Compare commits

..

18 Commits

Author SHA1 Message Date
Andrew Kane
f858705293 No dimensions 2023-10-16 16:20:10 -07:00
Andrew Kane
e630efd195 Version bump to 0.5.1 [skip ci] 2023-10-10 17:40:57 -07:00
Andrew Kane
b5b912906b Added check for MVCC-compliant snapshot and removed marking tuples as dead for IVFFlat index scans - closes #260 2023-10-10 17:28:48 -07:00
Andrew Kane
4b5db94307 Disable closer caching for new elements for now 2023-10-06 14:27:09 -07:00
Andrew Kane
65e70326b8 Updated comment [skip ci] 2023-10-06 14:07:35 -07:00
Andrew Kane
71641ed84e Updated comment [skip ci] 2023-10-06 13:58:07 -07:00
Andrew Kane
f3dba25036 Added comment [skip ci] 2023-10-06 13:56:25 -07:00
Andrew Kane
5588ba6410 Improved variable name [skip ci] 2023-10-06 13:46:19 -07:00
Andrew Kane
ec9fac5456 Improved closerSet logic 2023-10-06 13:39:55 -07:00
Andrew Kane
8085d3e538 Moved sorting logic into SelectNeighbors 2023-10-06 12:56:15 -07:00
Andrew Kane
cae162ffc6 Ensure order is deterministic for SelectNeighbors closer caching 2023-10-06 12:26:53 -07:00
Andrew Kane
62482e3760 Use e for consistency 2023-10-05 16:15:13 -07:00
Heikki Linnakangas
c81302b835 Improve HNSW index build performance more (#295)
This takes the approach from commit a713e2acaa further. Once we have
remove a candidate from the "closer" set, we still don't need to
recalculate everything that follows. Any candidates that were in the
closer set before still only need to be compared with any new
candidates that we have added.
2023-10-05 16:04:50 -07:00
Andrew Kane
a713e2acaa Improved performance of HNSW index builds - closes #292
Co-authored-by: Heikki Linnakangas <heikki.linnakangas@iki.fi>
2023-10-05 13:21:26 -07:00
Andrew Kane
6e1312ddbe DRY max size 2023-10-04 21:43:34 -07:00
Andrew Kane
4ef5bca275 Use BLCKSZ for consistency 2023-10-04 21:37:55 -07:00
Xiaoran Wang
1ecf6ada76 Include ItemIdData when computing the maxSize for the data in a page (#274)
As the data is aligned, for hnsw, the combined size won't be in the range
(8156 (maxSize exlucding `ItemIdData`), 8160]. So even if the
ItemIdData is not included in the maxSize, it works well now, but I
think it's better to make it correct.
2023-10-04 21:26:46 -07:00
Andrew Kane
564a3d45fc Added check for MVCC-compliant snapshot for HNSW index scans - closes #281 2023-10-04 20:14:50 -07:00
18 changed files with 277 additions and 184 deletions

View File

@@ -39,10 +39,6 @@ jobs:
sudo apt-get update
sudo apt-get install libipc-run-perl
- run: make prove_installcheck
- if: ${{ matrix.os == 'ubuntu-22.04' }}
run: make clean && make
env:
PG_CFLAGS: -Werror -fanalyzer
mac:
runs-on: macos-latest
if: ${{ !startsWith(github.ref_name, 'windows') }}

View File

@@ -1,6 +1,7 @@
## 0.5.1 (unreleased)
## 0.5.1 (2023-10-10)
- Improved performance of index scans for IVFFlat after updates and deletes
- Improved performance of HNSW index builds
- Added check for MVCC-compliant snapshot for index scans
## 0.5.0 (2023-08-28)

View File

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

View File

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

View File

@@ -1,5 +1,5 @@
EXTENSION = vector
EXTVERSION = 0.5.0
EXTVERSION = 0.5.1
OBJS = src\hnsw.obj src\hnswbuild.obj src\hnswinsert.obj src\hnswscan.obj src\hnswutils.obj src\hnswvacuum.obj src\ivfbuild.obj src\ivfflat.obj src\ivfinsert.obj src\ivfkmeans.obj src\ivfscan.obj src\ivfutils.obj src\ivfvacuum.obj src\vector.obj
HEADERS = src\vector.h

View File

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

View File

@@ -0,0 +1,2 @@
-- complain if script is sourced in psql, rather than via CREATE EXTENSION
\echo Use "ALTER EXTENSION vector UPDATE TO '0.5.1'" to load this file. \quit

View File

@@ -57,7 +57,9 @@
/* PROGRESS_CREATEIDX_SUBPHASE_INITIALIZE is 1 */
#define PROGRESS_HNSW_PHASE_LOAD 2
#define HNSW_ELEMENT_TUPLE_SIZE(_dim) MAXALIGN(offsetof(HnswElementTupleData, vec) + VECTOR_SIZE(_dim))
#define HNSW_MAX_SIZE (BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - sizeof(ItemIdData))
#define HNSW_ELEMENT_TUPLE_SIZE(_datum) MAXALIGN(offsetof(HnswElementTupleData, value) + VARSIZE_ANY(_datum))
#define HNSW_NEIGHBOR_TUPLE_SIZE(level, m) MAXALIGN(offsetof(HnswNeighborTupleData, indextids) + ((level) + 2) * (m) * sizeof(ItemPointerData))
#define HnswPageGetOpaque(page) ((HnswPageOpaque) PageGetSpecialPointer(page))
@@ -96,12 +98,13 @@ typedef struct HnswElementData
List *heaptids;
uint8 level;
uint8 deleted;
bool loaded;
HnswNeighborArray *neighbors;
BlockNumber blkno;
OffsetNumber offno;
OffsetNumber neighborOffno;
BlockNumber neighborPage;
Vector *vec;
Datum value;
} HnswElementData;
typedef HnswElementData * HnswElement;
@@ -110,11 +113,13 @@ typedef struct HnswCandidate
{
HnswElement element;
float distance;
bool closer;
} HnswCandidate;
typedef struct HnswNeighborArray
{
int length;
bool closerSet;
HnswCandidate *items;
} HnswNeighborArray;
@@ -200,7 +205,7 @@ typedef struct HnswElementTupleData
ItemPointerData heaptids[HNSW_HEAPTIDS];
ItemPointerData neighbortid;
uint16 unused2;
Vector vec;
char value[FLEXIBLE_ARRAY_MEMBER];
} HnswElementTupleData;
typedef HnswElementTupleData * HnswElementTuple;

View File

@@ -8,6 +8,7 @@
#include "lib/pairingheap.h"
#include "nodes/pg_list.h"
#include "storage/bufmgr.h"
#include "utils/datum.h"
#include "utils/memutils.h"
#if PG_VERSION_NUM >= 140000
@@ -105,8 +106,6 @@ CreateElementPages(HnswBuildState * buildstate)
{
Relation index = buildstate->index;
ForkNumber forkNum = buildstate->forkNum;
int dimensions = buildstate->dimensions;
Size etupSize;
Size maxSize;
HnswElementTuple etup;
HnswNeighborTuple ntup;
@@ -117,12 +116,11 @@ CreateElementPages(HnswBuildState * buildstate)
ListCell *lc;
/* Calculate sizes */
maxSize = BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData));
etupSize = HNSW_ELEMENT_TUPLE_SIZE(dimensions);
maxSize = HNSW_MAX_SIZE;
/* Allocate once */
etup = palloc0(etupSize);
ntup = palloc0(maxSize);
etup = palloc0(BLCKSZ);
ntup = palloc0(BLCKSZ);
/* Prepare first page */
buf = HnswNewBuffer(index, forkNum);
@@ -133,12 +131,14 @@ CreateElementPages(HnswBuildState * buildstate)
foreach(lc, buildstate->elements)
{
HnswElement element = lfirst(lc);
Size etupSize;
Size ntupSize;
Size combinedSize;
HnswSetElementTuple(etup, element);
/* Calculate sizes */
etupSize = HNSW_ELEMENT_TUPLE_SIZE(element->value);
ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(element->level, buildstate->m);
combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
@@ -273,18 +273,15 @@ InsertTuple(Relation index, Datum *values, HnswElement element, HnswBuildState *
int m = buildstate->m;
/* Detoast once for all calls */
Datum value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
element->value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
/* Normalize if needed */
if (buildstate->normprocinfo != NULL)
{
if (!HnswNormValue(buildstate->normprocinfo, collation, &value, buildstate->normvec))
if (!HnswNormValue(buildstate->normprocinfo, collation, &element->value, buildstate->normvec))
return false;
}
/* Copy value to element so accessible outside of memory context */
memcpy(element->vec, DatumGetVector(value), VECTOR_SIZE(buildstate->dimensions));
/* Insert element in graph */
HnswInsertElement(element, entryPoint, NULL, procinfo, collation, m, efConstruction, false);
@@ -360,7 +357,6 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
/* Allocate necessary memory outside of memory context */
element = HnswInitElement(tid, buildstate->m, buildstate->ml, buildstate->maxLevel);
element->vec = palloc(VECTOR_SIZE(buildstate->dimensions));
/* Use memory context since detoast can allocate */
oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
@@ -368,9 +364,8 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
/* Insert tuple */
inserted = InsertTuple(index, values, element, buildstate, &dup);
/* Reset memory context */
/* Switch memory context */
MemoryContextSwitchTo(oldCtx);
MemoryContextReset(buildstate->tmpCtx);
/* Add outside memory context */
if (dup != NULL)
@@ -378,9 +373,16 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
/* Add to buildstate or free */
if (inserted)
{
element->value = datumCopy(element->value, false, -1);
element->loaded = true;
buildstate->elements = lappend(buildstate->elements, element);
}
else
HnswFreeElement(element);
/* Reset memory context */
MemoryContextReset(buildstate->tmpCtx);
}
/*
@@ -395,6 +397,7 @@ HnswGetMaxInMemoryElements(int m, double ml, int dimensions)
elementSize += sizeof(HnswNeighborArray) * (avgLevel + 1);
elementSize += sizeof(HnswCandidate) * (m * (avgLevel + 2));
elementSize += sizeof(ItemPointerData);
/* TODO Handle non-vector types */
elementSize += VECTOR_SIZE(dimensions);
return (maintenance_work_mem * 1024L) / elementSize;
}
@@ -414,10 +417,6 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
buildstate->efConstruction = HnswGetEfConstruction(index);
buildstate->dimensions = TupleDescAttr(index->rd_att, 0)->atttypmod;
/* Require column to have dimensions to be indexed */
if (buildstate->dimensions < 0)
elog(ERROR, "column does not have dimensions");
if (buildstate->dimensions > HNSW_MAX_DIM)
elog(ERROR, "column cannot have more than %d dimensions for hnsw index", HNSW_MAX_DIM);
@@ -440,6 +439,7 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
buildstate->flushed = false;
/* Reuse for each tuple */
/* TODO Fix / replace with support function */
buildstate->normvec = InitVector(buildstate->dimensions);
buildstate->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,

View File

@@ -123,7 +123,6 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
Size minCombinedSize;
HnswElementTuple etup;
BlockNumber currentPage = insertPage;
int dimensions = e->vec->dim;
HnswNeighborTuple ntup;
Buffer nbuf;
Page npage;
@@ -132,10 +131,10 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
BlockNumber newInsertPage = InvalidBlockNumber;
/* Calculate sizes */
etupSize = HNSW_ELEMENT_TUPLE_SIZE(dimensions);
etupSize = HNSW_ELEMENT_TUPLE_SIZE(e->value);
ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(e->level, m);
combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
maxSize = BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData));
maxSize = HNSW_MAX_SIZE;
minCombinedSize = etupSize + HNSW_NEIGHBOR_TUPLE_SIZE(0, m) + sizeof(ItemIdData);
/* Prepare element tuple */
@@ -405,7 +404,7 @@ HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup)
Buffer buf;
Page page;
GenericXLogState *state;
Size etupSize = HNSW_ELEMENT_TUPLE_SIZE(dup->vec->dim);
Size etupSize = HNSW_ELEMENT_TUPLE_SIZE(dup->value);
HnswElementTuple etup;
int i;
@@ -515,7 +514,7 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
/* Create an element */
element = HnswInitElement(heap_tid, m, HnswGetMl(m), HnswGetMaxLevel(m));
element->vec = DatumGetVector(value);
element->value = value;
/* Prevent concurrent inserts when likely updating entry point */
if (entryPoint == NULL || element->level > entryPoint->level)

View File

@@ -160,6 +160,11 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
if (scan->orderByData == NULL)
elog(ERROR, "cannot scan hnsw index without order");
/* Requires MVCC-compliant snapshot as not able to maintain a pin */
/* https://www.postgresql.org/docs/current/index-locking.html */
if (!IsMVCCSnapshot(scan->xs_snapshot))
elog(ERROR, "non-MVCC snapshots are not supported with hnsw");
/* Get scan value */
value = GetScanValue(scan);
@@ -201,15 +206,6 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
scan->xs_ctup.t_self = *heaptid;
#endif
/*
* Typically, an index scan must maintain a pin on the index page
* holding the item last returned by amgettuple. However, this is not
* needed with the current vacuum strategy, which ensures scans do not
* visit tuples in danger of being marked as deleted.
*
* https://www.postgresql.org/docs/current/index-locking.html
*/
scan->xs_recheckorderby = false;
return true;
}

View File

@@ -4,6 +4,7 @@
#include "hnsw.h"
#include "storage/bufmgr.h"
#include "utils/datum.h"
#include "vector.h"
/*
@@ -139,6 +140,7 @@ HnswInitNeighbors(HnswElement element, int m)
a = &element->neighbors[lc];
a->length = 0;
a->items = palloc(sizeof(HnswCandidate) * lm);
a->closerSet = false;
}
}
@@ -186,7 +188,8 @@ HnswFreeElement(HnswElement element)
{
HnswFreeNeighbors(element);
list_free_deep(element->heaptids);
pfree(element->vec);
if (element->loaded)
pfree(DatumGetPointer(element->value));
pfree(element);
}
@@ -213,7 +216,7 @@ HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno)
element->blkno = blkno;
element->offno = offno;
element->neighbors = NULL;
element->vec = NULL;
element->loaded = false;
return element;
}
@@ -323,7 +326,7 @@ HnswSetElementTuple(HnswElementTuple etup, HnswElement element)
else
ItemPointerSetInvalid(&etup->heaptids[i]);
}
memcpy(&etup->vec, element->vec, VECTOR_SIZE(element->vec->dim));
memcpy(&etup->value, DatumGetPointer(element->value), VARSIZE_ANY(element->value));
}
/*
@@ -446,8 +449,10 @@ HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHe
if (loadVec)
{
element->vec = palloc(VECTOR_SIZE(etup->vec.dim));
memcpy(element->vec, &etup->vec, VECTOR_SIZE(etup->vec.dim));
Datum value = PointerGetDatum(&etup->value);
element->value = datumCopy(value, false, -1);
element->loaded = true;
}
}
@@ -475,7 +480,7 @@ HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index,
/* Calculate distance */
if (distance != NULL)
*distance = (float) DatumGetFloat8(FunctionCall2Coll(procinfo, collation, *q, PointerGetDatum(&etup->vec)));
*distance = (float) DatumGetFloat8(FunctionCall2Coll(procinfo, collation, *q, PointerGetDatum(&etup->value)));
UnlockReleaseBuffer(buf);
}
@@ -486,7 +491,7 @@ HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index,
static float
GetCandidateDistance(HnswCandidate * hc, Datum q, FmgrInfo *procinfo, Oid collation)
{
return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, q, PointerGetDatum(hc->element->vec)));
return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, q, hc->element->value));
}
/*
@@ -692,6 +697,34 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
return w;
}
/*
* Compare candidate distances
*/
static int
#if PG_VERSION_NUM >= 130000
CompareCandidateDistances(const ListCell *a, const ListCell *b)
#else
CompareCandidateDistances(const void *a, const void *b)
#endif
{
HnswCandidate *hca = lfirst((ListCell *) a);
HnswCandidate *hcb = lfirst((ListCell *) b);
if (hca->distance < hcb->distance)
return 1;
if (hca->distance > hcb->distance)
return -1;
if (hca->element < hcb->element)
return 1;
if (hca->element > hcb->element)
return -1;
return 0;
}
/*
* Calculate the distance between elements
*/
@@ -721,7 +754,7 @@ HnswGetDistance(HnswElement a, HnswElement b, int lc, FmgrInfo *procinfo, Oid co
}
}
return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(a->vec), PointerGetDatum(b->vec)));
return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, a->value, b->value));
}
/*
@@ -748,33 +781,77 @@ CheckElementCloser(HnswCandidate * e, List *r, int lc, FmgrInfo *procinfo, Oid c
* Algorithm 4 from paper
*/
static List *
SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswCandidate * *pruned)
SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswElement e2, HnswCandidate * newCandidate, HnswCandidate * *pruned, bool sortCandidates)
{
List *r = NIL;
List *w = list_copy(c);
pairingheap *wd;
bool mustCalculate = !e2->neighbors[lc].closerSet;
List *added = NIL;
bool removedAny = false;
if (list_length(w) <= m)
return w;
wd = pairingheap_allocate(CompareNearestCandidates, NULL);
/* Ensure order of candidates is deterministic for closer caching */
if (sortCandidates)
list_sort(w, CompareCandidateDistances);
while (list_length(w) > 0 && list_length(r) < m)
{
/* Assumes w is already ordered desc */
HnswCandidate *e = llast(w);
bool closer;
w = list_delete_last(w);
closer = CheckElementCloser(e, r, lc, procinfo, collation);
/* Use previous state of r and wd to skip work when possible */
if (mustCalculate)
e->closer = CheckElementCloser(e, r, lc, procinfo, collation);
else if (list_length(added) > 0)
{
/*
* If the current candidate was closer, we only need to compare it
* with the other candidates that we have added.
*/
if (e->closer)
{
e->closer = CheckElementCloser(e, added, lc, procinfo, collation);
if (closer)
if (!e->closer)
removedAny = true;
}
else
{
/*
* If we have removed any candidates from closer, a candidate
* that was not closer earlier might now be.
*/
if (removedAny)
{
e->closer = CheckElementCloser(e, r, lc, procinfo, collation);
if (e->closer)
added = lappend(added, e);
}
}
}
else if (e == newCandidate)
{
e->closer = CheckElementCloser(e, r, lc, procinfo, collation);
if (e->closer)
added = lappend(added, e);
}
if (e->closer)
r = lappend(r, e);
else
pairingheap_add(wd, &(CreatePairingHeapNode(e)->ph_node));
}
/* Cached value can only be used in future if sorted deterministically */
e2->neighbors[lc].closerSet = sortCandidates;
/* Keep pruned connections */
while (!pairingheap_is_empty(wd) && list_length(r) < m)
r = lappend(r, ((HnswPairingHeapNode *) pairingheap_remove_first(wd))->inner);
@@ -804,7 +881,7 @@ HnswFindDuplicate(HnswElement e)
HnswCandidate *neighbor = &neighbors->items[i];
/* Exit early since ordered by distance */
if (vector_cmp_internal(e->vec, neighbor->element->vec) != 0)
if (!datumIsEqual(e->value, neighbor->element->value, false, -1))
break;
/* Check for space */
@@ -828,28 +905,6 @@ AddConnections(HnswElement element, List *neighbors, int m, int lc)
a->items[a->length++] = *((HnswCandidate *) lfirst(lc2));
}
/*
* Compare candidate distances
*/
static int
#if PG_VERSION_NUM >= 130000
CompareCandidateDistances(const ListCell *a, const ListCell *b)
#else
CompareCandidateDistances(const void *a, const void *b)
#endif
{
HnswCandidate *hca = lfirst((ListCell *) a);
HnswCandidate *hcb = lfirst((ListCell *) b);
if (hca->distance < hcb->distance)
return 1;
if (hca->distance > hcb->distance)
return -1;
return 0;
}
/*
* Update connections
*/
@@ -879,13 +934,13 @@ HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int
/* Load elements on insert */
if (index != NULL)
{
Datum q = PointerGetDatum(hc->element->vec);
Datum q = hc->element->value;
for (int i = 0; i < currentNeighbors->length; i++)
{
HnswCandidate *hc3 = &currentNeighbors->items[i];
if (hc3->element->vec == NULL)
if (!hc3->element->loaded)
HnswLoadElement(hc3->element, &hc3->distance, &q, index, procinfo, collation, true);
else
hc3->distance = GetCandidateDistance(hc3, q, procinfo, collation);
@@ -903,13 +958,12 @@ HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int
{
List *c = NIL;
/* Add and sort candidates */
/* Add candidates */
for (int i = 0; i < currentNeighbors->length; i++)
c = lappend(c, &currentNeighbors->items[i]);
c = lappend(c, &hc2);
list_sort(c, CompareCandidateDistances);
SelectNeighbors(c, m, lc, procinfo, collation, &pruned);
SelectNeighbors(c, m, lc, procinfo, collation, hc->element, &hc2, &pruned, true);
/* Should not happen */
if (pruned == NULL)
@@ -967,7 +1021,7 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
List *w;
int level = element->level;
int entryLevel;
Datum q = PointerGetDatum(element->vec);
Datum q = element->value;
HnswElement skipElement = existing ? element : NULL;
/* No neighbors if no entry point */
@@ -1008,7 +1062,12 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
else
lw = w;
neighbors = SelectNeighbors(lw, lm, lc, procinfo, collation, NULL);
/*
* Candidates are sorted, but not deterministically. Could set
* sortCandidates to true for in-memory builds to enable closer
* caching, but there does not seem to be a difference in performance.
*/
neighbors = SelectNeighbors(lw, lm, lc, procinfo, collation, element, NULL, NULL, false);
AddConnections(element, neighbors, lm, lc);

View File

@@ -93,7 +93,7 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
if (itemUpdated)
{
Size etupSize = HNSW_ELEMENT_TUPLE_SIZE(etup->vec.dim);
Size etupSize = HNSW_ELEMENT_TUPLE_SIZE(PointerGetDatum(&etup->value));
/* Mark rest as invalid */
for (int i = idx; i < HNSW_HEAPTIDS; i++)
@@ -481,6 +481,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
HnswNeighborTuple ntup;
Size etupSize;
Size ntupSize;
Datum value;
Buffer nbuf;
Page npage;
BlockNumber neighborPage;
@@ -504,8 +505,11 @@ MarkDeleted(HnswVacuumState * vacuumstate)
if (ItemPointerIsValid(&etup->heaptids[0]))
continue;
/* Get datum */
value = PointerGetDatum(&etup->value);
/* Calculate sizes */
etupSize = HNSW_ELEMENT_TUPLE_SIZE(etup->vec.dim);
etupSize = HNSW_ELEMENT_TUPLE_SIZE(value);
ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(etup->level, vacuumstate->m);
/* Get neighbor page */
@@ -528,7 +532,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
/* Overwrite element */
etup->deleted = 1;
MemSet(&etup->vec.x, 0, etup->vec.dim * sizeof(float));
MemSet(&etup->value, 0, VARSIZE_ANY(value));
/* Overwrite neighbors */
for (int i = 0; i < ntup->count; i++)

View File

@@ -246,8 +246,6 @@ typedef struct IvfflatScanOpaqueData
int probes;
int dimensions;
bool first;
Buffer buf;
ItemPointerData heaptid;
/* Sorting */
Tuplesortstate *sortstate;

View File

@@ -99,7 +99,7 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
/* Get tuple size */
itemsz = MAXALIGN(IndexTupleSize(itup));
Assert(itemsz <= BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(IvfflatPageOpaqueData)));
Assert(itemsz <= BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(IvfflatPageOpaqueData)) - sizeof(ItemIdData));
/* Find a page to insert the item */
for (;;)

View File

@@ -143,10 +143,6 @@ GetScanItems(IndexScanDesc scan, Datum value)
bool isnull;
ItemId itemid = PageGetItemId(page, offno);
/* Skip dead tuples */
if (scan->ignore_killed_tuples && ItemIdIsDead(itemid))
continue;
itup = (IndexTuple) PageGetItem(page, itemid);
datum = index_getattr(itup, 1, tupdesc, &isnull);
@@ -161,8 +157,6 @@ GetScanItems(IndexScanDesc scan, Datum value)
slot->tts_isnull[0] = false;
slot->tts_values[1] = PointerGetDatum(&itup->t_tid);
slot->tts_isnull[1] = false;
slot->tts_values[2] = Int32GetDatum((int) searchPage);
slot->tts_isnull[2] = false;
ExecStoreVirtualTuple(slot);
tuplesort_puttupleslot(so->sortstate, slot);
@@ -187,55 +181,6 @@ GetScanItems(IndexScanDesc scan, Datum value)
tuplesort_performsort(so->sortstate);
}
/*
* Mark prior tuple as dead
*/
static void
MarkPriorTupleDead(IndexScanDesc scan)
{
IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
Buffer buf = so->buf;
Page page;
OffsetNumber maxoffno;
/* Safety check */
if (!BufferIsValid(so->buf) || !ItemPointerIsValid(&so->heaptid))
return;
/* Only a shared locked is needed for ItemIdMarkDead */
LockBuffer(buf, BUFFER_LOCK_SHARE);
page = BufferGetPage(buf);
maxoffno = PageGetMaxOffsetNumber(page);
for (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
{
ItemId itemid = PageGetItemId(page, offno);
IndexTuple itup = (IndexTuple) PageGetItem(page, itemid);
/*
* Find tuple. Since buffer has been pinned, tuple cannot have been
* vacuumed (and heap TID reused).
*/
if (ItemPointerEquals(&itup->t_tid, &so->heaptid))
{
/*
* Make sure tuple has not already been marked dead to avoid extra
* WAL if wal_log_hints or data checksums enabled
*/
if (!ItemIdIsDead(itemid))
{
ItemIdMarkDead(itemid);
MarkBufferDirtyHint(buf, true);
}
break;
}
}
/* Unlock buffer */
LockBuffer(buf, BUFFER_LOCK_UNLOCK);
}
/*
* Prepare for an index scan
*/
@@ -261,9 +206,7 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
probes = lists;
so = (IvfflatScanOpaque) palloc(offsetof(IvfflatScanOpaqueData, lists) + probes * sizeof(IvfflatScanList));
so->buf = InvalidBuffer;
so->first = true;
ItemPointerSetInvalid(&so->heaptid);
so->probes = probes;
so->dimensions = dimensions;
@@ -274,13 +217,12 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
/* Create tuple description for sorting */
#if PG_VERSION_NUM >= 120000
so->tupdesc = CreateTemplateTupleDesc(3);
so->tupdesc = CreateTemplateTupleDesc(2);
#else
so->tupdesc = CreateTemplateTupleDesc(3, false);
so->tupdesc = CreateTemplateTupleDesc(2, false);
#endif
TupleDescInitEntry(so->tupdesc, (AttrNumber) 1, "distance", FLOAT8OID, -1, 0);
TupleDescInitEntry(so->tupdesc, (AttrNumber) 2, "heaptid", TIDOID, -1, 0);
TupleDescInitEntry(so->tupdesc, (AttrNumber) 3, "indexblkno", INT4OID, -1, 0);
/* Prep sort */
so->sortstate = tuplesort_begin_heap(so->tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, work_mem, NULL, false);
@@ -312,7 +254,6 @@ ivfflatrescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys, int
#endif
so->first = true;
ItemPointerSetInvalid(&so->heaptid);
pairingheap_reset(so->listQueue);
if (keys && scan->numberOfKeys > 0)
@@ -347,6 +288,11 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
if (scan->orderByData == NULL)
elog(ERROR, "cannot scan ivfflat index without order");
/* Requires MVCC-compliant snapshot as not able to pin during sorting */
/* https://www.postgresql.org/docs/current/index-locking.html */
if (!IsMVCCSnapshot(scan->xs_snapshot))
elog(ERROR, "non-MVCC snapshots are not supported with ivfflat");
if (scan->orderByData->sk_flags & SK_ISNULL)
value = PointerGetDatum(InitVector(so->dimensions));
else
@@ -370,17 +316,10 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
if (value != scan->orderByData->sk_argument)
pfree(DatumGetPointer(value));
}
else
{
/* Mark prior tuple as dead */
if (scan->kill_prior_tuple)
MarkPriorTupleDead(scan);
}
if (tuplesort_gettupleslot(so->sortstate, true, false, so->slot, NULL))
{
ItemPointer heaptid = (ItemPointer) DatumGetPointer(slot_getattr(so->slot, 2, &so->isnull));
BlockNumber indexblkno = DatumGetInt32(slot_getattr(so->slot, 3, &so->isnull));
#if PG_VERSION_NUM >= 120000
scan->xs_heaptid = *heaptid;
@@ -388,21 +327,6 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
scan->xs_ctup.t_self = *heaptid;
#endif
/* Keep track of info needed to mark tuple as dead */
so->heaptid = *heaptid;
/* Unpin buffer */
if (BufferIsValid(so->buf))
ReleaseBuffer(so->buf);
/*
* An index scan must maintain a pin on the index page holding the
* item last returned by amgettuple
*
* https://www.postgresql.org/docs/current/index-locking.html
*/
so->buf = ReadBuffer(scan->indexRelation, indexblkno);
scan->xs_recheckorderby = false;
return true;
}
@@ -418,10 +342,6 @@ ivfflatendscan(IndexScanDesc scan)
{
IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
/* Release pin */
if (BufferIsValid(so->buf))
ReleaseBuffer(so->buf);
pairingheap_free(so->listQueue);
tuplesort_end(so->sortstate);

113
test/t/019_hnsw_array.pl Normal file
View File

@@ -0,0 +1,113 @@
use strict;
use warnings;
use PostgresNode;
use TestLib;
use Test::More;
my $node;
my @queries = ();
my @expected;
my $limit = 20;
sub test_recall
{
my ($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 ORDER BY v $operator '$queries[0]' LIMIT $limit;
));
like($explain, qr/Index Scan/);
for my $i (0 .. $#queries)
{
my $actual = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SELECT i FROM tst ORDER BY v $operator '$queries[$i]' LIMIT $limit;
));
my @actual_ids = split("\n", $actual);
my %actual_set = map { $_ => 1 } @actual_ids;
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 = get_new_node('node');
$node->init;
$node->start;
# Create table
$node->safe_psql("postgres", "CREATE EXTENSION vector;");
$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v float4[3]);");
$node->safe_psql("postgres",
"INSERT INTO tst SELECT i, ARRAY[random(), random(), random()] FROM generate_series(1, 10000) i;"
);
$node->safe_psql("postgres", qq(
CREATE FUNCTION float4_l2_distance(float4[], float4[]) RETURNS float8
AS 'BEGIN RETURN l2_distance(\$1::vector, \$2::vector); END;'
LANGUAGE plpgsql IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION float4_l2_squared_distance(float4[], float4[]) RETURNS float8
AS 'BEGIN RETURN vector_l2_squared_distance(\$1::vector, \$2::vector); END;'
LANGUAGE plpgsql IMMUTABLE STRICT PARALLEL SAFE;
CREATE OPERATOR <-> (
LEFTARG = float4[], RIGHTARG = float4[], PROCEDURE = float4_l2_distance,
COMMUTATOR = '<->'
);
CREATE OPERATOR CLASS float4_l2_ops
FOR TYPE float4[] USING hnsw AS
OPERATOR 1 <-> (float4[], float4[]) FOR ORDER BY float_ops,
FUNCTION 1 float4_l2_squared_distance(float4[], float4[]);
));
# Generate queries
for (1 .. 20)
{
my $r1 = rand();
my $r2 = rand();
my $r3 = rand();
push(@queries, "{$r1,$r2,$r3}");
}
# Check each index type
my @operators = ("<->");
my @opclasses = ("float4_l2_ops");
for my $i (0 .. $#operators)
{
my $operator = $operators[$i];
my $opclass = $opclasses[$i];
# Get exact results
@expected = ();
foreach (@queries)
{
my $res = $node->safe_psql("postgres", "SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;");
push(@expected, $res);
}
# Add index
$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v $opclass);");
my $min = $operator eq "<#>" ? 0.80 : 0.99;
test_recall($min, $operator);
}
done_testing();

View File

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