPage, ForkNumber forkNum, bool building);
 void		HnswSetNeighborTuple(char *base, HnswNeighborTuple ntup, HnswElement e, int m);
 void		HnswAddHeapTid(HnswElement element, ItemPointer heaptid);
 HnswNeighborArray *HnswInitNeighborArray(int lm, HnswAllocator * allocator);
 void		HnswInitNeighbors(char *base, HnswElement element, int m, HnswAllocator * alloc);
-bool		HnswInsertTupleOnDisk(Relation index, Datum value, Datum *values, bool *isnull, ItemPointer heap_tid, bool building);
-void		HnswUpdateNeighborsOnDisk(Relation index, FmgrInfo **procinfo, Oid *collation, HnswElement e, int m, bool checkExisting, bool building);
-void		HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec, Relation index);
-void		HnswLoadElement(HnswElement element, double *distance, bool *matches, Datum *q, IndexTuple qtup, ScanKeyData *keyData, Relation index, FmgrInfo **procinfo, Oid *collation, bool loadVec, double *maxDistance);
-void		HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element, bool useIndexTuple);
-void		HnswUpdateConnection(char *base, HnswNeighborArray * neighbors, HnswElement newElement, float distance, int lm, int *updateIdx, Relation index, FmgrInfo **procinfo, Oid *collation);
+bool		HnswInsertTupleOnDisk(Relation index, HnswSupport * support, Datum value, ItemPointer heaptid, bool building);
+void		HnswUpdateNeighborsOnDisk(Relation index, HnswSupport * support, HnswElement e, int m, bool checkExisting, bool building);
+void		HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec);
+void		HnswLoadElement(HnswElement element, double *distance, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec, double *maxDistance);
+bool		HnswFormIndexValue(Datum *out, Datum *values, bool *isnull, const HnswTypeInfo * typeInfo, HnswSupport * support);
+void		HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element);
+void		HnswUpdateConnection(char *base, HnswNeighborArray * neighbors, HnswElement newElement, float distance, int lm, int *updateIdx, Relation index, HnswSupport * support);
 bool		HnswLoadNeighborTids(HnswElement element, ItemPointerData *indextids, Relation index, int m, int lm, int lc);
 void		HnswInitLockTranche(void);
 const		HnswTypeInfo *HnswGetTypeInfo(Relation index);
 PGDLLEXPORT void HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc);
-void		HnswInitProcinfo(FmgrInfo **procinfo, Oid **collation, Relation index);
-Size		HnswGetElementTupleSize(char *base, HnswElement element, bool useIndexTuple);
-bool		HnswIndexTupleIsEqual(IndexTuple a, IndexTuple b, TupleDesc tupdesc);

 /* Index access methods */
 IndexBuildResult *hnswbuild(Relation heap, Relation index, IndexInfo *indexInfo);
--- a/src/hnswbuild.c
+++ b/src/hnswbuild.c
@@ -148,7 +148,6 @@ CreateGraphPages(HnswBuildState * buildstate)
 	Page		page;
 	HnswElementPtr iter = buildstate->graph->head;
 	char	   *base = buildstate->hnswarea;
-	bool		useIndexTuple = buildstate->useIndexTuple;

 	/* Calculate sizes */
 	maxSize = HNSW_MAX_SIZE;
@@ -168,6 +167,7 @@ CreateGraphPages(HnswBuildState * buildstate)
 		Size		etupSize;
 		Size		ntupSize;
 		Size		combinedSize;
+		Pointer		valuePtr = HnswPtrAccess(base, element->value);

 		/* Update iterator */
 		iter = element->next;
@@ -176,7 +176,7 @@ CreateGraphPages(HnswBuildState * buildstate)
 		MemSet(etup, 0, HNSW_TUPLE_ALLOC_SIZE);

 		/* Calculate sizes */
-		etupSize = HnswGetElementTupleSize(base, element, useIndexTuple);
+		etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(valuePtr));
 		ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(element->level, buildstate->m);
 		combinedSize = etupSize + ntupSize + sizeof(ItemIdData);

@@ -186,7 +186,7 @@ CreateGraphPages(HnswBuildState * buildstate)
 					(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
 					 errmsg("index tuple too large")));

-		HnswSetElementTuple(base, etup, element, useIndexTuple);
+		HnswSetElementTuple(base, etup, element);

 		/* Keep element and neighbors on the same page if possible */
 		if (PageGetFreeSpace(page) < etupSize || (combinedSize <= maxSize && PageGetFreeSpace(page) < combinedSize))
@@ -327,29 +327,20 @@ AddDuplicateInMemory(HnswElement element, HnswElement dup)
 * Find duplicate element
 */
 static bool
-FindDuplicateInMemory(char *base, HnswElement element, bool useIndexTuple, TupleDesc tupdesc)
+FindDuplicateInMemory(char *base, HnswElement element)
 {
 	HnswNeighborArray *neighbors = HnswGetNeighbors(base, element, 0);
 	Datum		value = HnswGetValue(base, element);
-	IndexTuple	itup = HnswPtrAccess(base, element->itup);

 	for (int i = 0; i < neighbors->length; i++)
 	{
 		HnswCandidate *neighbor = &neighbors->items[i];
 		HnswElement neighborElement = HnswPtrAccess(base, neighbor->element);
+		Datum		neighborValue = HnswGetValue(base, neighborElement);

-		if (useIndexTuple)
-		{
-			/* Exit early since ordered by distance */
-			if (!HnswIndexTupleIsEqual(itup, HnswPtrAccess(base, neighborElement->itup), tupdesc))
-				return false;
-		}
-		else
-		{
-			/* Exit early since ordered by distance */
-			if (!datumIsEqual(value, HnswGetValue(base, neighborElement), false, -1))
-				return false;
-		}
+		/* Exit early since ordered by distance */
+		if (!datumIsEqual(value, neighborValue, false, -1))
+			return false;

 		/* Check for space */
 		if (AddDuplicateInMemory(element, neighborElement))
@@ -375,7 +366,7 @@ AddElementInMemory(char *base, HnswGraph * graph, HnswElement element)
 * Update neighbors
 */
 static void
-UpdateNeighborsInMemory(char *base, Relation index, FmgrInfo **procinfo, Oid *collation, HnswElement e, int m)
+UpdateNeighborsInMemory(char *base, HnswSupport * support, HnswElement e, int m)
 {
 	for (int lc = e->level; lc >= 0; lc--)
 	{
@@ -397,7 +388,7 @@ UpdateNeighborsInMemory(char *base, Relation index, FmgrInfo **procinfo, Oid *co
 			Assert(neighborElement);

 			LWLockAcquire(&neighborElement->lock, LW_EXCLUSIVE);
-			HnswUpdateConnection(base, HnswGetNeighbors(base, neighborElement, lc), e, hc->distance, lm, NULL, index, procinfo, collation);
+			HnswUpdateConnection(base, HnswGetNeighbors(base, neighborElement, lc), e, hc->distance, lm, NULL, NULL, support);
 			LWLockRelease(&neighborElement->lock);
 		}
 	}
@@ -407,20 +398,20 @@ UpdateNeighborsInMemory(char *base, Relation index, FmgrInfo **procinfo, Oid *co
 * Update graph in memory
 */
 static void
-UpdateGraphInMemory(FmgrInfo **procinfo, Oid *collation, HnswElement element, int m, int efConstruction, HnswElement entryPoint, HnswBuildState * buildstate)
+UpdateGraphInMemory(HnswSupport * support, HnswElement element, int m, int efConstruction, HnswElement entryPoint, HnswBuildState * buildstate)
 {
 	HnswGraph  *graph = buildstate->graph;
 	char	   *base = buildstate->hnswarea;

 	/* Look for duplicate */
-	if (FindDuplicateInMemory(base, element, buildstate->useIndexTuple, buildstate->tupdesc))
+	if (FindDuplicateInMemory(base, element))
 		return;

 	/* Add element */
 	AddElementInMemory(base, graph, element);

 	/* Update neighbors */
-	UpdateNeighborsInMemory(base, buildstate->index, procinfo, collation, element, m);
+	UpdateNeighborsInMemory(base, support, element, m);

 	/* Update entry point if needed (already have lock) */
 	if (entryPoint == NULL || element->level > entryPoint->level)
@@ -433,10 +424,8 @@ UpdateGraphInMemory(FmgrInfo **procinfo, Oid *collation, HnswElement element, in
 static void
 InsertTupleInMemory(HnswBuildState * buildstate, HnswElement element)
 {
-	Relation	index = buildstate->index;
-	FmgrInfo  **procinfo = buildstate->procinfo;
-	Oid		   *collation = buildstate->collation;
 	HnswGraph  *graph = buildstate->graph;
+	HnswSupport *support = &buildstate->support;
 	HnswElement entryPoint;
 	LWLock	   *entryLock = &graph->entryLock;
 	LWLock	   *entryWaitLock = &graph->entryWaitLock;
@@ -468,10 +457,10 @@ InsertTupleInMemory(HnswBuildState * buildstate, HnswElement element)
 	}

 	/* Find neighbors for element */
-	HnswFindElementNeighbors(base, element, entryPoint, index, procinfo, collation, m, efConstruction, false, true);
+	HnswFindElementNeighbors(base, element, entryPoint, NULL, support, m, efConstruction, false);

 	/* Update graph in memory */
-	UpdateGraphInMemory(procinfo, collation, element, m, efConstruction, entryPoint, buildstate);
+	UpdateGraphInMemory(support, element, m, efConstruction, entryPoint, buildstate);

 	/* Release entry lock */
 	LWLockRelease(entryLock);
@@ -483,35 +472,19 @@ InsertTupleInMemory(HnswBuildState * buildstate, HnswElement element)
 static bool
 InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, HnswBuildState * buildstate)
 {
-	const		HnswTypeInfo *typeInfo = buildstate->typeInfo;
 	HnswGraph  *graph = buildstate->graph;
 	HnswElement element;
 	HnswAllocator *allocator = &buildstate->allocator;
+	HnswSupport *support = &buildstate->support;
 	Size		valueSize;
 	Pointer		valuePtr;
 	LWLock	   *flushLock = &graph->flushLock;
 	char	   *base = buildstate->hnswarea;
-	bool		useIndexTuple = buildstate->useIndexTuple;
-	TupleDesc	tupdesc = buildstate->tupdesc;
-	IndexTuple	itup;
-	Size		itupSize;
-	IndexTuple	itupPtr;
+	Datum		value;

-	/* Detoast once for all calls */
-	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
-
-	/* Check value */
-	if (typeInfo->checkValue != NULL)
-		typeInfo->checkValue(DatumGetPointer(value));
-
-	/* Normalize if needed */
-	if (buildstate->normprocinfo != NULL)
-	{
-		if (!HnswCheckNorm(buildstate->normprocinfo, buildstate->collation[0], value))
-			return false;
-
-		value = HnswNormValue(typeInfo, buildstate->collation[0], value);
-	}
+	/* Form index value */
+	if (!HnswFormIndexValue(&value, values, isnull, buildstate->typeInfo, support))
+		return false;

 	/* Get datum size */
 	valueSize = VARSIZE_ANY(DatumGetPointer(value));
@@ -524,7 +497,7 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
 	{
 		LWLockRelease(flushLock);

-		return HnswInsertTupleOnDisk(index, value, values, isnull, heaptid, true);
+		return HnswInsertTupleOnDisk(index, support, value, heaptid, true);
 	}

 	/*
@@ -556,22 +529,12 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn

 		LWLockRelease(flushLock);

-		return HnswInsertTupleOnDisk(index, value, values, isnull, heaptid, true);
+		return HnswInsertTupleOnDisk(index, support, value, heaptid, true);
 	}

 	/* Ok, we can proceed to allocate the element */
 	element = HnswInitElement(base, heaptid, buildstate->m, buildstate->ml, buildstate->maxLevel, allocator);
-
-	if (useIndexTuple)
-	{
-		/* TODO fix */
-		values[0] = value;
-		itup = index_form_tuple(tupdesc, values, isnull);
-		itupSize = IndexTupleSize(itup);
-		itupPtr = HnswAlloc(allocator, itupSize);
-	}
-	else
-		valuePtr = HnswAlloc(allocator, valueSize);
+	valuePtr = HnswAlloc(allocator, valueSize);

 	/*
 	 * We have now allocated the space needed for the element, so we don't
@@ -581,19 +544,8 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
 	LWLockRelease(&graph->allocatorLock);

 	/* Copy the datum */
-	if (useIndexTuple)
-	{
-		bool		unused;
-
-		memcpy(itupPtr, itup, itupSize);
-		HnswPtrStore(base, element->itup, itupPtr);
-		HnswPtrStore(base, element->value, DatumGetPointer(index_getattr(itupPtr, 1, tupdesc, &unused)));
-	}
-	else
-	{
-		memcpy(valuePtr, DatumGetPointer(value), valueSize);
-		HnswPtrStore(base, element->value, valuePtr);
-	}
+	memcpy(valuePtr, DatumGetPointer(value), valueSize);
+	HnswPtrStore(base, element->value, valuePtr);

 	/* Create a lock for the element */
 	LWLockInitialize(&element->lock, hnsw_lock_tranche_id);
@@ -720,19 +672,6 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 				 errmsg("type not supported for hnsw index")));

-	/* TODO See if needed */
-	if (IndexRelationGetNumberOfKeyAttributes(index) > 2)
-		elog(ERROR, "index cannot have more than two columns");
-
-	if (!OidIsValid(index_getprocid(index, 1, HNSW_DISTANCE_PROC)))
-		elog(ERROR, "first column must be a vector");
-
-	for (int i = 1; i < IndexRelationGetNumberOfKeyAttributes(index); i++)
-	{
-		if (!OidIsValid(index_getprocid(index, i + 1, HNSW_ATTRIBUTE_DISTANCE_PROC)))
-			elog(ERROR, "column %d cannot be a vector", i + 1);
-	}
-
 	/* Require column to have dimensions to be indexed */
 	if (buildstate->dimensions < 0)
 		ereport(ERROR,
@@ -753,15 +692,12 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 	buildstate->indtuples = 0;

 	/* Get support functions */
-	HnswInitProcinfo(buildstate->procinfo, &buildstate->collation, index);
-	buildstate->normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
+	HnswInitSupport(&buildstate->support, index);

 	InitGraph(&buildstate->graphData, NULL, (Size) maintenance_work_mem * 1024L);
 	buildstate->graph = &buildstate->graphData;
 	buildstate->ml = HnswGetMl(buildstate->m);
 	buildstate->maxLevel = HnswGetMaxLevel(buildstate->m);
-	buildstate->useIndexTuple = HnswUseIndexTuple(index);
-	buildstate->tupdesc = RelationGetDescr(index);

 	buildstate->graphCtx = GenerationContextCreate(CurrentMemoryContext,
 												   "Hnsw build graph context",
--- a/src/hnswinsert.c
+++ b/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)
@@ -153,11 +154,11 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B
 	OffsetNumber freeOffno = InvalidOffsetNumber;
 	OffsetNumber freeNeighborOffno = InvalidOffsetNumber;
 	BlockNumber newInsertPage = InvalidBlockNumber;
+	uint8		tupleVersion;
 	char	   *base = NULL;
-	bool		useIndexTuple = HnswUseIndexTuple(index);

 	/* Calculate sizes */
-	etupSize = HnswGetElementTupleSize(base, e, useIndexTuple);
+	etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(HnswPtrAccess(base, e->value)));
 	ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(e->level, m);
 	combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
 	maxSize = HNSW_MAX_SIZE;
@@ -165,7 +166,7 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B

 	/* Prepare element tuple */
 	etup = palloc0(etupSize);
-	HnswSetElementTuple(base, etup, e, useIndexTuple);
+	HnswSetElementTuple(base, etup, e);

 	/* Prepare neighbor tuple */
 	ntup = palloc0(ntupSize);
@@ -203,7 +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)
 			{
@@ -213,6 +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;
 		}

@@ -369,7 +374,7 @@ HnswLoadNeighbors(HnswElement element, Relation index, int m, int lm, int lc)
 * Load elements for insert
 */
 static void
-LoadElementsForInsert(HnswNeighborArray * neighbors, Datum q, IndexTuple qtup, int *idx, Relation index, FmgrInfo **procinfo, Oid *collation)
+LoadElementsForInsert(HnswNeighborArray * neighbors, HnswQuery * q, int *idx, Relation index, HnswSupport * support)
 {
 	char	   *base = NULL;

@@ -378,9 +383,8 @@ LoadElementsForInsert(HnswNeighborArray * neighbors, Datum q, IndexTuple qtup, i
 		HnswCandidate *hc = &neighbors->items[i];
 		HnswElement element = HnswPtrAccess(base, hc->element);
 		double		distance;
-		bool		matches;

-		HnswLoadElement(element, &distance, &matches, &q, qtup, NULL, index, procinfo, collation, true, NULL);
+		HnswLoadElement(element, &distance, q, index, support, true, NULL);
 		hc->distance = distance;

 		/* Prune element if being deleted */
@@ -396,7 +400,7 @@ LoadElementsForInsert(HnswNeighborArray * neighbors, Datum q, IndexTuple qtup, i
 * Get update index
 */
 static int
-GetUpdateIndex(HnswElement element, HnswElement newElement, float distance, int m, int lm, int lc, Relation index, FmgrInfo **procinfo, Oid *collation, MemoryContext updateCtx)
+GetUpdateIndex(HnswElement element, HnswElement newElement, float distance, int m, int lm, int lc, Relation index, HnswSupport * support, MemoryContext updateCtx)
 {
 	char	   *base = NULL;
 	int			idx = -1;
@@ -421,13 +425,14 @@ GetUpdateIndex(HnswElement element, HnswElement newElement, float distance, int
 		idx = -2;
 	else
 	{
-		Datum		q = HnswGetValue(base, element);
-		IndexTuple	qtup = HnswPtrAccess(base, element->itup);;
+		HnswQuery	q;

-		LoadElementsForInsert(neighbors, q, qtup, &idx, index, procinfo, collation);
+		q.value = HnswGetValue(base, element);
+
+		LoadElementsForInsert(neighbors, &q, &idx, index, support);

 		if (idx == -1)
-			HnswUpdateConnection(base, neighbors, newElement, distance, lm, &idx, index, procinfo, collation);
+			HnswUpdateConnection(base, neighbors, newElement, distance, lm, &idx, index, support);
 	}

 	MemoryContextSwitchTo(oldCtx);
@@ -532,7 +537,7 @@ UpdateNeighborOnDisk(HnswElement element, HnswElement newElement, int idx, int m
 * Update neighbors
 */
 void
-HnswUpdateNeighborsOnDisk(Relation index, FmgrInfo **procinfo, Oid *collation, HnswElement e, int m, bool checkExisting, bool building)
+HnswUpdateNeighborsOnDisk(Relation index, HnswSupport * support, HnswElement e, int m, bool checkExisting, bool building)
 {
 	char	   *base = NULL;

@@ -555,7 +560,7 @@ HnswUpdateNeighborsOnDisk(Relation index, FmgrInfo **procinfo, Oid *collation, H
 			HnswElement neighborElement = HnswPtrAccess(base, hc->element);
 			int			idx;

-			idx = GetUpdateIndex(neighborElement, e, hc->distance, m, lm, lc, index, procinfo, collation, updateCtx);
+			idx = GetUpdateIndex(neighborElement, e, hc->distance, m, lm, lc, index, support, updateCtx);

 			/* New element was not selected as a neighbor */
 			if (idx == -1)
@@ -633,26 +638,16 @@ FindDuplicateOnDisk(Relation index, HnswElement element, bool building)
 	char	   *base = NULL;
 	HnswNeighborArray *neighbors = HnswGetNeighbors(base, element, 0);
 	Datum		value = HnswGetValue(base, element);
-	IndexTuple	itup = HnswPtrAccess(base, element->itup);
-	TupleDesc	tupdesc = RelationGetDescr(index);

 	for (int i = 0; i < neighbors->length; i++)
 	{
 		HnswCandidate *neighbor = &neighbors->items[i];
 		HnswElement neighborElement = HnswPtrAccess(base, neighbor->element);
+		Datum		neighborValue = HnswGetValue(base, neighborElement);

-		if (HnswUseIndexTuple(index))
-		{
-			/* Exit early since ordered by distance */
-			if (!HnswIndexTupleIsEqual(itup, HnswPtrAccess(base, neighborElement->itup), tupdesc))
-				return false;
-		}
-		else
-		{
-			/* Exit early since ordered by distance */
-			if (!datumIsEqual(value, HnswGetValue(base, neighborElement), false, -1))
-				return false;
-		}
+		/* Exit early since ordered by distance */
+		if (!datumIsEqual(value, neighborValue, false, -1))
+			return false;

 		if (AddDuplicateOnDisk(index, element, neighborElement, building))
 			return true;
@@ -665,7 +660,7 @@ FindDuplicateOnDisk(Relation index, HnswElement element, bool building)
 * Update graph on disk
 */
 static void
-UpdateGraphOnDisk(Relation index, FmgrInfo **procinfo, Oid *collation, HnswElement element, int m, int efConstruction, HnswElement entryPoint, bool building)
+UpdateGraphOnDisk(Relation index, HnswSupport * support, HnswElement element, int m, int efConstruction, HnswElement entryPoint, bool building)
 {
 	BlockNumber newInsertPage = InvalidBlockNumber;

@@ -681,7 +676,7 @@ UpdateGraphOnDisk(Relation index, FmgrInfo **procinfo, Oid *collation, HnswEleme
 		HnswUpdateMetaPage(index, 0, NULL, newInsertPage, MAIN_FORKNUM, building);

 	/* Update neighbors */
-	HnswUpdateNeighborsOnDisk(index, procinfo, collation, element, m, false, building);
+	HnswUpdateNeighborsOnDisk(index, support, element, m, false, building);

 	/* Update entry point if needed */
 	if (entryPoint == NULL || element->level > entryPoint->level)
@@ -692,19 +687,15 @@ UpdateGraphOnDisk(Relation index, FmgrInfo **procinfo, Oid *collation, HnswEleme
 * Insert a tuple into the index
 */
 bool
-HnswInsertTupleOnDisk(Relation index, Datum value, Datum *values, bool *isnull, ItemPointer heap_tid, bool building)
+HnswInsertTupleOnDisk(Relation index, HnswSupport * support, Datum value, ItemPointer heaptid, bool building)
 {
 	HnswElement entryPoint;
 	HnswElement element;
 	int			m;
 	int			efConstruction = HnswGetEfConstruction(index);
-	FmgrInfo   *procinfo[2];
-	Oid		   *collation;
 	LOCKMODE	lockmode = ShareLock;
 	char	   *base = NULL;

-	HnswInitProcinfo(procinfo, &collation, index);
-
 	/*
 	 * Get a shared lock. This allows vacuum to ensure no in-flight inserts
 	 * before repairing graph. Use a page lock so it does not interfere with
@@ -716,24 +707,8 @@ HnswInsertTupleOnDisk(Relation index, Datum value, Datum *values, bool *isnull,
 	HnswGetMetaPageInfo(index, &m, &entryPoint);

 	/* Create an element */
-	element = HnswInitElement(base, heap_tid, m, HnswGetMl(m), HnswGetMaxLevel(m), NULL);
-	if (HnswUseIndexTuple(index))
-	{
-		/* TODO no toast */
-		TupleDesc	tupdesc = RelationGetDescr(index);
-		IndexTuple	itup;
-		bool		unused;
-
-		/* TODO fix */
-		values[0] = value;
-		itup = index_form_tuple(tupdesc, values, isnull);
-
-		HnswPtrStore(base, element->itup, itup);
-		HnswPtrStore(base, element->value, DatumGetPointer(index_getattr(itup, 1, tupdesc, &unused)));
-
-	}
-	else
-		HnswPtrStore(base, element->value, DatumGetPointer(value));
+	element = HnswInitElement(base, heaptid, m, HnswGetMl(m), HnswGetMaxLevel(m), NULL);
+	HnswPtrStore(base, element->value, DatumGetPointer(value));

 	/* Prevent concurrent inserts when likely updating entry point */
 	if (entryPoint == NULL || element->level > entryPoint->level)
@@ -750,10 +725,10 @@ HnswInsertTupleOnDisk(Relation index, Datum value, Datum *values, bool *isnull,
 	}

 	/* Find neighbors for element */
-	HnswFindElementNeighbors(base, element, entryPoint, index, procinfo, collation, m, efConstruction, false, false);
+	HnswFindElementNeighbors(base, element, entryPoint, index, support, m, efConstruction, false);

 	/* Update graph on disk */
-	UpdateGraphOnDisk(index, procinfo, collation, element, m, efConstruction, entryPoint, building);
+	UpdateGraphOnDisk(index, support, element, m, efConstruction, entryPoint, building);

 	/* Release lock */
 	UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
@@ -765,31 +740,19 @@ HnswInsertTupleOnDisk(Relation index, Datum value, Datum *values, bool *isnull,
 * Insert a tuple into the index
 */
 static void
-HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid)
+HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid)
 {
 	Datum		value;
 	const		HnswTypeInfo *typeInfo = HnswGetTypeInfo(index);
-	FmgrInfo   *normprocinfo;
-	Oid		   *collation = index->rd_indcollation;
+	HnswSupport support;

-	/* Detoast once for all calls */
-	value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
+	HnswInitSupport(&support, index);

-	/* Check value */
-	if (typeInfo->checkValue != NULL)
-		typeInfo->checkValue(DatumGetPointer(value));
+	/* Form index value */
+	if (!HnswFormIndexValue(&value, values, isnull, typeInfo, &support))
+		return;

-	/* Normalize if needed */
-	normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
-	if (normprocinfo != NULL)
-	{
-		if (!HnswCheckNorm(normprocinfo, collation[0], value))
-			return;
-
-		value = HnswNormValue(typeInfo, collation[0], value);
-	}
-
-	HnswInsertTupleOnDisk(index, value, values, isnull, heap_tid, false);
+	HnswInsertTupleOnDisk(index, &support, value, heaptid, false);
 }

 /*
--- a/src/hnswscan.c
+++ b/src/hnswscan.c
@@ -5,41 +5,74 @@
 #include "pgstat.h"
 #include "storage/bufmgr.h"
 #include "storage/lmgr.h"
+#include "utils/float.h"
 #include "utils/memutils.h"

 /*
 * Algorithm 5 from paper
 */
 static List *
-GetScanItems(IndexScanDesc scan, Datum q)
+GetScanItems(IndexScanDesc scan, Datum value)
 {
 	HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
 	Relation	index = scan->indexRelation;
-	FmgrInfo  **procinfo = so->procinfo;
-	Oid		   *collation = so->collation;
+	HnswSupport *support = &so->support;
 	List	   *ep;
 	List	   *w;
 	int			m;
 	HnswElement entryPoint;
 	char	   *base = NULL;
-	bool		inMemory = false;
-	ScanKeyData *keyData = scan->keyData;
+	HnswQuery  *q = &so->q;

 	/* Get m and entry point */
 	HnswGetMetaPageInfo(index, &m, &entryPoint);

+	q->value = value;
+	so->m = m;
+
 	if (entryPoint == NULL)
 		return NIL;

-	ep = list_make1(HnswEntryCandidate(base, entryPoint, q, NULL, keyData, index, procinfo, collation, false, inMemory));
+	ep = list_make1(HnswEntryCandidate(base, entryPoint, q, index, support, false));

 	for (int lc = entryPoint->level; lc >= 1; lc--)
 	{
-		w = HnswSearchLayer(base, q, NULL, keyData, ep, 1, lc, index, procinfo, collation, m, false, NULL, inMemory);
+		w = HnswSearchLayer(base, q, ep, 1, lc, index, support, m, false, NULL, NULL, NULL, true, NULL);
 		ep = w;
 	}

-	return HnswSearchLayer(base, q, NULL, keyData, ep, hnsw_ef_search, 0, index, procinfo, collation, m, false, NULL, inMemory);
+	return HnswSearchLayer(base, q, ep, hnsw_ef_search, 0, index, support, m, false, NULL, &so->v, hnsw_iterative_scan != HNSW_ITERATIVE_SCAN_OFF ? &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;
+	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 *sc;
+
+		if (pairingheap_is_empty(so->discarded))
+			break;
+
+		sc = HnswGetSearchCandidate(w_node, pairingheap_remove_first(so->discarded));
+
+		ep = lappend(ep, sc);
+	}
+
+	return HnswSearchLayer(base, &so->q, ep, batch_size, 0, index, &so->support, so->m, false, NULL, &so->v, &so->discarded, false, &so->tuples);
 }

 /*
@@ -62,13 +95,24 @@ GetScanValue(IndexScanDesc scan)
 		Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));

 		/* Normalize if needed */
-		if (so->normprocinfo != NULL)
-			value = HnswNormValue(so->typeInfo, so->collation[0], value);
+		if (so->support.normprocinfo != NULL)
+			value = HnswNormValue(so->typeInfo, so->support.collation, value);
 	}

 	return value;
 }

+#if defined(HNSW_MEMORY)
+/*
+ * Show memory usage
+ */
+static void
+ShowMemoryUsage(HnswScanOpaque so)
+{
+	elog(INFO, "memory: %zu KB, tuples: " INT64_FORMAT, MemoryContextMemAllocated(so->tmpCtx, false) / 1024, so->tuples);
+}
+#endif
+
 /*
 * Prepare for an index scan
 */
@@ -77,19 +121,28 @@ hnswbeginscan(Relation index, int nkeys, int norderbys)
 {
 	IndexScanDesc scan;
 	HnswScanOpaque so;
+	double		maxMemory;

 	scan = RelationGetIndexScan(index, nkeys, norderbys);

 	so = (HnswScanOpaque) palloc(sizeof(HnswScanOpaqueData));
 	so->typeInfo = HnswGetTypeInfo(index);
-	so->first = true;
-	so->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
-									   "Hnsw scan temporary context",
-									   ALLOCSET_DEFAULT_SIZES);

 	/* Set support functions */
-	HnswInitProcinfo(so->procinfo, &so->collation, index);
-	so->normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
+	HnswInitSupport(&so->support, index);
+
+	/*
+	 * Use a lower max allocation size than default to allow scanning more
+	 * tuples for iterative search before exceeding work_mem
+	 */
+	so->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
+									   "Hnsw scan temporary context",
+									   0, 8 * 1024, 256 * 1024);
+
+	/* Calculate max memory */
+	/* Add 256 extra bytes to fill last block when close */
+	maxMemory = (double) work_mem * hnsw_scan_mem_multiplier * 1024.0 + 256;
+	so->maxMemory = Min(maxMemory, (double) SIZE_MAX);

 	scan->opaque = so;

@@ -105,6 +158,11 @@ hnswrescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys, int no
 	HnswScanOpaque so = (HnswScanOpaque) scan->opaque;

 	so->first = true;
+	/* v and discarded are allocated in tmpCtx */
+	so->v.tids = NULL;
+	so->discarded = NULL;
+	so->tuples = 0;
+	so->previousDistance = -get_float8_infinity();
 	MemoryContextReset(so->tmpCtx);

 	if (keys && scan->numberOfKeys > 0)
@@ -162,26 +220,104 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
 		so->first = false;

 #if defined(HNSW_MEMORY)
-		elog(INFO, "memory: %zu KB", MemoryContextMemAllocated(so->tmpCtx, false) / 1024);
+		ShowMemoryUsage(so);
 #endif
 	}

-	while (list_length(so->w) > 0)
+	for (;;)
 	{
 		char	   *base = NULL;
-		HnswSearchCandidate *hc = llast(so->w);
-		HnswElement element = HnswPtrAccess(base, hc->element);
+		HnswSearchCandidate *sc;
+		HnswElement element;
 		ItemPointer heaptid;

+		if (list_length(so->w) == 0)
+		{
+			if (hnsw_iterative_scan == HNSW_ITERATIVE_SCAN_OFF)
+				break;
+
+			/* Empty index */
+			if (so->discarded == NULL)
+				break;
+
+			/* Reached max number of tuples */
+			if (so->tuples >= hnsw_max_scan_tuples)
+			{
+				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) > so->maxMemory)
+			{
+				if (pairingheap_is_empty(so->discarded))
+				{
+					ereport(DEBUG1,
+							(errmsg("hnsw index scan reached memory limit after " INT64_FORMAT " tuples", so->tuples),
+							 errhint("Increase hnsw.scan_mem_multiplier 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);
+
+				so->w = ResumeScanItems(scan);
+
+				UnlockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
+
+#if defined(HNSW_MEMORY)
+				ShowMemoryUsage(so);
+#endif
+			}
+
+			if (list_length(so->w) == 0)
+				break;
+		}
+
+		sc = llast(so->w);
+		element = HnswPtrAccess(base, sc->element);
+
 		/* Move to next element if no valid heap TIDs */
-		if (!hc->matches || element->heaptidsLength == 0)
+		if (element->heaptidsLength == 0)
 		{
 			so->w = list_delete_last(so->w);
+
+			/* Mark memory as free for next iteration */
+			if (hnsw_iterative_scan != HNSW_ITERATIVE_SCAN_OFF)
+			{
+				pfree(element);
+				pfree(sc);
+			}
+
 			continue;
 		}

 		heaptid = &element->heaptids[--element->heaptidsLength];

+		if (hnsw_iterative_scan == HNSW_ITERATIVE_SCAN_STRICT)
+		{
+			if (sc->distance < so->previousDistance)
+				continue;
+
+			so->previousDistance = sc->distance;
+		}
+
 		MemoryContextSwitchTo(oldCtx);

 		scan->xs_heaptid = *heaptid;
--- a/src/hnswutils.c
+++ b/src/hnswutils.c
@@ -100,19 +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;
-	ItemPointerData indextid;
-}			HnswUnvisited;
-
 /*
 * Get the max number of connections in an upper layer for each element in the index
 */
@@ -154,41 +141,14 @@ HnswOptionalProcInfo(Relation index, uint16 procnum)
 }

 /*
- * Init procinfo
+ * Init support functions
 */
 void
-HnswInitProcinfo(FmgrInfo **procinfo, Oid **collation, Relation index)
+HnswInitSupport(HnswSupport * support, Relation index)
 {
-	procinfo[0] = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
-
-	if (IndexRelationGetNumberOfKeyAttributes(index) > 1)
-		procinfo[1] = index_getprocinfo(index, 2, HNSW_ATTRIBUTE_DISTANCE_PROC);
-
-	*collation = index->rd_indcollation;
-}
-
-/*
- * Get element tuple size
- */
-Size
-HnswGetElementTupleSize(char *base, HnswElement element, bool useIndexTuple)
-{
-	Size		size;
-
-	if (useIndexTuple)
-	{
-		IndexTuple	itup = HnswPtrAccess(base, element->itup);
-
-		size = IndexTupleSize(itup);
-	}
-	else
-	{
-		Pointer		valuePtr = HnswPtrAccess(base, element->value);
-
-		size = VARSIZE_ANY(valuePtr);
-	}
-
-	return HNSW_ELEMENT_TUPLE_SIZE(size);
+	support->procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
+	support->collation = index->rd_indcollation[0];
+	support->normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
 }

 /*
@@ -204,40 +164,9 @@ HnswNormValue(const HnswTypeInfo * typeInfo, Oid collation, Datum value)
 * Check if non-zero norm
 */
 bool
-HnswCheckNorm(FmgrInfo *procinfo, Oid collation, Datum value)
+HnswCheckNorm(HnswSupport * support, Datum value)
 {
-	return DatumGetFloat8(FunctionCall1Coll(procinfo, collation, value)) > 0;
-}
-
-/*
- * Check if index tuples are equal
- */
-bool
-HnswIndexTupleIsEqual(IndexTuple a, IndexTuple b, TupleDesc tupdesc)
-{
-	for (int i = 0; i < tupdesc->natts; i++)
-	{
-		bool		nullA;
-		bool		nullB;
-
-		Datum		datumA = index_getattr(a, i + 1, tupdesc, &nullA);
-		Datum		datumB = index_getattr(b, i + 1, tupdesc, &nullB);
-
-		if (nullA || nullB)
-		{
-			if (nullA != nullB)
-				return false;
-		}
-		else
-		{
-			Form_pg_attribute att = TupleDescAttr(tupdesc, i);
-
-			if (!datumIsEqual(datumA, datumB, att->attbyval, att->attlen))
-				return false;
-		}
-	}
-
-	return true;
+	return DatumGetFloat8(FunctionCall1Coll(support->normprocinfo, support->collation, value)) > 0;
 }

 /*
@@ -322,11 +251,12 @@ 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);

 	HnswPtrStore(base, element->value, (Pointer) NULL);
-	HnswPtrStore(base, element->itup, (IndexTuple) NULL);

 	return element;
 }
@@ -353,7 +283,6 @@ HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno)
 	element->offno = offno;
 	HnswPtrStore(base, element->neighbors, (HnswNeighborArrayPtr *) NULL);
 	HnswPtrStore(base, element->value, (Pointer) NULL);
-	HnswPtrStore(base, element->itup, (IndexTuple) NULL);
 	return element;
 }

@@ -465,15 +394,45 @@ HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, Bloc
 	UnlockReleaseBuffer(buf);
 }

+/*
+ * Form index value
+ */
+bool
+HnswFormIndexValue(Datum *out, Datum *values, bool *isnull, const HnswTypeInfo * typeInfo, HnswSupport * support)
+{
+	/* Detoast once for all calls */
+	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
+
+	/* Check value */
+	if (typeInfo->checkValue != NULL)
+		typeInfo->checkValue(DatumGetPointer(value));
+
+	/* Normalize if needed */
+	if (support->normprocinfo != NULL)
+	{
+		if (!HnswCheckNorm(support, value))
+			return false;
+
+		value = HnswNormValue(typeInfo, support->collation, value);
+	}
+
+	*out = value;
+
+	return true;
+}
+
 /*
 * Set element tuple, except for neighbor info
 */
 void
-HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element, bool useIndexTuple)
+HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element)
 {
+	Pointer		valuePtr = HnswPtrAccess(base, element->value);
+
 	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)
@@ -481,19 +440,7 @@ HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element, bool
 		else
 			ItemPointerSetInvalid(&etup->heaptids[i]);
 	}
-
-	if (useIndexTuple)
-	{
-		IndexTuple	itup = HnswPtrAccess(base, element->itup);
-
-		memcpy(&etup->data, itup, IndexTupleSize(itup));
-	}
-	else
-	{
-		Pointer		valuePtr = HnswPtrAccess(base, element->value);
-
-		memcpy(&etup->data, valuePtr, VARSIZE_ANY(valuePtr));
-	}
+	memcpy(&etup->data, valuePtr, VARSIZE_ANY(valuePtr));
 }

 /*
@@ -528,16 +475,18 @@ HnswSetNeighborTuple(char *base, HnswNeighborTuple ntup, HnswElement e, int m)
 	}

 	ntup->count = idx;
+	ntup->version = e->version;
 }

 /*
 * Load an element from a tuple
 */
 void
-HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec, Relation index)
+HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec)
 {
 	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;
@@ -557,128 +506,26 @@ HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHe
 	if (loadVec)
 	{
 		char	   *base = NULL;
+		Datum		value = datumCopy(PointerGetDatum(&etup->data), false, -1);

-		if (HnswUseIndexTuple(index))
-		{
-			IndexTuple	itup = CopyIndexTuple((IndexTuple) &etup->data);
-			TupleDesc	tupdesc = RelationGetDescr(index);
-			bool		unused;
-
-			HnswPtrStore(base, element->itup, itup);
-			HnswPtrStore(base, element->value, DatumGetPointer(index_getattr(itup, 1, tupdesc, &unused)));
-		}
-		else
-		{
-			Datum		value = datumCopy(PointerGetDatum(&etup->data), false, -1);
-
-			HnswPtrStore(base, element->value, DatumGetPointer(value));
-		}
+		HnswPtrStore(base, element->value, DatumGetPointer(value));
 	}
 }

-/*
- * Get the attribute distance
- */
-static inline double
-AttributeDistance(double e)
-{
-	/* TODO Better bias */
-	/* must be >> max(w * g) + 1 / log10(2) */
-	double		bias = 4.32;
-
-	return e > 0 ? bias - 1.0 / log10(e + 1) : 0;
-}
-
 /*
 * Calculate the distance between values
 */
-static double
-HnswGetDistance(IndexTuple itup, Datum vec, Datum q, IndexTuple qtup, ScanKeyData *keyData, Relation index, FmgrInfo **procinfo, Oid *collation, bool *matches)
+static inline double
+HnswGetDistance(Datum a, Datum b, HnswSupport * support)
 {
-	double		g;
-
-	if (DatumGetPointer(q) == NULL)
-		g = 0;
-	else
-		g = DatumGetFloat8(FunctionCall2Coll(procinfo[0], collation[0], q, vec));
-
-	Assert(PointerIsValid(matches));
-	*matches = true;
-
-	if (IndexRelationGetNumberOfKeyAttributes(index) > 1)
-	{
-		double		w = 0.25;
-		double		e = 0.0;
-		TupleDesc	tupdesc = RelationGetDescr(index);
-
-		if (keyData)
-		{
-			/* TODO need to pass length of key data */
-			int			keyCount = 1;
-
-			for (int i = 0; i < keyCount; i++)
-			{
-				ScanKey		key = &keyData[i];
-				bool		isnull;
-				Datum		value = index_getattr(itup, key->sk_attno, tupdesc, &isnull);
-				bool		attnull = key->sk_flags & SK_ISNULL;
-
-				if (isnull || attnull)
-				{
-					if (isnull != attnull)
-					{
-						e += 1000;
-						*matches = false;
-					}
-				}
-				else if (!DatumGetBool(FunctionCall2Coll(&key->sk_func, key->sk_collation, value, key->sk_argument)))
-				{
-					double		ei = fabs(DatumGetFloat8(FunctionCall2Coll(procinfo[key->sk_attno - 1], collation[key->sk_attno - 1], value, key->sk_argument)));
-
-					if (ei > 0)
-						e += ei;
-					else
-						/* Distance is zero for inequality */
-						e += 1000;
-
-					*matches = false;
-				}
-			}
-
-			return w * g + AttributeDistance(e);
-		}
-		else if (qtup)
-		{
-			int			keyCount = IndexRelationGetNumberOfKeyAttributes(index) - 1;
-
-			for (int i = 0; i < keyCount; i++)
-			{
-				bool		isnull;
-				bool		attnull;
-				Datum		value = index_getattr(itup, i + 2, tupdesc, &isnull);
-				Datum		value2 = index_getattr(qtup, i + 2, tupdesc, &attnull);
-
-				if (isnull || attnull)
-				{
-					if (isnull != attnull)
-						e += 1000;
-				}
-				else
-					e += fabs(DatumGetFloat8(FunctionCall2Coll(procinfo[i + 1], collation[i + 1], value, value2)));
-			}
-
-			return w * g + AttributeDistance(e);
-		}
-	}
-
-	return g;
+	return DatumGetFloat8(FunctionCall2Coll(support->procinfo, support->collation, a, b));
 }

 /*
 * Load an element and optionally get its distance from q
 */
 static void
-HnswLoadElementImpl(BlockNumber blkno, OffsetNumber offno, double *distance, bool *matches, Datum *q, IndexTuple qtup, ScanKeyData *keyData, Relation index, FmgrInfo **procinfo, Oid *collation, bool loadVec, double *maxDistance, HnswElement * element)
+HnswLoadElementImpl(BlockNumber blkno, OffsetNumber offno, double *distance, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec, double *maxDistance, HnswElement * element)
 {
 	Buffer		buf;
 	Page		page;
@@ -696,23 +543,10 @@ HnswLoadElementImpl(BlockNumber blkno, OffsetNumber offno, double *distance, boo
 	/* Calculate distance */
 	if (distance != NULL)
 	{
-		IndexTuple	itup = NULL;
-		Datum		value;
-
-		if (HnswUseIndexTuple(index))
-		{
-			TupleDesc	tupdesc = RelationGetDescr(index);
-			bool		unused;
-
-			itup = (IndexTuple) &etup->data;
-			value = index_getattr(itup, 1, tupdesc, &unused);
-		}
+		if (DatumGetPointer(q->value) == NULL)
+			*distance = 0;
 		else
-		{
-			value = PointerGetDatum(&etup->data);
-		}
-
-		*distance = HnswGetDistance(itup, value, *q, qtup, keyData, index, procinfo, collation, matches);
+			*distance = HnswGetDistance(q->value, PointerGetDatum(&etup->data), support);
 	}

 	/* Load element */
@@ -721,7 +555,7 @@ HnswLoadElementImpl(BlockNumber blkno, OffsetNumber offno, double *distance, boo
 		if (*element == NULL)
 			*element = HnswInitElementFromBlock(blkno, offno);

-		HnswLoadElementFromTuple(*element, etup, true, loadVec, index);
+		HnswLoadElementFromTuple(*element, etup, true, loadVec);
 	}

 	UnlockReleaseBuffer(buf);
@@ -731,41 +565,51 @@ HnswLoadElementImpl(BlockNumber blkno, OffsetNumber offno, double *distance, boo
 * Load an element and optionally get its distance from q
 */
 void
-HnswLoadElement(HnswElement element, double *distance, bool *matches, Datum *q, IndexTuple qtup, ScanKeyData *keyData, Relation index, FmgrInfo **procinfo, Oid *collation, bool loadVec, double *maxDistance)
+HnswLoadElement(HnswElement element, double *distance, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec, double *maxDistance)
 {
-	HnswLoadElementImpl(element->blkno, element->offno, distance, matches, q, qtup, keyData, index, procinfo, collation, loadVec, maxDistance, &element);
+	HnswLoadElementImpl(element->blkno, element->offno, distance, q, index, support, loadVec, maxDistance, &element);
 }

 /*
 * Get the distance for an element
 */
 static double
-GetElementDistance(char *base, HnswElement element, bool *matches, Datum q, IndexTuple qtup, ScanKeyData *keyData, Relation index, FmgrInfo **procinfo, Oid *collation)
+GetElementDistance(char *base, HnswElement element, HnswQuery * q, HnswSupport * support)
 {
 	Datum		value = HnswGetValue(base, element);
-	IndexTuple	itup = HnswPtrAccess(base, element->itup);

-	return HnswGetDistance(itup, value, q, qtup, keyData, index, procinfo, collation, matches);
+	return HnswGetDistance(q->value, value, support);
+}
+
+/*
+ * Allocate a search candidate
+ */
+static HnswSearchCandidate *
+HnswInitSearchCandidate(char *base, HnswElement element, double distance)
+{
+	HnswSearchCandidate *sc = palloc(sizeof(HnswSearchCandidate));
+
+	HnswPtrStore(base, sc->element, element);
+	sc->distance = distance;
+	return sc;
 }

 /*
 * Create a candidate for the entry point
 */
 HnswSearchCandidate *
-HnswEntryCandidate(char *base, HnswElement entryPoint, Datum q, IndexTuple qtup, ScanKeyData *keyData, Relation index, FmgrInfo **procinfo, Oid *collation, bool loadVec, bool inMemory)
+HnswEntryCandidate(char *base, HnswElement entryPoint, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec)
 {
-	HnswSearchCandidate *sc = palloc(sizeof(HnswSearchCandidate));
+	bool		inMemory = index == NULL;
+	double		distance;

-	HnswPtrStore(base, sc->element, entryPoint);
 	if (inMemory)
-		sc->distance = GetElementDistance(base, entryPoint, &sc->matches, q, qtup, keyData, index, procinfo, collation);
+		distance = GetElementDistance(base, entryPoint, q, support);
 	else
-		HnswLoadElement(entryPoint, &sc->distance, &sc->matches, &q, qtup, keyData, index, procinfo, collation, loadVec, NULL);
-	return sc;
-}
+		HnswLoadElement(entryPoint, &distance, q, index, support, loadVec, NULL);

-#define HnswGetSearchCandidate(membername, ptr) pairingheap_container(HnswSearchCandidate, membername, ptr)
-#define HnswGetSearchCandidateConst(membername, ptr) pairingheap_const_container(HnswSearchCandidate, membername, ptr)
+	return HnswInitSearchCandidate(base, entryPoint, distance);
+}

 /*
 * Compare candidate distances
@@ -782,6 +626,21 @@ 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
 */
@@ -902,8 +761,11 @@ HnswLoadNeighborTids(HnswElement element, ItemPointerData *indextids, Relation i

 	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 false;
@@ -949,23 +811,34 @@ HnswLoadUnvisitedFromDisk(HnswElement element, HnswUnvisited * unvisited, int *u
 * Algorithm 2 from paper
 */
 List *
-HnswSearchLayer(char *base, Datum q, IndexTuple qtup, ScanKeyData *keyData, List *ep, int ef, int lc, Relation index, FmgrInfo **procinfo, Oid *collation, int m, bool inserting, HnswElement skipElement, bool inMemory)
+HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation index, HnswSupport * support, 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;
 	int			lm = HnswGetLayerM(m, lc);
 	HnswUnvisited *unvisited = palloc(lm * sizeof(HnswUnvisited));
 	int			unvisitedLength;
-	uint64		additional = 0;
-	uint64		maxAdditional = keyData && lc == 0 ? 10000 : 0;
+	bool		inMemory = index == NULL;

-	InitVisited(base, &v, inMemory, ef, m);
+	if (v == NULL)
+	{
+		v = &vh;
+		initVisited = true;
+	}
+
+	if (initVisited)
+	{
+		InitVisited(base, v, inMemory, ef, m);
+
+		if (discarded != NULL)
+			*discarded = pairingheap_allocate(CompareNearestDiscardedCandidates, NULL);
+	}

 	/* Create local memory for neighborhood if needed */
 	if (inMemory)
@@ -980,15 +853,18 @@ HnswSearchLayer(char *base, Datum q, IndexTuple qtup, ScanKeyData *keyData, List
 		HnswSearchCandidate *sc = (HnswSearchCandidate *) lfirst(lc2);
 		bool		found;

-		AddToVisited(base, &v, sc->element, inMemory, &found);
+		if (initVisited)
+		{
+			AddToVisited(base, v, sc->element, inMemory, &found);
+
+			/* OK to count elements instead of tuples */
+			if (tuples != NULL)
+				(*tuples)++;
+		}

 		pairingheap_add(C, &sc->c_node);
 		pairingheap_add(W, &sc->w_node);

-		/* Do not count elements that do not match filter towards ef */
-		if (!sc->matches && ++additional <= maxAdditional)
-			continue;
-
 		/*
 		 * Do not count elements being deleted towards ef when vacuuming. It
 		 * would be ideal to do this for inserts as well, but this could
@@ -1010,16 +886,19 @@ HnswSearchLayer(char *base, Datum q, IndexTuple qtup, ScanKeyData *keyData, List
 		cElement = HnswPtrAccess(base, c->element);

 		if (inMemory)
-			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);
+
+		/* OK to count elements instead of tuples */
+		if (tuples != NULL)
+			(*tuples) += unvisitedLength;

 		for (int i = 0; i < unvisitedLength; i++)
 		{
 			HnswElement eElement;
 			HnswSearchCandidate *e;
 			double		eDistance;
-			bool		eMatches;
 			bool		alwaysAdd = wlen < ef;

 			f = HnswGetSearchCandidate(w_node, pairingheap_first(W));
@@ -1027,7 +906,7 @@ HnswSearchLayer(char *base, Datum q, IndexTuple qtup, ScanKeyData *keyData, List
 			if (inMemory)
 			{
 				eElement = unvisited[i].element;
-				eDistance = GetElementDistance(base, eElement, &eMatches, q, qtup, keyData, index, procinfo, collation);
+				eDistance = GetElementDistance(base, eElement, q, support);
 			}
 			else
 			{
@@ -1037,26 +916,30 @@ HnswSearchLayer(char *base, Datum q, IndexTuple qtup, ScanKeyData *keyData, List

 				/* Avoid any allocations if not adding */
 				eElement = NULL;
-				HnswLoadElementImpl(blkno, offno, &eDistance, &eMatches, &q, qtup, keyData, index, procinfo, collation, inserting, alwaysAdd ? NULL : &f->distance, &eElement);
+				HnswLoadElementImpl(blkno, offno, &eDistance, q, index, support, inserting, alwaysAdd || discarded != NULL ? NULL : &f->distance, &eElement);

 				if (eElement == NULL)
 					continue;
 			}

-			if (!(eDistance < f->distance || alwaysAdd))
-				continue;
+			if (eElement == NULL || !(eDistance < f->distance || alwaysAdd))
+			{
+				if (discarded != NULL)
+				{
+					/* Create a new candidate */
+					e = HnswInitSearchCandidate(base, eElement, eDistance);
+					pairingheap_add(*discarded, &e->w_node);
+				}

-			Assert(!eElement->deleted);
+				continue;
+			}

 			/* Make robust to issues */
 			if (eElement->level < lc)
 				continue;

 			/* Create a new candidate */
-			e = palloc(sizeof(HnswSearchCandidate));
-			HnswPtrStore(base, e->element, eElement);
-			e->distance = eDistance;
-			e->matches = eMatches;
+			e = HnswInitSearchCandidate(base, eElement, eDistance);
 			pairingheap_add(C, &e->c_node);
 			pairingheap_add(W, &e->w_node);

@@ -1067,15 +950,16 @@ HnswSearchLayer(char *base, Datum q, IndexTuple qtup, ScanKeyData *keyData, List
 			 */
 			if (CountElement(skipElement, eElement))
 			{
-				/* Do not count elements that do not match filter towards ef */
-				if (!e->matches && ++additional <= maxAdditional)
-					continue;
-
 				wlen++;

 				/* 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);
+				}
 			}
 		}
 	}
@@ -1143,11 +1027,10 @@ CompareCandidateDistancesOffset(const ListCell *a, const ListCell *b)
 * Check if an element is closer to q than any element from R
 */
 static bool
-CheckElementCloser(char *base, HnswCandidate * e, List *r, Relation index, FmgrInfo **procinfo, Oid *collation)
+CheckElementCloser(char *base, HnswCandidate * e, List *r, HnswSupport * support)
 {
 	HnswElement eElement = HnswPtrAccess(base, e->element);
 	Datum		eValue = HnswGetValue(base, eElement);
-	IndexTuple	etup = HnswPtrAccess(base, eElement->itup);
 	ListCell   *lc2;

 	foreach(lc2, r)
@@ -1155,9 +1038,7 @@ CheckElementCloser(char *base, HnswCandidate * e, List *r, Relation index, FmgrI
 		HnswCandidate *ri = lfirst(lc2);
 		HnswElement riElement = HnswPtrAccess(base, ri->element);
 		Datum		riValue = HnswGetValue(base, riElement);
-		IndexTuple	ritup = HnswPtrAccess(base, riElement->itup);
-		bool		matches;
-		float		distance = HnswGetDistance(etup, eValue, riValue, ritup, NULL, index, procinfo, collation, &matches);
+		float		distance = HnswGetDistance(eValue, riValue, support);

 		if (distance <= e->distance)
 			return false;
@@ -1170,7 +1051,7 @@ CheckElementCloser(char *base, HnswCandidate * e, List *r, Relation index, FmgrI
 * Algorithm 4 from paper
 */
 static List *
-SelectNeighbors(char *base, List *c, int lm, Relation index, FmgrInfo **procinfo, Oid *collation, bool *closerSet, HnswCandidate * newCandidate, HnswCandidate * *pruned, bool sortCandidates)
+SelectNeighbors(char *base, List *c, int lm, HnswSupport * support, bool *closerSet, HnswCandidate * newCandidate, HnswCandidate * *pruned, bool sortCandidates)
 {
 	List	   *r = NIL;
 	List	   *w = list_copy(c);
@@ -1204,7 +1085,7 @@ SelectNeighbors(char *base, List *c, int lm, Relation index, FmgrInfo **procinfo

 		/* Use previous state of r and wd to skip work when possible */
 		if (mustCalculate)
-			e->closer = CheckElementCloser(base, e, r, index, procinfo, collation);
+			e->closer = CheckElementCloser(base, e, r, support);
 		else if (list_length(added) > 0)
 		{
 			/* Keep Valgrind happy for in-memory, parallel builds */
@@ -1217,7 +1098,7 @@ SelectNeighbors(char *base, List *c, int lm, Relation index, FmgrInfo **procinfo
 			 */
 			if (e->closer)
 			{
-				e->closer = CheckElementCloser(base, e, added, index, procinfo, collation);
+				e->closer = CheckElementCloser(base, e, added, support);

 				if (!e->closer)
 					removedAny = true;
@@ -1230,7 +1111,7 @@ SelectNeighbors(char *base, List *c, int lm, Relation index, FmgrInfo **procinfo
 				 */
 				if (removedAny)
 				{
-					e->closer = CheckElementCloser(base, e, r, index, procinfo, collation);
+					e->closer = CheckElementCloser(base, e, r, support);
 					if (e->closer)
 						added = lappend(added, e);
 				}
@@ -1238,7 +1119,7 @@ SelectNeighbors(char *base, List *c, int lm, Relation index, FmgrInfo **procinfo
 		}
 		else if (e == newCandidate)
 		{
-			e->closer = CheckElementCloser(base, e, r, index, procinfo, collation);
+			e->closer = CheckElementCloser(base, e, r, support);
 			if (e->closer)
 				added = lappend(added, e);
 		}
@@ -1289,7 +1170,7 @@ AddConnections(char *base, HnswElement element, List *neighbors, int lc)
 * Update connections
 */
 void
-HnswUpdateConnection(char *base, HnswNeighborArray * neighbors, HnswElement newElement, float distance, int lm, int *updateIdx, Relation index, FmgrInfo **procinfo, Oid *collation)
+HnswUpdateConnection(char *base, HnswNeighborArray * neighbors, HnswElement newElement, float distance, int lm, int *updateIdx, Relation index, HnswSupport * support)
 {
 	HnswCandidate newHc;

@@ -1315,7 +1196,7 @@ HnswUpdateConnection(char *base, HnswNeighborArray * neighbors, HnswElement newE
 			c = lappend(c, &neighbors->items[i]);
 		c = lappend(c, &newHc);

-		SelectNeighbors(base, c, lm, index, procinfo, collation, &neighbors->closerSet, &newHc, &pruned, true);
+		SelectNeighbors(base, c, lm, support, &neighbors->closerSet, &newHc, &pruned, true);

 		/* Should not happen */
 		if (pruned == NULL)
@@ -1386,16 +1267,17 @@ PrecomputeHash(char *base, HnswElement element)
 * Algorithm 1 from paper
 */
 void
-HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo **procinfo, Oid *collation, int m, int efConstruction, bool existing, bool inMemory)
+HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint, Relation index, HnswSupport * support, int m, int efConstruction, bool existing)
 {
 	List	   *ep;
 	List	   *w;
 	int			level = element->level;
 	int			entryLevel;
-	Datum		q = HnswGetValue(base, element);
-	IndexTuple	qtup = HnswPtrAccess(base, element->itup);
-	ScanKeyData *keyData = NULL;
+	HnswQuery	q;
 	HnswElement skipElement = existing ? element : NULL;
+	bool		inMemory = index == NULL;
+
+	q.value = HnswGetValue(base, element);

 	/* Precompute hash */
 	if (inMemory)
@@ -1406,13 +1288,13 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
 		return;

 	/* Get entry point and level */
-	ep = list_make1(HnswEntryCandidate(base, entryPoint, q, qtup, keyData, index, procinfo, collation, true, inMemory));
+	ep = list_make1(HnswEntryCandidate(base, entryPoint, &q, index, support, true));
 	entryLevel = entryPoint->level;

 	/* 1st phase: greedy search to insert level */
 	for (int lc = entryLevel; lc >= level + 1; lc--)
 	{
-		w = HnswSearchLayer(base, q, qtup, keyData, ep, 1, lc, index, procinfo, collation, m, true, skipElement, inMemory);
+		w = HnswSearchLayer(base, &q, ep, 1, lc, index, support, m, true, skipElement, NULL, NULL, true, NULL);
 		ep = w;
 	}

@@ -1431,7 +1313,7 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
 		List	   *lw = NIL;
 		ListCell   *lc2;

-		w = HnswSearchLayer(base, q, qtup, keyData, ep, efConstruction, lc, index, procinfo, collation, m, true, skipElement, inMemory);
+		w = HnswSearchLayer(base, &q, ep, efConstruction, lc, index, support, m, true, skipElement, NULL, NULL, true, NULL);

 		/* Convert search candidates to candidates */
 		foreach(lc2, w)
@@ -1455,7 +1337,7 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
 		 * sortCandidates to true for in-memory builds to enable closer
 		 * caching, but there does not seem to be a difference in performance.
 		 */
-		neighbors = SelectNeighbors(base, lw, lm, index, procinfo, collation, &HnswGetNeighbors(base, element, lc)->closerSet, NULL, NULL, false);
+		neighbors = SelectNeighbors(base, lw, lm, support, &HnswGetNeighbors(base, element, lc)->closerSet, NULL, NULL, false);

 		AddConnections(base, element, neighbors, lc);

--- a/src/hnswvacuum.c
+++ b/src/hnswvacuum.c
@@ -184,13 +184,12 @@ static void
 RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswElement entryPoint)
 {
 	Relation	index = vacuumstate->index;
+	HnswSupport *support = &vacuumstate->support;
 	Buffer		buf;
 	Page		page;
 	GenericXLogState *state;
 	int			m = vacuumstate->m;
 	int			efConstruction = vacuumstate->efConstruction;
-	FmgrInfo  **procinfo = vacuumstate->procinfo;
-	Oid		   *collation = vacuumstate->collation;
 	BufferAccessStrategy bas = vacuumstate->bas;
 	HnswNeighborTuple ntup = vacuumstate->ntup;
 	Size		ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(element->level, m);
@@ -205,7 +204,7 @@ RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswEleme
 	element->heaptidsLength = 0;

 	/* Find neighbors for element, skipping itself */
-	HnswFindElementNeighbors(base, element, entryPoint, index, procinfo, collation, m, efConstruction, true, false);
+	HnswFindElementNeighbors(base, element, entryPoint, index, support, m, efConstruction, true);

 	/* Zero memory for each element */
 	MemSet(ntup, 0, HNSW_TUPLE_ALLOC_SIZE);
@@ -229,7 +228,7 @@ RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswEleme
 	UnlockReleaseBuffer(buf);

 	/* Update neighbors */
-	HnswUpdateNeighborsOnDisk(index, procinfo, collation, element, m, true, false);
+	HnswUpdateNeighborsOnDisk(index, support, element, m, true, false);
 }

 /*
@@ -239,6 +238,7 @@ static void
 RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
 {
 	Relation	index = vacuumstate->index;
+	HnswSupport *support = &vacuumstate->support;
 	HnswElement highestPoint = &vacuumstate->highestPoint;
 	HnswElement entryPoint;
 	MemoryContext oldCtx = MemoryContextSwitchTo(vacuumstate->tmpCtx);
@@ -256,7 +256,7 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
 		LockPage(index, HNSW_UPDATE_LOCK, ShareLock);

 		/* Load element */
-		HnswLoadElement(highestPoint, NULL, NULL, NULL, NULL, NULL, index, vacuumstate->procinfo, vacuumstate->collation, true, NULL);
+		HnswLoadElement(highestPoint, NULL, NULL, index, support, true, NULL);

 		/* Repair if needed */
 		if (NeedsUpdated(vacuumstate, highestPoint))
@@ -294,7 +294,7 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
 			 * is outdated, this can remove connections at higher levels in
 			 * the graph until they are repaired, but this should be fine.
 			 */
-			HnswLoadElement(entryPoint, NULL, NULL, NULL, NULL, NULL, index, vacuumstate->procinfo, vacuumstate->collation, true, NULL);
+			HnswLoadElement(entryPoint, NULL, NULL, index, support, true, NULL);

 			if (NeedsUpdated(vacuumstate, entryPoint))
 			{
@@ -370,7 +370,7 @@ RepairGraph(HnswVacuumState * vacuumstate)

 			/* Create an element */
 			element = HnswInitElementFromBlock(blkno, offno);
-			HnswLoadElementFromTuple(element, etup, false, true, index);
+			HnswLoadElementFromTuple(element, etup, false, true);

 			elements = lappend(elements, element);
 		}
@@ -440,7 +440,6 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 	BlockNumber insertPage = InvalidBlockNumber;
 	Relation	index = vacuumstate->index;
 	BufferAccessStrategy bas = vacuumstate->bas;
-	bool		useIndexTuple = HnswUseIndexTuple(index);

 	/*
 	 * Wait for index scans to complete. Scans before this point may contain
@@ -522,19 +521,20 @@ MarkDeleted(HnswVacuumState * vacuumstate)

 			/* Overwrite element */
 			etup->deleted = 1;
-			if (useIndexTuple)
-			{
-				IndexTuple	itup = (IndexTuple) &etup->data;
-
-				MemSet(itup, 0, IndexTupleSize(itup));
-			}
-			else
-				MemSet(&etup->data, 0, VARSIZE_ANY(&etup->data));
+			MemSet(&etup->data, 0, VARSIZE_ANY(&etup->data));

 			/* Overwrite neighbors */
 			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
@@ -581,12 +581,13 @@ InitVacuumState(HnswVacuumState * vacuumstate, IndexVacuumInfo *info, IndexBulkD
 	vacuumstate->callback_state = callback_state;
 	vacuumstate->efConstruction = HnswGetEfConstruction(index);
 	vacuumstate->bas = GetAccessStrategy(BAS_BULKREAD);
-	HnswInitProcinfo(vacuumstate->procinfo, &vacuumstate->collation, index);
 	vacuumstate->ntup = palloc0(HNSW_TUPLE_ALLOC_SIZE);
 	vacuumstate->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
 												"Hnsw vacuum temporary context",
 												ALLOCSET_DEFAULT_SIZES);

+	HnswInitSupport(&vacuumstate->support, index);
+
 	/* Get m from metapage */
 	HnswGetMetaPageInfo(index, &vacuumstate->m, NULL);

--- a/src/ivfbuild.c
+++ b/src/ivfbuild.c
@@ -228,11 +228,11 @@ BuildCallback(Relation index, ItemPointer tid, Datum *values,
 static inline void
 GetNextTuple(Tuplesortstate *sortstate, TupleDesc tupdesc, TupleTableSlot *slot, IndexTuple *itup, int *list)
 {
-	Datum		value;
-	bool		isnull;
-
 	if (tuplesort_gettupleslot(sortstate, true, false, slot, NULL))
 	{
+		Datum		value;
+		bool		isnull;
+
 		*list = DatumGetInt32(slot_getattr(slot, 1, &isnull));
 		value = slot_getattr(slot, 3, &isnull);

@@ -254,8 +254,8 @@ InsertTuples(Relation index, IvfflatBuildState * buildstate, ForkNumber forkNum)
 	IndexTuple	itup = NULL;	/* silence compiler warning */
 	int64		inserted = 0;

-	TupleTableSlot *slot = MakeSingleTupleTableSlot(buildstate->tupdesc, &TTSOpsMinimalTuple);
-	TupleDesc	tupdesc = RelationGetDescr(index);
+	TupleTableSlot *slot = MakeSingleTupleTableSlot(buildstate->sortdesc, &TTSOpsMinimalTuple);
+	TupleDesc	tupdesc = buildstate->tupdesc;

 	pgstat_progress_update_param(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_IVFFLAT_PHASE_LOAD);

@@ -319,6 +319,7 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
 	buildstate->index = index;
 	buildstate->indexInfo = indexInfo;
 	buildstate->typeInfo = IvfflatGetTypeInfo(index);
+	buildstate->tupdesc = RelationGetDescr(index);

 	buildstate->lists = IvfflatGetLists(index);
 	buildstate->dimensions = TupleDescAttr(index->rd_att, 0)->atttypmod;
@@ -356,12 +357,12 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
 				 errmsg("dimensions must be greater than one for this opclass")));

 	/* Create tuple description for sorting */
-	buildstate->tupdesc = CreateTemplateTupleDesc(3);
-	TupleDescInitEntry(buildstate->tupdesc, (AttrNumber) 1, "list", INT4OID, -1, 0);
-	TupleDescInitEntry(buildstate->tupdesc, (AttrNumber) 2, "tid", TIDOID, -1, 0);
-	TupleDescInitEntry(buildstate->tupdesc, (AttrNumber) 3, "vector", RelationGetDescr(index)->attrs[0].atttypid, -1, 0);
+	buildstate->sortdesc = CreateTemplateTupleDesc(3);
+	TupleDescInitEntry(buildstate->sortdesc, (AttrNumber) 1, "list", INT4OID, -1, 0);
+	TupleDescInitEntry(buildstate->sortdesc, (AttrNumber) 2, "tid", TIDOID, -1, 0);
+	TupleDescInitEntry(buildstate->sortdesc, (AttrNumber) 3, "vector", buildstate->tupdesc->attrs[0].atttypid, -1, 0);

-	buildstate->slot = MakeSingleTupleTableSlot(buildstate->tupdesc, &TTSOpsVirtual);
+	buildstate->slot = MakeSingleTupleTableSlot(buildstate->sortdesc, &TTSOpsVirtual);

 	buildstate->centers = VectorArrayInit(buildstate->lists, buildstate->dimensions, buildstate->typeInfo->itemSize(buildstate->dimensions));
 	buildstate->listInfo = palloc(sizeof(ListInfo) * buildstate->lists);
@@ -633,7 +634,7 @@ IvfflatParallelScanAndSort(IvfflatSpool * ivfspool, IvfflatShared * ivfshared, S
 	InitBuildState(&buildstate, ivfspool->heap, ivfspool->index, indexInfo);
 	memcpy(buildstate.centers->items, ivfcenters, buildstate.centers->itemsize * buildstate.centers->maxlen);
 	buildstate.centers->length = buildstate.centers->maxlen;
-	ivfspool->sortstate = InitBuildSortState(buildstate.tupdesc, sortmem, coordinate);
+	ivfspool->sortstate = InitBuildSortState(buildstate.sortdesc, sortmem, coordinate);
 	buildstate.sortstate = ivfspool->sortstate;
 	scan = table_beginscan_parallel(ivfspool->heap,
 									ParallelTableScanFromIvfflatShared(ivfshared));
@@ -950,7 +951,7 @@ AssignTuples(IvfflatBuildState * buildstate)
 	}

 	/* Begin serial/leader tuplesort */
-	buildstate->sortstate = InitBuildSortState(buildstate->tupdesc, maintenance_work_mem, coordinate);
+	buildstate->sortstate = InitBuildSortState(buildstate->sortdesc, maintenance_work_mem, coordinate);

 	/* Add tuples to sort */
 	if (buildstate->heap != NULL)
--- a/src/ivfflat.c
+++ b/src/ivfflat.c
@@ -17,8 +17,16 @@
 #endif

 int			ivfflat_probes;
+int			ivfflat_iterative_scan;
+int			ivfflat_max_probes;
 static relopt_kind ivfflat_relopt_kind;

+static const struct config_enum_entry ivfflat_iterative_scan_options[] = {
+	{"off", IVFFLAT_ITERATIVE_SCAN_OFF, false},
+	{"relaxed_order", IVFFLAT_ITERATIVE_SCAN_RELAXED, false},
+	{NULL, 0, false}
+};
+
 /*
 * Initialize index options and variables
 */
@@ -31,7 +39,16 @@ IvfflatInit(void)

 	DefineCustomIntVariable("ivfflat.probes", "Sets the number of probes",
 							"Valid range is 1..lists.", &ivfflat_probes,
-							IVFFLAT_DEFAULT_PROBES, IVFFLAT_MIN_LISTS, IVFFLAT_MAX_LISTS, PGC_USERSET, 0, NULL, NULL, NULL);
+							IVFFLAT_DEFAULT_PROBES, IVFFLAT_MIN_LISTS, IVFFLAT_MAX_LISTS, PGC_USERSET, GUC_EXPLAIN, NULL, NULL, NULL);
+
+	DefineCustomEnumVariable("ivfflat.iterative_scan", "Sets the mode for iterative scans",
+							 NULL, &ivfflat_iterative_scan,
+							 IVFFLAT_ITERATIVE_SCAN_OFF, ivfflat_iterative_scan_options, PGC_USERSET, GUC_EXPLAIN, NULL, NULL, NULL);
+
+	/* If this is less than probes, probes is used */
+	DefineCustomIntVariable("ivfflat.max_probes", "Sets the max number of probes for iterative scans",
+							NULL, &ivfflat_max_probes,
+							IVFFLAT_MAX_LISTS, IVFFLAT_MIN_LISTS, IVFFLAT_MAX_LISTS, PGC_USERSET, GUC_EXPLAIN, NULL, NULL, NULL);

 	MarkGUCPrefixReserved("ivfflat");
 }
@@ -82,6 +99,10 @@ ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 		*indexSelectivity = 0;
 		*indexCorrelation = 0;
 		*indexPages = 0;
+#if PG_VERSION_NUM >= 180000
+		/* See "On disable_cost" thread on pgsql-hackers */
+		path->path.disabled_nodes = 2;
+#endif
 		return;
 	}

--- a/src/ivfflat.h
+++ b/src/ivfflat.h
@@ -80,6 +80,14 @@

 /* Variables */
 extern int	ivfflat_probes;
+extern int	ivfflat_iterative_scan;
+extern int	ivfflat_max_probes;
+
+typedef enum IvfflatIterativeScanMode
+{
+	IVFFLAT_ITERATIVE_SCAN_OFF,
+	IVFFLAT_ITERATIVE_SCAN_RELAXED
+}			IvfflatIterativeScanMode;

 typedef struct VectorArrayData
 {
@@ -165,6 +173,7 @@ typedef struct IvfflatBuildState
 	Relation	index;
 	IndexInfo  *indexInfo;
 	const		IvfflatTypeInfo *typeInfo;
+	TupleDesc	tupdesc;

 	/* Settings */
 	int			dimensions;
@@ -198,7 +207,7 @@ typedef struct IvfflatBuildState

 	/* Sorting */
 	Tuplesortstate *sortstate;
-	TupleDesc	tupdesc;
+	TupleDesc	sortdesc;
 	TupleTableSlot *slot;

 	/* Memory */
@@ -247,8 +256,11 @@ typedef struct IvfflatScanOpaqueData
 {
 	const		IvfflatTypeInfo *typeInfo;
 	int			probes;
+	int			maxProbes;
 	int			dimensions;
 	bool		first;
+	Datum		value;
+	MemoryContext tmpCtx;

 	/* Sorting */
 	Tuplesortstate *sortstate;
@@ -265,7 +277,9 @@ typedef struct IvfflatScanOpaqueData

 	/* Lists */
 	pairingheap *listQueue;
-	IvfflatScanList lists[FLEXIBLE_ARRAY_MEMBER];	/* must come last */
+	BlockNumber *listPages;
+	int			listIndex;
+	IvfflatScanList *lists;
 }			IvfflatScanOpaqueData;

 typedef IvfflatScanOpaqueData * IvfflatScanOpaque;
--- a/src/ivfinsert.c
+++ b/src/ivfinsert.c
@@ -98,7 +98,7 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
 	IvfflatGetMetaPageInfo(index, NULL, NULL);

 	/* Find the insert page - sets the page and list info */
-	FindInsertPage(index, values, &insertPage, &listInfo);
+	FindInsertPage(index, &value, &insertPage, &listInfo);
 	Assert(BlockNumberIsValid(insertPage));
 	originalInsertPage = insertPage;

--- a/src/ivfscan.c
+++ b/src/ivfscan.c
@@ -10,10 +10,7 @@
 #include "miscadmin.h"
 #include "pgstat.h"
 #include "storage/bufmgr.h"
-
-#ifdef IVFFLAT_MEMORY
 #include "utils/memutils.h"
-#endif

 #define GetScanList(ptr) pairingheap_container(IvfflatScanList, ph_node, ptr)
 #define GetScanListConst(ptr) pairingheap_const_container(IvfflatScanList, ph_node, ptr)
@@ -65,7 +62,7 @@ GetScanLists(IndexScanDesc scan, Datum value)
 			/* Use procinfo from the index instead of scan key for performance */
 			distance = DatumGetFloat8(so->distfunc(so->procinfo, so->collation, PointerGetDatum(&list->center), value));

-			if (listCount < so->probes)
+			if (listCount < so->maxProbes)
 			{
 				IvfflatScanList *scanlist;

@@ -78,7 +75,7 @@ GetScanLists(IndexScanDesc scan, Datum value)
 				pairingheap_add(so->listQueue, &scanlist->ph_node);

 				/* Calculate max distance */
-				if (listCount == so->probes)
+				if (listCount == so->maxProbes)
 					maxDistance = GetScanList(pairingheap_first(so->listQueue))->distance;
 			}
 			else if (distance < maxDistance)
@@ -102,6 +99,11 @@ GetScanLists(IndexScanDesc scan, Datum value)

 		UnlockReleaseBuffer(cbuf);
 	}
+
+	for (int i = listCount - 1; i >= 0; i--)
+		so->listPages[i] = GetScanList(pairingheap_remove_first(so->listQueue))->startPage;
+
+	Assert(pairingheap_is_empty(so->listQueue));
 }

 /*
@@ -114,11 +116,14 @@ GetScanItems(IndexScanDesc scan, Datum value)
 	TupleDesc	tupdesc = RelationGetDescr(scan->indexRelation);
 	double		tuples = 0;
 	TupleTableSlot *slot = so->vslot;
+	int			batchProbes = 0;
+
+	tuplesort_reset(so->sortstate);

 	/* Search closest probes lists */
-	while (!pairingheap_is_empty(so->listQueue))
+	while (so->listIndex < so->maxProbes && (++batchProbes) <= so->probes)
 	{
-		BlockNumber searchPage = GetScanList(pairingheap_remove_first(so->listQueue))->startPage;
+		BlockNumber searchPage = so->listPages[so->listIndex++];

 		/* Search all entry pages for list */
 		while (BlockNumberIsValid(searchPage))
@@ -166,13 +171,17 @@ GetScanItems(IndexScanDesc scan, Datum value)
 		}
 	}

-	if (tuples < 100)
+	if (tuples < 100 && ivfflat_iterative_scan == IVFFLAT_ITERATIVE_SCAN_OFF)
 		ereport(DEBUG1,
 				(errmsg("index scan found few tuples"),
 				 errdetail("Index may have been created with little data."),
 				 errhint("Recreate the index and possibly decrease lists.")));

 	tuplesort_performsort(so->sortstate);
+
+#if defined(IVFFLAT_MEMORY)
+	elog(INFO, "memory: %zu MB", MemoryContextMemAllocated(CurrentMemoryContext, true) / (1024 * 1024));
+#endif
 }

 /*
@@ -209,7 +218,13 @@ GetScanValue(IndexScanDesc scan)

 		/* Normalize if needed */
 		if (so->normprocinfo != NULL)
+		{
+			MemoryContext oldCtx = MemoryContextSwitchTo(so->tmpCtx);
+
 			value = IvfflatNormValue(so->typeInfo, so->collation, value);
+
+			MemoryContextSwitchTo(oldCtx);
+		}
 	}

 	return value;
@@ -240,19 +255,30 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 	int			lists;
 	int			dimensions;
 	int			probes = ivfflat_probes;
+	int			maxProbes;
+	MemoryContext oldCtx;

 	scan = RelationGetIndexScan(index, nkeys, norderbys);

 	/* Get lists and dimensions from metapage */
 	IvfflatGetMetaPageInfo(index, &lists, &dimensions);

+	if (ivfflat_iterative_scan != IVFFLAT_ITERATIVE_SCAN_OFF)
+		maxProbes = Max(ivfflat_max_probes, probes);
+	else
+		maxProbes = probes;
+
 	if (probes > lists)
 		probes = lists;

-	so = (IvfflatScanOpaque) palloc(offsetof(IvfflatScanOpaqueData, lists) + probes * sizeof(IvfflatScanList));
+	if (maxProbes > lists)
+		maxProbes = lists;
+
+	so = (IvfflatScanOpaque) palloc(sizeof(IvfflatScanOpaqueData));
 	so->typeInfo = IvfflatGetTypeInfo(index);
 	so->first = true;
 	so->probes = probes;
+	so->maxProbes = maxProbes;
 	so->dimensions = dimensions;

 	/* Set support functions */
@@ -260,6 +286,12 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 	so->normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
 	so->collation = index->rd_indcollation[0];

+	so->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
+									   "Ivfflat scan temporary context",
+									   ALLOCSET_DEFAULT_SIZES);
+
+	oldCtx = MemoryContextSwitchTo(so->tmpCtx);
+
 	/* Create tuple description for sorting */
 	so->tupdesc = CreateTemplateTupleDesc(2);
 	TupleDescInitEntry(so->tupdesc, (AttrNumber) 1, "distance", FLOAT8OID, -1, 0);
@@ -280,6 +312,11 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 	so->bas = GetAccessStrategy(BAS_BULKREAD);

 	so->listQueue = pairingheap_allocate(CompareLists, scan);
+	so->listPages = palloc(maxProbes * sizeof(BlockNumber));
+	so->listIndex = 0;
+	so->lists = palloc(maxProbes * sizeof(IvfflatScanList));
+
+	MemoryContextSwitchTo(oldCtx);

 	scan->opaque = so;

@@ -294,11 +331,9 @@ ivfflatrescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys, int
 {
 	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;

-	if (!so->first)
-		tuplesort_reset(so->sortstate);
-
 	so->first = true;
 	pairingheap_reset(so->listQueue);
+	so->listIndex = 0;

 	if (keys && scan->numberOfKeys > 0)
 		memmove(scan->keyData, keys, scan->numberOfKeys * sizeof(ScanKeyData));
@@ -314,6 +349,8 @@ bool
 ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 {
 	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
+	ItemPointer heaptid;
+	bool		isnull;

 	/*
 	 * Index can be used to scan backward, but Postgres doesn't support
@@ -341,28 +378,23 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 		IvfflatBench("GetScanLists", GetScanLists(scan, value));
 		IvfflatBench("GetScanItems", GetScanItems(scan, value));
 		so->first = false;
-
-#if defined(IVFFLAT_MEMORY)
-		elog(INFO, "memory: %zu MB", MemoryContextMemAllocated(CurrentMemoryContext, true) / (1024 * 1024));
-#endif
-
-		/* Clean up if we allocated a new value */
-		if (value != scan->orderByData->sk_argument)
-			pfree(DatumGetPointer(value));
+		so->value = value;
 	}

-	if (tuplesort_gettupleslot(so->sortstate, true, false, so->mslot, NULL))
+	while (!tuplesort_gettupleslot(so->sortstate, true, false, so->mslot, NULL))
 	{
-		bool		isnull;
-		ItemPointer heaptid = (ItemPointer) DatumGetPointer(slot_getattr(so->mslot, 2, &isnull));
+		if (so->listIndex == so->maxProbes)
+			return false;

-		scan->xs_heaptid = *heaptid;
-		scan->xs_recheck = false;
-		scan->xs_recheckorderby = false;
-		return true;
+		IvfflatBench("GetScanItems", GetScanItems(scan, so->value));
 	}

-	return false;
+	heaptid = (ItemPointer) DatumGetPointer(slot_getattr(so->mslot, 2, &isnull));
+
+	scan->xs_heaptid = *heaptid;
+	scan->xs_recheck = false;
+	scan->xs_recheckorderby = false;
+	return true;
 }

 /*
@@ -373,12 +405,10 @@ ivfflatendscan(IndexScanDesc scan)
 {
 	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;

-	pairingheap_free(so->listQueue);
+	/* Free any temporary files */
 	tuplesort_end(so->sortstate);
-	FreeAccessStrategy(so->bas);
-	FreeTupleDesc(so->tupdesc);

-	/* TODO Free vslot and mslot without freeing TupleDesc */
+	MemoryContextDelete(so->tmpCtx);

 	pfree(so);
 	scan->opaque = NULL;
--- a/src/ivfvacuum.c
+++ b/src/ivfvacuum.c
@@ -26,7 +26,7 @@ ivfflatbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 		Page		cpage;
 		OffsetNumber coffno;
 		OffsetNumber cmaxoffno;
-		BlockNumber startPages[MaxOffsetNumber];
+		BlockNumber listPages[MaxOffsetNumber];
 		ListInfo	listInfo;

 		cbuf = ReadBuffer(index, blkno);
@@ -40,7 +40,7 @@ ivfflatbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 		{
 			IvfflatList list = (IvfflatList) PageGetItem(cpage, PageGetItemId(cpage, coffno));

-			startPages[coffno - FirstOffsetNumber] = list->startPage;
+			listPages[coffno - FirstOffsetNumber] = list->startPage;
 		}

 		listInfo.blkno = blkno;
@@ -50,7 +50,7 @@ ivfflatbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,

 		for (coffno = FirstOffsetNumber; coffno <= cmaxoffno; coffno = OffsetNumberNext(coffno))
 		{
-			BlockNumber searchPage = startPages[coffno - FirstOffsetNumber];
+			BlockNumber searchPage = listPages[coffno - FirstOffsetNumber];
 			BlockNumber insertPage = InvalidBlockNumber;

 			/* Iterate over entry pages */
--- a/test/expected/hnsw_vector.out
+++ b/test/expected/hnsw_vector.out
@@ -99,6 +99,32 @@ SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <+> (SELECT NULL::vector)) t2
     4
 (1 row)

+DROP TABLE t;
+-- iterative
+CREATE TABLE t (val vector(3));
+INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
+CREATE INDEX ON t USING hnsw (val vector_l2_ops);
+SET hnsw.iterative_scan = strict_order;
+SET hnsw.ef_search = 1;
+SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+   val   
+---------
+ [1,2,3]
+ [1,1,1]
+ [0,0,0]
+(3 rows)
+
+SET hnsw.iterative_scan = relaxed_order;
+SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+   val   
+---------
+ [1,2,3]
+ [1,1,1]
+ [0,0,0]
+(3 rows)
+
+RESET hnsw.iterative_scan;
+RESET hnsw.ef_search;
 DROP TABLE t;
 -- unlogged
 CREATE UNLOGGED TABLE t (val vector(3));
@@ -139,4 +165,29 @@ SET hnsw.ef_search = 0;
 ERROR:  0 is outside the valid range for parameter "hnsw.ef_search" (1 .. 1000)
 SET hnsw.ef_search = 1001;
 ERROR:  1001 is outside the valid range for parameter "hnsw.ef_search" (1 .. 1000)
+SHOW hnsw.iterative_scan;
+ hnsw.iterative_scan 
+---------------------
+ off
+(1 row)
+
+SET hnsw.iterative_scan = on;
+ERROR:  invalid value for parameter "hnsw.iterative_scan": "on"
+HINT:  Available values: off, relaxed_order, strict_order.
+SHOW hnsw.max_scan_tuples;
+ hnsw.max_scan_tuples 
+----------------------
+ 20000
+(1 row)
+
+SET hnsw.max_scan_tuples = 0;
+ERROR:  0 is outside the valid range for parameter "hnsw.max_scan_tuples" (1 .. 2147483647)
+SHOW hnsw.scan_mem_multiplier;
+ hnsw.scan_mem_multiplier 
+--------------------------
+ 1
+(1 row)
+
+SET hnsw.scan_mem_multiplier = 0;
+ERROR:  0 is outside the valid range for parameter "hnsw.scan_mem_multiplier" (1 .. 1000)
 DROP TABLE t;
--- a/test/expected/ivfflat_vector.out
+++ b/test/expected/ivfflat_vector.out
@@ -81,6 +81,37 @@ SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector)) t2
     3
 (1 row)

+DROP TABLE t;
+-- iterative
+CREATE TABLE t (val vector(3));
+INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
+CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 3);
+SET ivfflat.iterative_scan = relaxed_order;
+SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+   val   
+---------
+ [1,2,3]
+ [1,1,1]
+ [0,0,0]
+(3 rows)
+
+SET ivfflat.max_probes = 1;
+SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+   val   
+---------
+ [1,2,3]
+(1 row)
+
+SET ivfflat.max_probes = 2;
+SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+   val   
+---------
+ [1,2,3]
+ [1,1,1]
+(2 rows)
+
+RESET ivfflat.iterative_scan;
+RESET ivfflat.max_probes;
 DROP TABLE t;
 -- unlogged
 CREATE UNLOGGED TABLE t (val vector(3));
@@ -109,4 +140,27 @@ SHOW ivfflat.probes;
 1
 (1 row)

+SET ivfflat.probes = 0;
+ERROR:  0 is outside the valid range for parameter "ivfflat.probes" (1 .. 32768)
+SET ivfflat.probes = 32769;
+ERROR:  32769 is outside the valid range for parameter "ivfflat.probes" (1 .. 32768)
+SHOW ivfflat.iterative_scan;
+ ivfflat.iterative_scan 
+------------------------
+ off
+(1 row)
+
+SET ivfflat.iterative_scan = on;
+ERROR:  invalid value for parameter "ivfflat.iterative_scan": "on"
+HINT:  Available values: off, relaxed_order.
+SHOW ivfflat.max_probes;
+ ivfflat.max_probes 
+--------------------
+ 32768
+(1 row)
+
+SET ivfflat.max_probes = 0;
+ERROR:  0 is outside the valid range for parameter "ivfflat.max_probes" (1 .. 32768)
+SET ivfflat.max_probes = 32769;
+ERROR:  32769 is outside the valid range for parameter "ivfflat.max_probes" (1 .. 32768)
 DROP TABLE t;
--- a/test/sql/hnsw_vector.sql
+++ b/test/sql/hnsw_vector.sql
@@ -57,6 +57,23 @@ SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <+> (SELECT NULL::vector)) t2

 DROP TABLE t;

+-- iterative
+
+CREATE TABLE t (val vector(3));
+INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
+CREATE INDEX ON t USING hnsw (val vector_l2_ops);
+
+SET hnsw.iterative_scan = strict_order;
+SET hnsw.ef_search = 1;
+SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+
+SET hnsw.iterative_scan = relaxed_order;
+SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+
+RESET hnsw.iterative_scan;
+RESET hnsw.ef_search;
+DROP TABLE t;
+
 -- unlogged

 CREATE UNLOGGED TABLE t (val vector(3));
@@ -81,4 +98,16 @@ SHOW hnsw.ef_search;
 SET hnsw.ef_search = 0;
 SET hnsw.ef_search = 1001;

+SHOW hnsw.iterative_scan;
+
+SET hnsw.iterative_scan = on;
+
+SHOW hnsw.max_scan_tuples;
+
+SET hnsw.max_scan_tuples = 0;
+
+SHOW hnsw.scan_mem_multiplier;
+
+SET hnsw.scan_mem_multiplier = 0;
+
 DROP TABLE t;
--- a/test/sql/ivfflat_vector.sql
+++ b/test/sql/ivfflat_vector.sql
@@ -44,6 +44,25 @@ SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector)) t2

 DROP TABLE t;

+-- iterative
+
+CREATE TABLE t (val vector(3));
+INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
+CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 3);
+
+SET ivfflat.iterative_scan = relaxed_order;
+SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+
+SET ivfflat.max_probes = 1;
+SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+
+SET ivfflat.max_probes = 2;
+SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+
+RESET ivfflat.iterative_scan;
+RESET ivfflat.max_probes;
+DROP TABLE t;
+
 -- unlogged

 CREATE UNLOGGED TABLE t (val vector(3));
@@ -62,4 +81,16 @@ CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 32769);

 SHOW ivfflat.probes;

+SET ivfflat.probes = 0;
+SET ivfflat.probes = 32769;
+
+SHOW ivfflat.iterative_scan;
+
+SET ivfflat.iterative_scan = on;
+
+SHOW ivfflat.max_probes;
+
+SET ivfflat.max_probes = 0;
+SET ivfflat.max_probes = 32769;
+
 DROP TABLE t;
--- a/test/t/002_ivfflat_vacuum.pl
+++ b/test/t/002_ivfflat_vacuum.pl
@@ -6,13 +6,7 @@ use Test::More;

 my $dim = 3;

-my @r = ();
-for (1 .. $dim)
-{
-	my $v = int(rand(1000)) + 1;
-	push(@r, "i % $v");
-}
-my $array_sql = join(", ", @r);
+my $array_sql = join(",", ('random()') x $dim);

 # Initialize node
 my $node = PostgreSQL::Test::Cluster->new('node');
@@ -23,19 +17,20 @@ $node->start;
 $node->safe_psql("postgres", "CREATE EXTENSION vector;");
 $node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector($dim));");
 $node->safe_psql("postgres",
-	"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
+	"INSERT INTO tst SELECT i, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
 );
 $node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops);");

 # Get size
 my $size = $node->safe_psql("postgres", "SELECT pg_total_relation_size('tst_v_idx');");

+# Store values
+$node->safe_psql("postgres", "CREATE TABLE tmp AS SELECT * FROM tst;");
+
 # Delete all, vacuum, and insert same data
 $node->safe_psql("postgres", "DELETE FROM tst;");
 $node->safe_psql("postgres", "VACUUM tst;");
-$node->safe_psql("postgres",
-	"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
-);
+$node->safe_psql("postgres", "INSERT INTO tst SELECT * FROM tmp;");

 # Check size
 my $new_size = $node->safe_psql("postgres", "SELECT pg_total_relation_size('tst_v_idx');");
--- a/test/t/041_hnsw_filtering.pl
+++ b/test/t/041_hnsw_filtering.pl
@@ -1,109 +0,0 @@
-use strict;
-use warnings FATAL => 'all';
-use PostgreSQL::Test::Cluster;
-use PostgreSQL::Test::Utils;
-use Test::More;
-
-my $node;
-my @queries = ();
-my @cs = ();
-my @expected;
-my $limit = 20;
-my $dim = 3;
-my $array_sql = join(",", ('random()') x $dim);
-my $nc = 50;
-
-sub test_recall
-{
-	my ($min, $operator) = @_;
-	my $correct = 0;
-	my $total = 0;
-
-	my $explain = $node->safe_psql("postgres", qq(
-		SET enable_seqscan = off;
-		EXPLAIN ANALYZE SELECT i FROM tst WHERE c = $cs[0] ORDER BY v $operator '$queries[0]' LIMIT $limit;
-	));
-	like($explain, qr/Index Cond/);
-
-	for my $i (0 .. $#queries)
-	{
-		my $actual = $node->safe_psql("postgres", qq(
-			SET enable_seqscan = off;
-			SELECT i FROM tst WHERE c = $cs[$i] ORDER BY v $operator '$queries[$i]' LIMIT $limit;
-		));
-		my @actual_ids = split("\n", $actual);
-		my %actual_set = map { $_ => 1 } @actual_ids;
-
-		is(scalar(@actual_ids), $limit);
-
-		my @expected_ids = split("\n", $expected[$i]);
-
-		foreach (@expected_ids)
-		{
-			if (exists($actual_set{$_}))
-			{
-				$correct++;
-			}
-			$total++;
-		}
-	}
-
-	cmp_ok($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($dim), c int4);");
-$node->safe_psql("postgres",
-	"INSERT INTO tst SELECT i, ARRAY[$array_sql], i % $nc FROM generate_series(1, 20000) i;"
-);
-
-# Generate queries
-for (1 .. 20)
-{
-	my @r = ();
-	for (1 .. $dim)
-	{
-		push(@r, rand());
-	}
-	push(@queries, "[" . join(",", @r) . "]");
-	push(@cs, int(rand() * $nc));
-}
-
-# Get exact results
-@expected = ();
-for my $i (0 .. $#queries)
-{
-	my $res = $node->safe_psql("postgres", "SELECT i FROM tst WHERE c = $cs[$i] ORDER BY v <-> '$queries[$i]' LIMIT $limit;");
-	push(@expected, $res);
-}
-
-# Add index
-$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v vector_l2_ops, c);");
-
-# Test recall
-test_recall(0.99, '<->');
-
-# Test vacuum
-$node->safe_psql("postgres", "DELETE FROM tst WHERE c > 5;");
-$node->safe_psql("postgres", "VACUUM tst;");
-
-# Test columns
-my ($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING hnsw (c);");
-like($stderr, qr/first column must be a vector/);
-
-($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING hnsw (c, v vector_l2_ops);");
-like($stderr, qr/first column must be a vector/);
-
-($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING hnsw (v vector_l2_ops, c, c);");
-like($stderr, qr/index cannot have more than two columns/);
-
-($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING hnsw (v vector_l2_ops, v vector_l2_ops);");
-like($stderr, qr/column 2 cannot be a vector/);
-
-done_testing();
--- a/test/t/041_ivfflat_iterative_scan.pl
+++ b/test/t/041_ivfflat_iterative_scan.pl
@@ -0,0 +1,54 @@
+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", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops);");
+
+my $count = $node->safe_psql("postgres", qq(
+	SET enable_seqscan = off;
+	SET ivfflat.probes = 10;
+	SET ivfflat.iterative_scan = relaxed_order;
+	SELECT COUNT(*) FROM (SELECT v FROM tst WHERE i % 10000 = 0 ORDER BY v <-> (SELECT v FROM tst LIMIT 1) LIMIT 11) t;
+));
+is($count, 10);
+
+foreach ((30, 50, 70))
+{
+	my $max_probes = $_;
+	my $expected = $max_probes / 10;
+	my $sum = 0;
+
+	for my $i (1 .. 20)
+	{
+		$count = $node->safe_psql("postgres", qq(
+			SET enable_seqscan = off;
+			SET ivfflat.probes = 10;
+			SET ivfflat.iterative_scan = relaxed_order;
+			SET ivfflat.max_probes = $max_probes;
+			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);
+}
+
+done_testing();
--- a/test/t/042_ivfflat_iterative_scan_recall.pl
+++ b/test/t/042_ivfflat_iterative_scan_recall.pl
@@ -0,0 +1,125 @@
+use strict;
+use warnings FATAL => 'all';
+use PostgreSQL::Test::Cluster;
+use PostgreSQL::Test::Utils;
+use Test::More;
+
+my $node;
+my @queries = ();
+my @expected;
+my $limit = 20;
+my @cs = (100, 1000);
+
+sub test_recall
+{
+	my ($c, $probes, $min, $operator) = @_;
+	my $correct = 0;
+	my $total = 0;
+
+	my $explain = $node->safe_psql("postgres", qq(
+		SET enable_seqscan = off;
+		SET ivfflat.probes = $probes;
+		SET ivfflat.iterative_scan = relaxed_order;
+		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 ivfflat.probes = $probes;
+			SET ivfflat.iterative_scan = relaxed_order;
+			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 $c");
+}
+
+# Initialize node
+$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(3));");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT i, ARRAY[random(), random(), random()] FROM generate_series(1, 100000) i;"
+);
+
+# Generate queries
+for (1 .. 20)
+{
+	my $r1 = rand();
+	my $r2 = rand();
+	my $r3 = rand();
+	push(@queries, "[$r1,$r2,$r3]");
+}
+
+# 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", "CREATE INDEX idx ON tst USING ivfflat (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, 1, 0.57, $operator);
+			test_recall($c, 10, 0.98, $operator);
+		}
+		else
+		{
+			if ($operator eq "<->")
+			{
+				test_recall($c, 1, 0.80, $operator);
+			}
+			else
+			{
+				test_recall($c, 1, 0.88, $operator);
+			}
+		}
+	}
+
+	$node->safe_psql("postgres", "DROP INDEX idx;");
+}
+
+done_testing();
--- a/test/t/043_hnsw_iterative_scan.pl
+++ b/test/t/043_hnsw_iterative_scan.pl
@@ -0,0 +1,68 @@
+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 = $node->safe_psql("postgres", qq(
+	SET enable_seqscan = off;
+	SET hnsw.iterative_scan = relaxed_order;
+	SET hnsw.max_scan_tuples = 100000;
+	SET hnsw.scan_mem_multiplier = 2;
+	SELECT COUNT(*) FROM (SELECT v FROM tst WHERE i % 10000 = 0 ORDER BY v <-> (SELECT v FROM tst LIMIT 1) LIMIT 11) t;
+));
+is($count, 10);
+
+foreach ((30000, 50000, 70000))
+{
+	my $max_tuples = $_;
+	my $expected = $max_tuples / 10000;
+	my $sum = 0;
+
+	for my $i (1 .. 20)
+	{
+		$count = $node->safe_psql("postgres", qq(
+			SET enable_seqscan = off;
+			SET hnsw.iterative_scan = relaxed_order;
+			SET hnsw.max_scan_tuples = $max_tuples;
+			SET hnsw.scan_mem_multiplier = 2;
+			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) = $node->psql("postgres", qq(
+	SET enable_seqscan = off;
+	SET hnsw.iterative_scan = relaxed_order;
+	SET client_min_messages = debug1;
+	SET work_mem = '1MB';
+	SELECT COUNT(*) FROM (SELECT v FROM tst WHERE i % 10000 = 0 ORDER BY v <-> (SELECT v FROM tst LIMIT 1) LIMIT 11) t;
+));
+like($stderr, qr/hnsw index scan reached memory limit after \d+ tuples/);
+
+done_testing();
--- a/test/t/044_hnsw_iterative_scan_recall.pl
+++ b/test/t/044_hnsw_iterative_scan_recall.pl
@@ -0,0 +1,118 @@
+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 $dim);
+my @cs = (50, 500);
+
+sub test_recall
+{
+	my ($c, $ef_search, $min, $operator, $mode) = @_;
+	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.iterative_scan = $mode;
+		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.iterative_scan = $mode;
+			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 $mode $c");
+}
+
+# Initialize node
+$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, 50000) 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, $res);
+		}
+
+		test_recall($c, 40, 0.99, $operator, "strict_order");
+		test_recall($c, 40, 0.99, $operator, "relaxed_order");
+	}
+
+	$node->safe_psql("postgres", "DROP INDEX idx;");
+}
+
+done_testing();
Author	SHA1	Message	Date
Andrew Kane	fc0d3e7fdb	Added search parameters to EXPLAIN [skip ci]	2024-10-28 01:20:59 -07:00
Andrew Kane	c530a3c490	Updated comment [skip ci]	2024-10-28 00:56:10 -07:00
Andrew Kane	d8b9e8ef73	Updated readme [skip ci]	2024-10-28 00:55:25 -07:00
Andrew Kane	00894efed5	Updated readme [skip ci]	2024-10-28 00:45:26 -07:00
Andrew Kane	0aa0f6619b	Updated readme [skip ci]	2024-10-28 00:43:39 -07:00
Andrew Kane	38d053001e	Updated readme [skip ci]	2024-10-28 00:32:12 -07:00
Andrew Kane	ccb95407e7	Updated readme [skip ci]	2024-10-28 00:09:26 -07:00
Andrew Kane	04d5e934a1	Scan 2000 more tuples with default work_mem	2024-10-27 22:23:28 -07:00
Andrew Kane	b163b5b196	Moved code [skip ci]	2024-10-27 22:21:43 -07:00
Andrew Kane	6a30c1e824	Fixed test [skip ci]	2024-10-27 21:10:51 -07:00
Andrew Kane	2db1b19644	Use greater than [skip ci]	2024-10-27 21:07:22 -07:00
Andrew Kane	305d62146e	Updated comment [skip ci]	2024-10-27 21:05:32 -07:00
Andrew Kane	f9d627c9a9	Updated default value of hnsw.scan_mem_multiplier [skip ci]	2024-10-27 21:05:04 -07:00
Andrew Kane	38f42820be	Added test for hnsw.scan_mem_multiplier [skip ci]	2024-10-27 20:05:58 -07:00
Andrew Kane	15c8245b42	Updated readme [skip ci]	2024-10-27 19:30:40 -07:00
Andrew Kane	572a9ab404	Updated readme [skip ci]	2024-10-27 18:58:50 -07:00
Andrew Kane	00492d7e57	Ensure max memory fits into Size for HNSW index scans	2024-10-27 14:21:15 -07:00
Andrew Kane	857d716d9e	Renamed iterative_search to iterative_scan	2024-10-27 14:02:22 -07:00
Andrew Kane	c5dd2af750	Added comments [skip ci]	2024-10-25 21:39:03 -07:00
Andrew Kane	78b877bdaf	Revert "Renamed iterative_search to iterative_scan" This reverts commit `7043cce893`.	2024-10-24 20:32:07 -07:00
Andrew Kane	7043cce893	Renamed iterative_search to iterative_scan	2024-10-24 20:31:43 -07:00
Andrew Kane	62039d74f6	Added iterative search section to readme [skip ci]	2024-10-24 18:05:29 -07:00
Andrew Kane	ac6576e53a	Added hnsw.search_mem_multiplier option	2024-10-24 18:02:20 -07:00
Andrew Kane	67eff41c44	Updated changelog [skip ci]	2024-10-23 15:12:40 -07:00
Andrew Kane	1291b12090	Added Postgres 18 to CI [skip ci]	2024-10-22 00:38:19 -07:00
Andrew Kane	24522700b8	Improved hnswrescan	2024-10-21 23:41:32 -07:00
Andrew Kane	bfb3a45b31	Use consistent order [skip ci]	2024-10-21 21:47:03 -07:00
Andrew Kane	e718eb8da4	Updated range and defaults for iterative search parameters	2024-10-21 20:38:50 -07:00
Andrew Kane	049972a4a3	Improved test output [skip ci]	2024-10-13 17:22:49 -07:00
Andrew Kane	61027645e9	Improved test output [skip ci]	2024-10-13 17:21:38 -07:00
Andrew Kane	a41b327b33	Speed up test [skip ci]	2024-10-13 17:12:12 -07:00
Andrew Kane	7f735ebd9b	Added test for strict order [skip ci]	2024-10-13 17:04:03 -07:00
Andrew Kane	02b01e1ca9	Show tuples with memory usage [skip ci]	2024-10-12 22:05:13 -07:00
Andrew Kane	388e42f6e6	Fixed flaky test [skip ci]	2024-10-11 15:48:19 -07:00
Andrew Kane	bf379eed86	Use a memory context for IVFFlat index scans	2024-10-11 15:46:38 -07:00
Andrew Kane	e1bc929429	Simplify lists for IvfflatScanOpaque [skip ci]	2024-10-11 15:29:23 -07:00
Andrew Kane	38285aacc7	Revert "Updated IVFFlat to support multiple attributes (not enabled yet)" This reverts commit `772ab69de6`.	2024-10-11 15:01:54 -07:00
Andrew Kane	a2408e60fa	Revert "Added IndexTuple to HNSW elements (first step to support multiple attributes)" This reverts commit `53a8734bac`.	2024-10-11 14:57:57 -07:00
Andrew Kane	53a8734bac	Added IndexTuple to HNSW elements (first step to support multiple attributes)	2024-10-11 14:12:01 -07:00
Andrew Kane	7484625227	Added comments [skip ci]	2024-10-11 11:59:36 -07:00
Andrew Kane	d1ebb8db73	Use -1 for no limit for ivfflat.max_probes [skip ci]	2024-10-11 11:43:32 -07:00
Andrew Kane	42af8aa1d1	Updated GUC descriptions [skip ci]	2024-10-11 11:26:27 -07:00
Andrew Kane	9d15a76b60	Improved enum naming [skip ci]	2024-10-11 11:20:36 -07:00
Andrew Kane	a3a20f9816	Simplified GUC names [skip ci]	2024-10-11 11:18:01 -07:00
Andrew Kane	b26a21b848	Added regression tests for iterative search [skip ci]	2024-10-11 11:07:11 -07:00
Andrew Kane	2dc392ed6c	Updated GUC names [skip ci]	2024-10-10 23:50:11 -07:00
Andrew Kane	960d2848cb	Updated comment [skip ci]	2024-10-10 21:02:33 -07:00
Andrew Kane	8e88b481a6	Use a lower max allocation size than default to allow scanning more tuples for iterative search before exceeding work_mem [skip ci]	2024-10-10 20:57:57 -07:00
Andrew Kane	124018b8dd	Added HnswInitSearchCandidate function	2024-10-10 19:30:47 -07:00
Andrew Kane	35b252a3e3	Switched to macos-13 on CI [skip ci]	2024-10-10 18:33:35 -07:00
Andrew Kane	2832e746f0	Use get_float8_infinity instead of INFINITY Co-authored-by: "Jonathan S. Katz" <jkatz@amazon.com>	2024-10-10 18:16:39 -07:00
Andrew Kane	961cb17d80	Added iterative search for HNSW [skip ci]	2024-10-10 18:14:39 -07:00
Andrew Kane	c91ed7b2c3	Added iterative search for IVFFlat [skip ci]	2024-10-10 18:12:27 -07:00
Andrew Kane	48fe70c219	Fixed sortdesc for multiple attributes [skip ci]	2024-10-10 15:22:23 -07:00
Andrew Kane	29908405ab	List assignment does not change with scaling [skip ci]	2024-10-10 15:16:08 -07:00
Andrew Kane	08d0340655	Improved IVFFlat vacuum test [skip ci]	2024-10-10 14:24:26 -07:00
Andrew Kane	7d2eb49c2a	Fixed list assignment for cosine distance and non-normalized vectors for IVFFlat	2024-10-10 12:54:13 -07:00
Andrew Kane	772ab69de6	Updated IVFFlat to support multiple attributes (not enabled yet)	2024-10-10 12:34:03 -07:00
Andrew Kane	e13e9a9614	Improved scoping [skip ci]	2024-10-10 12:19:08 -07:00
Andrew Kane	e2fab306ac	Added tupdesc to buildstate [skip ci]	2024-10-10 12:18:41 -07:00
Andrew Kane	edc2126a4a	Renamed tupdesc to sortdesc [skip ci]	2024-10-10 12:16:57 -07:00
Andrew Kane	2cbd08b6c0	Moved unions and macros [skip ci]	2024-10-10 09:41:26 -07:00
Andrew Kane	fa6782985a	Added HnswQuery struct for query data	2024-10-09 23:45:47 -07:00
Andrew Kane	32ab27d72a	Added HnswSupport struct for support functions	2024-10-09 23:10:26 -07:00
Andrew Kane	064db12de7	Moved procinfo initialization for inserts [skip ci]	2024-10-09 21:59:21 -07:00
Andrew Kane	45a6eef9e0	Improved variable name [skip ci]	2024-10-09 21:52:10 -07:00
Andrew Kane	17266ed409	Use inMemory for conditionals	2024-10-09 21:49:32 -07:00
Andrew Kane	a98534e5ab	DRY HNSW procinfo	2024-10-09 21:03:18 -07:00
Andrew Kane	57c05c59a2	DRY code for forming index value	2024-10-09 20:50:17 -07:00