Compare commits

..

56 Commits

Author SHA1 Message Date
Andrew Kane
5934aeb30d Skip more work 2023-10-05 10:37:57 -07:00
Andrew Kane
115c2bd91d Fixed alignment [skip ci] 2023-10-04 23:09:31 -07:00
Andrew Kane
97ac01773d Speed up HNSW index build 2023-10-04 19:43:25 -07:00
Andrew Kane
8d7abb6590 Revert "Fixed locking for index scans for HNSW - #256"
This reverts commit d032726976.
2023-09-26 23:00:14 -07:00
jeff-davis
b247b688a8 No need to MarkBufferDirty(); GenericXLogFinish() does that. (#265) 2023-09-15 13:14:10 -07:00
Andrew Kane
9672446a4c Updated order [skip ci] 2023-09-12 19:53:12 -07:00
Andrew Kane
334614b7f7 Added HnswFreeNeighbors function [skip ci] 2023-09-12 19:49:31 -07:00
Andrew Kane
643eacd9dc Improved variable name [skip ci] 2023-09-12 19:37:42 -07:00
Andrew Kane
bca50a03fa Use consistent variable name 2023-09-12 19:24:31 -07:00
Andrew Kane
d87833cacc Improved variable scoping [skip ci] 2023-09-12 19:16:55 -07:00
Andrew Kane
4c01073ac1 Improved variable scoping [skip ci] 2023-09-12 18:43:43 -07:00
Andrew Kane
6fed8f1e78 Improved types and scoping for k-means [skip ci] 2023-09-12 18:37:56 -07:00
Andrew Kane
611f5b1047 Improved variable scoping [skip ci] 2023-09-12 18:35:53 -07:00
Andrew Kane
e367155732 Improved types and scoping for k-means 2023-09-12 18:33:31 -07:00
Andrew Kane
466c556b1a Improved variable scoping [skip ci] 2023-09-12 18:24:46 -07:00
Andrew Kane
75e09265d6 Improved variable scoping [skip ci] 2023-09-12 18:14:20 -07:00
Andrew Kane
77c5070fb6 Improved variable scoping [skip ci] 2023-09-12 16:50:51 -07:00
Andrew Kane
1006fdf3f2 Improved variable scoping [skip ci] 2023-09-12 16:44:09 -07:00
Andrew Kane
4209c9b3af Improved variable scoping 2023-09-12 16:13:10 -07:00
Andrew Kane
ec0bb4e4ff Improved code 2023-09-12 15:43:28 -07:00
Andrew Kane
b164833933 Removed pinning for HNSW index scan 2023-09-11 12:12:28 -07:00
Andrew Kane
30fb4dd602 Updated comments [skip ci] 2023-09-07 15:29:54 -07:00
Andrew Kane
d032726976 Fixed locking for index scans for HNSW - #256 2023-09-07 15:27:26 -07:00
Andrew Kane
8fa9001474 Improved Makefiles 2023-09-05 16:43:23 -07:00
Andrew Kane
3431acef94 Improved variable names 2023-09-03 16:49:24 -07:00
Andrew Kane
41bdf24cb7 Fixed flaky test 2023-09-03 15:24:40 -07:00
Andrew Kane
3469a0e74c Simplified test [skip ci] 2023-09-03 15:18:23 -07:00
Andrew Kane
0fe43ca675 Added test for marking tuples as dead 2023-09-03 15:15:02 -07:00
Andrew Kane
bbbe1db72b Improved performance of index scans for IVFFlat after updates and deletes 2023-09-03 14:21:57 -07:00
Andrew Kane
bab5fea9e7 Improved variable name [skip ci] 2023-09-03 12:32:59 -07:00
Andrew Kane
b1f9519689 Get info from metapage to determine cost 2023-09-03 12:31:01 -07:00
Andrew Kane
4571fccc60 Fixed previous commit 2023-09-03 10:51:56 -07:00
Andrew Kane
db747e5aa0 Get lists from metapage 2023-09-03 10:34:44 -07:00
Andrew Kane
2179414c05 Updated extension comment [skip ci] 2023-09-03 03:08:35 -07:00
Andrew Kane
8426ee61d2 Improved upgrading instructions [skip ci] 2023-09-03 03:02:10 -07:00
Andrew Kane
c98c4e13aa Added query for checking version to readme [skip ci] 2023-09-03 02:57:56 -07:00
Andrew Kane
04312f6638 Simplified HNSW vacuum logic 2023-09-03 02:25:19 -07:00
Andrew Kane
72ea3c1210 Added GetScanValue function 2023-09-03 01:59:27 -07:00
Andrew Kane
b0801b8833 Fixed vacuum from previous commit 2023-09-03 01:58:45 -07:00
Andrew Kane
d05d6ee83d Get m from metapage 2023-09-03 01:35:21 -07:00
Andrew Kane
4022bb66a0 Improved variable scoping 2023-09-02 21:14:31 -07:00
Andrew Kane
034d4acaea Removed comment [skip ci] 2023-09-02 18:23:08 -07:00
Andrew Kane
01f58e470a Revert "Use int64 for wlen"
This reverts commit dbef8d1ad1.
2023-09-02 18:09:10 -07:00
Andrew Kane
dbef8d1ad1 Use int64 for wlen 2023-09-02 17:55:48 -07:00
Andrew Kane
5c005cf57c Revert "No need to increment wlen when removing"
This reverts commit 6b2e215447.
2023-09-02 17:41:31 -07:00
Andrew Kane
5665a11a05 Updated comment [skip ci] 2023-09-02 15:34:33 -07:00
Andrew Kane
6b2e215447 No need to increment wlen when removing 2023-09-02 15:33:40 -07:00
Andrew Kane
0d86191eaf Improved test for cosine distance [skip ci] 2023-09-01 19:59:21 -07:00
Andrew Kane
cf9f7aeea9 Added another test for cosine distance [skip ci] 2023-09-01 19:57:28 -07:00
Andrew Kane
0b0e542ce6 Fixed auto-vectorization for vector_spherical_distance with MSVC 2023-09-01 18:42:37 -07:00
Andrew Kane
a4590d2d9d Simplified WAL tests [skip ci] 2023-09-01 15:49:52 -07:00
Andrew Kane
9ebec1529b Updated comments [skip ci] 2023-09-01 00:35:06 -07:00
Andrew Kane
77ff4c18f0 Updated comments [skip ci] 2023-09-01 00:32:42 -07:00
Andrew Kane
88dabaa41c Added test for IVFFlat insert recall 2023-09-01 00:30:02 -07:00
Andrew Kane
1809ffa52b Renamed test [skip ci] 2023-09-01 00:15:07 -07:00
Andrew Kane
024f283ee8 Updated header order [skip ci] 2023-09-01 00:14:03 -07:00
30 changed files with 529 additions and 682 deletions

View File

@@ -1,3 +1,7 @@
## 0.5.1 (unreleased)
- Improved performance of index scans for IVFFlat after updates and deletes
## 0.5.0 (2023-08-28)
- Added HNSW index type

View File

@@ -8,7 +8,7 @@ HEADERS = src/vector.h
TESTS = $(wildcard test/sql/*.sql)
REGRESS = $(patsubst test/sql/%.sql,%,$(TESTS))
REGRESS_OPTS = --inputdir=test --load-extension=vector
REGRESS_OPTS = --inputdir=test --load-extension=$(EXTENSION)
OPTFLAGS = -march=native

View File

@@ -5,7 +5,7 @@ OBJS = src\hnsw.obj src\hnswbuild.obj src\hnswinsert.obj src\hnswscan.obj src\hn
HEADERS = src\vector.h
REGRESS = btree cast copy functions input ivfflat_cosine ivfflat_ip ivfflat_l2 ivfflat_options ivfflat_unlogged
REGRESS_OPTS = --inputdir=test --load-extension=vector
REGRESS_OPTS = --inputdir=test --load-extension=$(EXTENSION)
# For /arch flags
# https://learn.microsoft.com/en-us/cpp/build/reference/arch-minimum-cpu-architecture

View File

@@ -595,12 +595,18 @@ pgvector is available on [these providers](https://github.com/pgvector/pgvector/
## Upgrading
Install the latest version and run:
Install the latest version. Then in each database you want to upgrade, run:
```sql
ALTER EXTENSION vector UPDATE;
```
You can check the version in the current database with:
```sql
SELECT extversion FROM pg_extension WHERE extname = 'vector';
```
## Upgrade Notes
### 0.4.0

View File

@@ -91,7 +91,7 @@ hnswcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
MemSet(&costs, 0, sizeof(costs));
index = index_open(path->indexinfo->indexoid, NoLock);
m = HnswGetM(index);
HnswGetMetaPageInfo(index, &m, NULL);
index_close(index, NoLock);
/* Approximate entry level */
@@ -196,7 +196,7 @@ hnswhandler(PG_FUNCTION_ARGS)
amroutine->aminsert = hnswinsert;
amroutine->ambulkdelete = hnswbulkdelete;
amroutine->amvacuumcleanup = hnswvacuumcleanup;
amroutine->amcanreturn = NULL; /* tuple not included in heapsort */
amroutine->amcanreturn = NULL;
amroutine->amcostestimate = hnswcostestimate;
amroutine->amoptions = hnswoptions;
amroutine->amproperty = NULL; /* TODO AMPROP_DISTANCE_ORDERABLE */

View File

@@ -4,7 +4,6 @@
#include "postgres.h"
#include "access/generic_xlog.h"
#include "access/parallel.h"
#include "access/reloptions.h"
#include "nodes/execnodes.h"
#include "port.h" /* for random() */
@@ -15,10 +14,6 @@
#error "Requires PostgreSQL 11+"
#endif
#if PG_VERSION_NUM < 120000
#include "access/relscan.h"
#endif
#define HNSW_MAX_DIM 2000
/* Support functions */
@@ -115,11 +110,13 @@ typedef struct HnswCandidate
{
HnswElement element;
float distance;
bool closer;
} HnswCandidate;
typedef struct HnswNeighborArray
{
int length;
bool closerSet;
HnswCandidate *items;
} HnswNeighborArray;
@@ -137,49 +134,6 @@ typedef struct HnswOptions
int efConstruction; /* size of dynamic candidate list */
} HnswOptions;
typedef struct HnswSpool
{
Relation heap;
Relation index;
} HnswSpool;
typedef struct HnswShared
{
/* Immutable state */
Oid heaprelid;
Oid indexrelid;
bool isconcurrent;
int scantuplesortstates;
/* Worker progress */
ConditionVariable workersdonecv;
/* Mutex for mutable state */
slock_t mutex;
/* Mutable state */
int nparticipantsdone;
double reltuples;
double indtuples;
#if PG_VERSION_NUM < 120000
ParallelHeapScanDescData heapdesc; /* must come last */
#endif
} HnswShared;
#if PG_VERSION_NUM >= 120000
#define ParallelTableScanFromHnswShared(shared) \
(ParallelTableScanDesc) ((char *) (shared) + BUFFERALIGN(sizeof(HnswShared)))
#endif
typedef struct HnswLeader
{
ParallelContext *pcxt;
int nparticipanttuplesorts;
HnswShared *hnswshared;
Snapshot snapshot;
} HnswLeader;
typedef struct HnswBuildState
{
/* Info */
@@ -213,9 +167,6 @@ typedef struct HnswBuildState
/* Memory */
MemoryContext tmpCtx;
/* Parallel builds */
HnswLeader *hnswleader;
} HnswBuildState;
typedef struct HnswMetaPageData
@@ -269,7 +220,6 @@ typedef HnswNeighborTupleData * HnswNeighborTuple;
typedef struct HnswScanOpaqueData
{
bool first;
Buffer buf;
List *w;
MemoryContext tmpCtx;
@@ -310,15 +260,16 @@ typedef struct HnswVacuumState
/* Methods */
int HnswGetM(Relation index);
int HnswGetEfConstruction(Relation index);
FmgrInfo *HnswOptionalProcInfo(Relation rel, uint16 procnum);
FmgrInfo *HnswOptionalProcInfo(Relation index, uint16 procnum);
bool HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
void HnswCommitBuffer(Buffer buf, GenericXLogState *state);
Buffer HnswNewBuffer(Relation index, ForkNumber forkNum);
void HnswInitPage(Buffer buf, Page page);
void HnswInitRegisterPage(Relation index, Buffer *buf, Page *page, GenericXLogState **state);
void HnswInit(void);
List *HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, bool inserting, HnswElement skipElement);
List *HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement);
HnswElement HnswGetEntryPoint(Relation index);
void HnswGetMetaPageInfo(Relation index, int *m, HnswElement * entryPoint);
HnswElement HnswInitElement(ItemPointer tid, int m, double ml, int maxLevel);
void HnswFreeElement(HnswElement element);
HnswElement HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno);
@@ -335,8 +286,7 @@ void HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool
void HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index, FmgrInfo *procinfo, Oid collation, bool loadVec);
void HnswSetElementTuple(HnswElementTuple etup, HnswElement element);
void HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int *updateIdx, Relation index, FmgrInfo *procinfo, Oid collation);
void HnswLoadNeighbors(HnswElement element, Relation index);
PGDLLEXPORT void HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc);
void HnswLoadNeighbors(HnswElement element, Relation index, int m);
/* Index access methods */
IndexBuildResult *hnswbuild(Relation heap, Relation index, IndexInfo *indexInfo);

View File

@@ -2,15 +2,12 @@
#include <math.h>
#include "access/parallel.h"
#include "access/xact.h"
#include "catalog/index.h"
#include "hnsw.h"
#include "miscadmin.h"
#include "lib/pairingheap.h"
#include "nodes/pg_list.h"
#include "storage/bufmgr.h"
#include "tcop/tcopprot.h"
#include "utils/memutils.h"
#if PG_VERSION_NUM >= 140000
@@ -38,23 +35,6 @@
#define UpdateProgress(index, val) ((void)val)
#endif
#if PG_VERSION_NUM >= 140000
#include "utils/backend_status.h"
#include "utils/wait_event.h"
#endif
#if PG_VERSION_NUM >= 120000
#include "access/table.h"
#include "optimizer/optimizer.h"
#else
#include "access/heapam.h"
#include "optimizer/planner.h"
#include "pgstat.h"
#endif
#define PARALLEL_KEY_HNSW_SHARED UINT64CONST(0xA000000000000001)
#define PARALLEL_KEY_QUERY_TEXT UINT64CONST(0xA000000000000002)
/*
* Create the metapage
*/
@@ -101,7 +81,6 @@ HnswBuildAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **
HnswPageGetOpaque(*page)->nextblkno = BufferGetBlockNumber(newbuf);
/* Commit */
MarkBufferDirty(*buf);
GenericXLogFinish(*state);
UnlockReleaseBuffer(*buf);
@@ -199,7 +178,6 @@ CreateElementPages(HnswBuildState * buildstate)
insertPage = BufferGetBlockNumber(buf);
/* Commit */
MarkBufferDirty(buf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
@@ -247,7 +225,6 @@ CreateNeighborPages(HnswBuildState * buildstate)
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
/* Commit */
MarkBufferDirty(buf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
}
@@ -371,7 +348,6 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
/* TODO Fix progress for parallel builds */
if (HnswInsertTuple(buildstate->index, values, isnull, tid, buildstate->heap))
UpdateProgress(PROGRESS_CREATEIDX_TUPLES_DONE, ++buildstate->indtuples);
@@ -469,8 +445,6 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
buildstate->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
"Hnsw build temporary context",
ALLOCSET_DEFAULT_SIZES);
buildstate->hnswleader = NULL;
}
/*
@@ -483,373 +457,21 @@ FreeBuildState(HnswBuildState * buildstate)
MemoryContextDelete(buildstate->tmpCtx);
}
/*
* Within leader, wait for end of heap scan
*/
static double
ParallelHeapScan(HnswBuildState * buildstate)
{
HnswShared *hnswshared = buildstate->hnswleader->hnswshared;
int nparticipanttuplesorts;
double reltuples;
nparticipanttuplesorts = buildstate->hnswleader->nparticipanttuplesorts;
for (;;)
{
SpinLockAcquire(&hnswshared->mutex);
if (hnswshared->nparticipantsdone == nparticipanttuplesorts)
{
buildstate->indtuples = hnswshared->indtuples;
reltuples = hnswshared->reltuples;
SpinLockRelease(&hnswshared->mutex);
break;
}
SpinLockRelease(&hnswshared->mutex);
ConditionVariableSleep(&hnswshared->workersdonecv,
WAIT_EVENT_PARALLEL_CREATE_INDEX_SCAN);
}
ConditionVariableCancelSleep();
return reltuples;
}
/*
* Perform a worker's portion of a parallel insert
*/
static void
HnswParallelScanAndInsert(HnswSpool * hnswspool, HnswShared * hnswshared, bool progress)
{
HnswBuildState buildstate;
#if PG_VERSION_NUM >= 120000
TableScanDesc scan;
#else
HeapScanDesc scan;
#endif
double reltuples;
IndexInfo *indexInfo;
/* Join parallel scan */
indexInfo = BuildIndexInfo(hnswspool->index);
indexInfo->ii_Concurrent = hnswshared->isconcurrent;
InitBuildState(&buildstate, hnswspool->heap, hnswspool->index, indexInfo, MAIN_FORKNUM);
/* TODO Support in-memory builds */
buildstate.maxInMemoryElements = 0;
buildstate.flushed = true;
#if PG_VERSION_NUM >= 120000
scan = table_beginscan_parallel(hnswspool->heap,
ParallelTableScanFromHnswShared(hnswshared));
reltuples = table_index_build_scan(hnswspool->heap, hnswspool->index, indexInfo,
true, progress, BuildCallback,
(void *) &buildstate, scan);
#else
scan = heap_beginscan_parallel(hnswspool->heap, &hnswshared->heapdesc);
reltuples = IndexBuildHeapScan(hnswspool->heap, hnswspool->index, indexInfo,
true, BuildCallback,
(void *) &buildstate, scan);
#endif
/* Record statistics */
SpinLockAcquire(&hnswshared->mutex);
hnswshared->nparticipantsdone++;
hnswshared->reltuples += reltuples;
hnswshared->indtuples += buildstate.indtuples;
SpinLockRelease(&hnswshared->mutex);
/* Log statistics */
if (progress)
ereport(DEBUG1, (errmsg("leader processed " INT64_FORMAT " tuples", (int64) reltuples)));
else
ereport(DEBUG1, (errmsg("worker processed " INT64_FORMAT " tuples", (int64) reltuples)));
/* Notify leader */
ConditionVariableSignal(&hnswshared->workersdonecv);
FreeBuildState(&buildstate);
}
/*
* Perform work within a launched parallel process
*/
void
HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc)
{
char *sharedquery;
HnswSpool *hnswspool;
HnswShared *hnswshared;
Relation heapRel;
Relation indexRel;
LOCKMODE heapLockmode;
LOCKMODE indexLockmode;
/* Set debug_query_string for individual workers first */
sharedquery = shm_toc_lookup(toc, PARALLEL_KEY_QUERY_TEXT, true);
debug_query_string = sharedquery;
/* Report the query string from leader */
pgstat_report_activity(STATE_RUNNING, debug_query_string);
/* Look up shared state */
hnswshared = shm_toc_lookup(toc, PARALLEL_KEY_HNSW_SHARED, false);
/* Open relations using lock modes known to be obtained by index.c */
if (!hnswshared->isconcurrent)
{
heapLockmode = ShareLock;
indexLockmode = AccessExclusiveLock;
}
else
{
heapLockmode = ShareUpdateExclusiveLock;
indexLockmode = RowExclusiveLock;
}
/* Open relations within worker */
#if PG_VERSION_NUM >= 120000
heapRel = table_open(hnswshared->heaprelid, heapLockmode);
#else
heapRel = heap_open(hnswshared->heaprelid, heapLockmode);
#endif
indexRel = index_open(hnswshared->indexrelid, indexLockmode);
/* Initialize worker's own spool */
hnswspool = (HnswSpool *) palloc0(sizeof(HnswSpool));
hnswspool->heap = heapRel;
hnswspool->index = indexRel;
/* Perform inserts */
HnswParallelScanAndInsert(hnswspool, hnswshared, false);
/* Close relations within worker */
index_close(indexRel, indexLockmode);
#if PG_VERSION_NUM >= 120000
table_close(heapRel, heapLockmode);
#else
heap_close(heapRel, heapLockmode);
#endif
}
/*
* End parallel build
*/
static void
HnswEndParallel(HnswLeader * hnswleader)
{
/* Shutdown worker processes */
WaitForParallelWorkersToFinish(hnswleader->pcxt);
/* Free last reference to MVCC snapshot, if one was used */
if (IsMVCCSnapshot(hnswleader->snapshot))
UnregisterSnapshot(hnswleader->snapshot);
DestroyParallelContext(hnswleader->pcxt);
ExitParallelMode();
}
/*
* Return size of shared memory required for parallel index build
*/
static Size
ParallelEstimateShared(Relation heap, Snapshot snapshot)
{
#if PG_VERSION_NUM >= 120000
return add_size(BUFFERALIGN(sizeof(HnswShared)), table_parallelscan_estimate(heap, snapshot));
#else
if (!IsMVCCSnapshot(snapshot))
{
Assert(snapshot == SnapshotAny);
return sizeof(HnswShared);
}
return add_size(offsetof(HnswShared, heapdesc) +
offsetof(ParallelHeapScanDescData, phs_snapshot_data),
EstimateSnapshotSpace(snapshot));
#endif
}
/*
* Within leader, participate as a parallel worker
*/
static void
HnswLeaderParticipateAsWorker(HnswBuildState * buildstate)
{
HnswLeader *hnswleader = buildstate->hnswleader;
HnswSpool *leaderworker;
/* Allocate memory and initialize private spool */
leaderworker = (HnswSpool *) palloc0(sizeof(HnswSpool));
leaderworker->heap = buildstate->heap;
leaderworker->index = buildstate->index;
/* Perform work common to all participants */
HnswParallelScanAndInsert(leaderworker, hnswleader->hnswshared, true);
}
/*
* Begin parallel build
*/
static void
HnswBeginParallel(HnswBuildState * buildstate, bool isconcurrent, int request)
{
ParallelContext *pcxt;
int scantuplesortstates;
Snapshot snapshot;
Size esthnswshared;
HnswShared *hnswshared;
HnswLeader *hnswleader = (HnswLeader *) palloc0(sizeof(HnswLeader));
bool leaderparticipates = true;
int querylen;
#ifdef DISABLE_LEADER_PARTICIPATION
leaderparticipates = false;
#endif
/* Enter parallel mode and create context */
EnterParallelMode();
Assert(request > 0);
#if PG_VERSION_NUM >= 120000
pcxt = CreateParallelContext("vector", "HnswParallelBuildMain", request);
#else
pcxt = CreateParallelContext("vector", "HnswParallelBuildMain", request, true);
#endif
scantuplesortstates = leaderparticipates ? request + 1 : request;
/* Get snapshot for table scan */
if (!isconcurrent)
snapshot = SnapshotAny;
else
snapshot = RegisterSnapshot(GetTransactionSnapshot());
/* Estimate size of workspaces */
esthnswshared = ParallelEstimateShared(buildstate->heap, snapshot);
shm_toc_estimate_chunk(&pcxt->estimator, esthnswshared);
shm_toc_estimate_keys(&pcxt->estimator, 1);
/* Finally, estimate PARALLEL_KEY_QUERY_TEXT space */
if (debug_query_string)
{
querylen = strlen(debug_query_string);
shm_toc_estimate_chunk(&pcxt->estimator, querylen + 1);
shm_toc_estimate_keys(&pcxt->estimator, 1);
}
else
querylen = 0; /* keep compiler quiet */
/* Everyone's had a chance to ask for space, so now create the DSM */
InitializeParallelDSM(pcxt);
/* If no DSM segment was available, back out (do serial build) */
if (pcxt->seg == NULL)
{
if (IsMVCCSnapshot(snapshot))
UnregisterSnapshot(snapshot);
DestroyParallelContext(pcxt);
ExitParallelMode();
return;
}
/* Store shared build state, for which we reserved space */
hnswshared = (HnswShared *) shm_toc_allocate(pcxt->toc, esthnswshared);
/* Initialize immutable state */
hnswshared->heaprelid = RelationGetRelid(buildstate->heap);
hnswshared->indexrelid = RelationGetRelid(buildstate->index);
hnswshared->isconcurrent = isconcurrent;
hnswshared->scantuplesortstates = scantuplesortstates;
ConditionVariableInit(&hnswshared->workersdonecv);
SpinLockInit(&hnswshared->mutex);
/* Initialize mutable state */
hnswshared->nparticipantsdone = 0;
hnswshared->reltuples = 0;
hnswshared->indtuples = 0;
#if PG_VERSION_NUM >= 120000
table_parallelscan_initialize(buildstate->heap,
ParallelTableScanFromHnswShared(hnswshared),
snapshot);
#else
heap_parallelscan_initialize(&hnswshared->heapdesc, buildstate->heap, snapshot);
#endif
shm_toc_insert(pcxt->toc, PARALLEL_KEY_HNSW_SHARED, hnswshared);
/* Store query string for workers */
if (debug_query_string)
{
char *sharedquery;
sharedquery = (char *) shm_toc_allocate(pcxt->toc, querylen + 1);
memcpy(sharedquery, debug_query_string, querylen + 1);
shm_toc_insert(pcxt->toc, PARALLEL_KEY_QUERY_TEXT, sharedquery);
}
/* Launch workers, saving status for leader/caller */
LaunchParallelWorkers(pcxt);
hnswleader->pcxt = pcxt;
hnswleader->nparticipanttuplesorts = pcxt->nworkers_launched;
if (leaderparticipates)
hnswleader->nparticipanttuplesorts++;
hnswleader->hnswshared = hnswshared;
hnswleader->snapshot = snapshot;
/* If no workers were successfully launched, back out (do serial build) */
if (pcxt->nworkers_launched == 0)
{
HnswEndParallel(hnswleader);
return;
}
/* Log participants */
ereport(DEBUG1, (errmsg("using %d parallel workers", pcxt->nworkers_launched)));
/* Save leader state now that it's clear build will be parallel */
buildstate->hnswleader = hnswleader;
/* Join heap scan ourselves */
if (leaderparticipates)
HnswLeaderParticipateAsWorker(buildstate);
/* Wait for all launched workers */
WaitForParallelWorkersToAttach(pcxt);
}
/*
* Build graph
*/
static void
BuildGraph(HnswBuildState * buildstate, ForkNumber forkNum)
{
int parallel_workers = 0;
UpdateProgress(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_HNSW_PHASE_LOAD);
/* Calculate parallel workers */
parallel_workers = plan_create_index_workers(RelationGetRelid(buildstate->heap), RelationGetRelid(buildstate->index));
/* Attempt to launch parallel worker scan when required */
if (parallel_workers > 0)
{
/* TODO Support in-memory builds */
FlushPages(buildstate);
HnswBeginParallel(buildstate, buildstate->indexInfo->ii_Concurrent, parallel_workers);
}
/* Add tuples to sort */
if (buildstate->hnswleader)
buildstate->reltuples = ParallelHeapScan(buildstate);
else
{
#if PG_VERSION_NUM >= 120000
buildstate->reltuples = table_index_build_scan(buildstate->heap, buildstate->index, buildstate->indexInfo,
true, true, BuildCallback, (void *) buildstate, NULL);
buildstate->reltuples = table_index_build_scan(buildstate->heap, buildstate->index, buildstate->indexInfo,
true, true, BuildCallback, (void *) buildstate, NULL);
#else
buildstate->reltuples = IndexBuildHeapScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
true, BuildCallback, (void *) buildstate, NULL);
buildstate->reltuples = IndexBuildHeapScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
true, BuildCallback, (void *) buildstate, NULL);
#endif
}
/* End parallel build */
if (buildstate->hnswleader)
HnswEndParallel(buildstate->hnswleader);
}
/*

View File

@@ -202,8 +202,6 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
HnswInsertAppendPage(index, &newbuf, &newpage, state, page);
/* Commit */
MarkBufferDirty(newbuf);
MarkBufferDirty(buf);
GenericXLogFinish(state);
/* Unlock previous buffer */
@@ -270,9 +268,6 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
}
/* Commit */
MarkBufferDirty(buf);
if (nbuf != buf)
MarkBufferDirty(nbuf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
if (nbuf != buf)
@@ -329,7 +324,7 @@ HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswE
/* Get latest neighbors since they may have changed */
/* Do not lock yet since selecting neighbors can take time */
HnswLoadNeighbors(hc->element, index);
HnswLoadNeighbors(hc->element, index, m);
/*
* Could improve performance for vacuuming by checking neighbors
@@ -391,7 +386,6 @@ HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswE
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
/* Commit */
MarkBufferDirty(buf);
GenericXLogFinish(state);
}
else
@@ -445,7 +439,6 @@ HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup)
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
/* Commit */
MarkBufferDirty(buf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
@@ -492,9 +485,8 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
FmgrInfo *normprocinfo;
HnswElement entryPoint;
HnswElement element;
int m = HnswGetM(index);
int m;
int efConstruction = HnswGetEfConstruction(index);
double ml = HnswGetMl(m);
FmgrInfo *procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
Oid collation = index->rd_indcollation[0];
HnswElement dup;
@@ -511,10 +503,6 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
return false;
}
/* Create an element */
element = HnswInitElement(heap_tid, m, ml, HnswGetMaxLevel(m));
element->vec = DatumGetVector(value);
/*
* 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
@@ -522,8 +510,12 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
*/
LockPage(index, HNSW_UPDATE_LOCK, lockmode);
/* Get entry point */
entryPoint = HnswGetEntryPoint(index);
/* Get m and entry point */
HnswGetMetaPageInfo(index, &m, &entryPoint);
/* Create an element */
element = HnswInitElement(heap_tid, m, HnswGetMl(m), HnswGetMaxLevel(m));
element->vec = DatumGetVector(value);
/* Prevent concurrent inserts when likely updating entry point */
if (entryPoint == NULL || element->level > entryPoint->level)

View File

@@ -19,7 +19,11 @@ GetScanItems(IndexScanDesc scan, Datum q)
Oid collation = so->collation;
List *ep;
List *w;
HnswElement entryPoint = HnswGetEntryPoint(index);
int m;
HnswElement entryPoint;
/* Get m and entry point */
HnswGetMetaPageInfo(index, &m, &entryPoint);
if (entryPoint == NULL)
return NIL;
@@ -28,11 +32,11 @@ GetScanItems(IndexScanDesc scan, Datum q)
for (int lc = entryPoint->level; lc >= 1; lc--)
{
w = HnswSearchLayer(q, ep, 1, lc, index, procinfo, collation, false, NULL);
w = HnswSearchLayer(q, ep, 1, lc, index, procinfo, collation, m, false, NULL);
ep = w;
}
return HnswSearchLayer(q, ep, hnsw_ef_search, 0, index, procinfo, collation, false, NULL);
return HnswSearchLayer(q, ep, hnsw_ef_search, 0, index, procinfo, collation, m, false, NULL);
}
/*
@@ -58,6 +62,33 @@ GetDimensions(Relation index)
return dimensions;
}
/*
* Get scan value
*/
static Datum
GetScanValue(IndexScanDesc scan)
{
HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
Datum value;
if (scan->orderByData->sk_flags & SK_ISNULL)
value = PointerGetDatum(InitVector(GetDimensions(scan->indexRelation)));
else
{
value = scan->orderByData->sk_argument;
/* Value should not be compressed or toasted */
Assert(!VARATT_IS_COMPRESSED(DatumGetPointer(value)));
Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));
/* Fine if normalization fails */
if (so->normprocinfo != NULL)
HnswNormValue(so->normprocinfo, so->collation, &value, NULL);
}
return value;
}
/*
* Prepare for an index scan
*/
@@ -70,7 +101,6 @@ hnswbeginscan(Relation index, int nkeys, int norderbys)
scan = RelationGetIndexScan(index, nkeys, norderbys);
so = (HnswScanOpaque) palloc(sizeof(HnswScanOpaqueData));
so->buf = InvalidBuffer;
so->first = true;
so->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
"Hnsw scan temporary context",
@@ -130,20 +160,8 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
if (scan->orderByData == NULL)
elog(ERROR, "cannot scan hnsw index without order");
if (scan->orderByData->sk_flags & SK_ISNULL)
value = PointerGetDatum(InitVector(GetDimensions(scan->indexRelation)));
else
{
value = scan->orderByData->sk_argument;
/* Value should not be compressed or toasted */
Assert(!VARATT_IS_COMPRESSED(DatumGetPointer(value)));
Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));
/* Fine if normalization fails */
if (so->normprocinfo != NULL)
HnswNormValue(so->normprocinfo, so->collation, &value, NULL);
}
/* Get scan value */
value = GetScanValue(scan);
/*
* Get a shared lock. This allows vacuum to ensure no in-flight scans
@@ -162,40 +180,35 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
while (list_length(so->w) > 0)
{
HnswCandidate *hc = llast(so->w);
ItemPointer tid;
BlockNumber indexblkno;
ItemPointer heaptid;
/* Move to next element if no valid heap tids */
/* Move to next element if no valid heap TIDs */
if (list_length(hc->element->heaptids) == 0)
{
so->w = list_delete_last(so->w);
continue;
}
tid = llast(hc->element->heaptids);
indexblkno = hc->element->blkno;
heaptid = llast(hc->element->heaptids);
hc->element->heaptids = list_delete_last(hc->element->heaptids);
MemoryContextSwitchTo(oldCtx);
#if PG_VERSION_NUM >= 120000
scan->xs_heaptid = *tid;
scan->xs_heaptid = *heaptid;
#else
scan->xs_ctup.t_self = *tid;
scan->xs_ctup.t_self = *heaptid;
#endif
/* 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
* 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
*/
so->buf = ReadBuffer(scan->indexRelation, indexblkno);
scan->xs_recheckorderby = false;
return true;
@@ -213,10 +226,6 @@ hnswendscan(IndexScanDesc scan)
{
HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
/* Release pin */
if (BufferIsValid(so->buf))
ReleaseBuffer(so->buf);
MemoryContextDelete(so->tmpCtx);
pfree(so);

View File

@@ -38,12 +38,12 @@ HnswGetEfConstruction(Relation index)
* Get proc
*/
FmgrInfo *
HnswOptionalProcInfo(Relation rel, uint16 procnum)
HnswOptionalProcInfo(Relation index, uint16 procnum)
{
if (!OidIsValid(index_getprocid(rel, 1, procnum)))
if (!OidIsValid(index_getprocid(index, 1, procnum)))
return NULL;
return index_getprocinfo(rel, 1, procnum);
return index_getprocinfo(index, 1, procnum);
}
/*
@@ -117,7 +117,6 @@ HnswInitRegisterPage(Relation index, Buffer *buf, Page *page, GenericXLogState *
void
HnswCommitBuffer(Buffer buf, GenericXLogState *state)
{
MarkBufferDirty(buf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
}
@@ -140,9 +139,21 @@ HnswInitNeighbors(HnswElement element, int m)
a = &element->neighbors[lc];
a->length = 0;
a->items = palloc(sizeof(HnswCandidate) * lm);
a->closerSet = false;
}
}
/*
* Free neighbors
*/
static void
HnswFreeNeighbors(HnswElement element)
{
for (int lc = 0; lc <= element->level; lc++)
pfree(element->neighbors[lc].items);
pfree(element->neighbors);
}
/*
* Allocate an element
*/
@@ -174,10 +185,8 @@ HnswInitElement(ItemPointer heaptid, int m, double ml, int maxLevel)
void
HnswFreeElement(HnswElement element)
{
HnswFreeNeighbors(element);
list_free_deep(element->heaptids);
for (int lc = 0; lc <= element->level; lc++)
pfree(element->neighbors[lc].items);
pfree(element->neighbors);
pfree(element->vec);
pfree(element);
}
@@ -210,25 +219,43 @@ HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno)
}
/*
* Get the entry point
* Get the metapage info
*/
HnswElement
HnswGetEntryPoint(Relation index)
void
HnswGetMetaPageInfo(Relation index, int *m, HnswElement * entryPoint)
{
Buffer buf;
Page page;
HnswMetaPage metap;
HnswElement entryPoint = NULL;
buf = ReadBuffer(index, HNSW_METAPAGE_BLKNO);
LockBuffer(buf, BUFFER_LOCK_SHARE);
page = BufferGetPage(buf);
metap = HnswPageGetMeta(page);
if (BlockNumberIsValid(metap->entryBlkno))
entryPoint = HnswInitElementFromBlock(metap->entryBlkno, metap->entryOffno);
if (m != NULL)
*m = metap->m;
if (entryPoint != NULL)
{
if (BlockNumberIsValid(metap->entryBlkno))
*entryPoint = HnswInitElementFromBlock(metap->entryBlkno, metap->entryOffno);
else
*entryPoint = NULL;
}
UnlockReleaseBuffer(buf);
}
/*
* Get the entry point
*/
HnswElement
HnswGetEntryPoint(Relation index)
{
HnswElement entryPoint;
HnswGetMetaPageInfo(index, NULL, &entryPoint);
return entryPoint;
}
@@ -337,10 +364,9 @@ HnswSetNeighborTuple(HnswNeighborTuple ntup, HnswElement e, int m)
* Load neighbors from page
*/
static void
LoadNeighborsFromPage(HnswElement element, Relation index, Page page)
LoadNeighborsFromPage(HnswElement element, Relation index, Page page, int m)
{
HnswNeighborTuple ntup = (HnswNeighborTuple) PageGetItem(page, PageGetItemId(page, element->neighborOffno));
int m = HnswGetM(index);
int neighborCount = (element->level + 2) * m;
Assert(HnswIsNeighborTuple(ntup));
@@ -381,7 +407,7 @@ LoadNeighborsFromPage(HnswElement element, Relation index, Page page)
* Load neighbors
*/
void
HnswLoadNeighbors(HnswElement element, Relation index)
HnswLoadNeighbors(HnswElement element, Relation index, int m)
{
Buffer buf;
Page page;
@@ -390,7 +416,7 @@ HnswLoadNeighbors(HnswElement element, Relation index)
LockBuffer(buf, BUFFER_LOCK_SHARE);
page = BufferGetPage(buf);
LoadNeighborsFromPage(element, index, page);
LoadNeighborsFromPage(element, index, page, m);
UnlockReleaseBuffer(buf);
}
@@ -543,7 +569,7 @@ AddToVisited(HTAB *v, HnswCandidate * hc, Relation index, bool *found)
* Algorithm 2 from paper
*/
List *
HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, bool inserting, HnswElement skipElement)
HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement)
{
ListCell *lc2;
@@ -598,7 +624,7 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
break;
if (c->element->neighbors == NULL)
HnswLoadNeighbors(c->element, index);
HnswLoadNeighbors(c->element, index, m);
/* Get the neighborhood at layer lc */
neighborhood = &c->element->neighbors[lc];
@@ -723,11 +749,13 @@ 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)
{
List *r = NIL;
List *w = list_copy(c);
pairingheap *wd;
bool mustCalculate = !e2->neighbors[lc].closerSet;
bool foundNew = false;
if (list_length(w) <= m)
return w;
@@ -738,18 +766,40 @@ SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswC
{
/* 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 (foundNew)
{
/* If new or current candidate is not closer, no change in state */
if (newCandidate->closer && e->closer)
{
/* Only need to compare with new candidate */
float distance = HnswGetDistance(e->element, newCandidate->element, lc, procinfo, collation);
if (closer)
e->closer = e->distance < distance;
if (!e->closer)
mustCalculate = true;
}
}
else if (e == newCandidate)
{
e->closer = CheckElementCloser(e, r, lc, procinfo, collation);
foundNew = true;
}
if (e->closer)
r = lappend(r, e);
else
pairingheap_add(wd, &(CreatePairingHeapNode(e)->ph_node));
}
e2->neighbors[lc].closerSet = true;
/* Keep pruned connections */
while (!pairingheap_is_empty(wd) && list_length(r) < m)
r = lappend(r, ((HnswPairingHeapNode *) pairingheap_remove_first(wd))->inner);
@@ -884,7 +934,7 @@ HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int
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);
/* Should not happen */
if (pruned == NULL)
@@ -956,7 +1006,7 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
/* 1st phase: greedy search to insert level */
for (int lc = entryLevel; lc >= level + 1; lc--)
{
w = HnswSearchLayer(q, ep, 1, lc, index, procinfo, collation, true, skipElement);
w = HnswSearchLayer(q, ep, 1, lc, index, procinfo, collation, m, true, skipElement);
ep = w;
}
@@ -974,7 +1024,7 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
List *neighbors;
List *lw;
w = HnswSearchLayer(q, ep, efConstruction, lc, index, procinfo, collation, true, skipElement);
w = HnswSearchLayer(q, ep, efConstruction, lc, index, procinfo, collation, m, true, skipElement);
/* Elements being deleted or skipped can help with search */
/* but should be removed before selecting neighbors */
@@ -983,7 +1033,7 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
else
lw = w;
neighbors = SelectNeighbors(lw, lm, lc, procinfo, collation, NULL);
neighbors = SelectNeighbors(lw, lm, lc, procinfo, collation, element, NULL, NULL);
AddConnections(element, neighbors, lm, lc);

View File

@@ -128,10 +128,7 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
blkno = HnswPageGetOpaque(page)->nextblkno;
if (updated)
{
MarkBufferDirty(buf);
GenericXLogFinish(state);
}
else
GenericXLogAbort(state);
@@ -229,7 +226,6 @@ RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswEleme
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
/* Commit */
MarkBufferDirty(buf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
@@ -330,7 +326,10 @@ RepairGraph(HnswVacuumState * vacuumstate)
BufferAccessStrategy bas = vacuumstate->bas;
BlockNumber blkno = HNSW_HEAD_BLKNO;
/* Wait for inserts to complete */
/*
* Wait for inserts to complete. Inserts before this point may have
* neighbors about to be deleted. Inserts after this point will not.
*/
LockPage(index, HNSW_UPDATE_LOCK, ExclusiveLock);
UnlockPage(index, HNSW_UPDATE_LOCK, ExclusiveLock);
@@ -443,7 +442,11 @@ MarkDeleted(HnswVacuumState * vacuumstate)
Relation index = vacuumstate->index;
BufferAccessStrategy bas = vacuumstate->bas;
/* Wait for selects to complete */
/*
* Wait for index scans to complete. Scans before this point may contain
* tuples about to be deleted. Scans after this point will not, since the
* graph has been repaired.
*/
LockPage(index, HNSW_SCAN_LOCK, ExclusiveLock);
UnlockPage(index, HNSW_SCAN_LOCK, ExclusiveLock);
@@ -540,9 +543,6 @@ MarkDeleted(HnswVacuumState * vacuumstate)
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
/* Commit */
MarkBufferDirty(buf);
if (nbuf != buf)
MarkBufferDirty(nbuf);
GenericXLogFinish(state);
if (nbuf != buf)
UnlockReleaseBuffer(nbuf);
@@ -582,7 +582,6 @@ InitVacuumState(HnswVacuumState * vacuumstate, IndexVacuumInfo *info, IndexBulkD
vacuumstate->stats = stats;
vacuumstate->callback = callback;
vacuumstate->callback_state = callback_state;
vacuumstate->m = HnswGetM(index);
vacuumstate->efConstruction = HnswGetEfConstruction(index);
vacuumstate->bas = GetAccessStrategy(BAS_BULKREAD);
vacuumstate->procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
@@ -592,6 +591,9 @@ InitVacuumState(HnswVacuumState * vacuumstate, IndexVacuumInfo *info, IndexBulkD
"Hnsw vacuum temporary context",
ALLOCSET_DEFAULT_SIZES);
/* Get m from metapage */
HnswGetMetaPageInfo(index, &vacuumstate->m, NULL);
/* Create hash table */
hash_ctl.keysize = sizeof(ItemPointerData);
hash_ctl.entrysize = sizeof(ItemPointerData);

View File

@@ -10,8 +10,8 @@
#include "ivfflat.h"
#include "miscadmin.h"
#include "storage/bufmgr.h"
#include "utils/memutils.h"
#include "tcop/tcopprot.h"
#include "utils/memutils.h"
#if PG_VERSION_NUM >= 140000
#include "utils/backend_progress.h"
@@ -506,29 +506,30 @@ CreateListPages(Relation index, VectorArray centers, int dimensions,
Buffer buf;
Page page;
GenericXLogState *state;
OffsetNumber offno;
Size itemsz;
Size listSize;
IvfflatList list;
itemsz = MAXALIGN(IVFFLAT_LIST_SIZE(dimensions));
list = palloc(itemsz);
listSize = MAXALIGN(IVFFLAT_LIST_SIZE(dimensions));
list = palloc(listSize);
buf = IvfflatNewBuffer(index, forkNum);
IvfflatInitRegisterPage(index, &buf, &page, &state);
for (int i = 0; i < lists; i++)
{
OffsetNumber offno;
/* Load list */
list->startPage = InvalidBlockNumber;
list->insertPage = InvalidBlockNumber;
memcpy(&list->center, VectorArrayGet(centers, i), VECTOR_SIZE(dimensions));
/* Ensure free space */
if (PageGetFreeSpace(page) < itemsz)
if (PageGetFreeSpace(page) < listSize)
IvfflatAppendPage(index, &buf, &page, &state, forkNum);
/* Add the item */
offno = PageAddItem(page, (Item) list, itemsz, InvalidOffsetNumber, false, false);
offno = PageAddItem(page, (Item) list, listSize, InvalidOffsetNumber, false, false);
if (offno == InvalidOffsetNumber)
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));

View File

@@ -71,7 +71,7 @@ ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
int lists;
double ratio;
double spc_seq_page_cost;
Relation indexRel;
Relation index;
#if PG_VERSION_NUM < 120000
List *qinfos;
#endif
@@ -89,9 +89,9 @@ ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
MemSet(&costs, 0, sizeof(costs));
indexRel = index_open(path->indexinfo->indexoid, NoLock);
lists = IvfflatGetLists(indexRel);
index_close(indexRel, NoLock);
index = index_open(path->indexinfo->indexoid, NoLock);
IvfflatGetMetaPageInfo(index, &lists, NULL);
index_close(index, NoLock);
/* Get the ratio of lists that we need to visit */
ratio = ((double) ivfflat_probes) / lists;

View File

@@ -244,8 +244,10 @@ typedef struct IvfflatScanList
typedef struct IvfflatScanOpaqueData
{
int probes;
int dimensions;
bool first;
Buffer buf;
ItemPointerData heaptid;
/* Sorting */
Tuplesortstate *sortstate;
@@ -275,9 +277,10 @@ VectorArray VectorArrayInit(int maxlen, int dimensions);
void VectorArrayFree(VectorArray arr);
void PrintVectorArray(char *msg, VectorArray arr);
void IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers);
FmgrInfo *IvfflatOptionalProcInfo(Relation rel, uint16 procnum);
FmgrInfo *IvfflatOptionalProcInfo(Relation index, uint16 procnum);
bool IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
int IvfflatGetLists(Relation index);
void IvfflatGetMetaPageInfo(Relation index, int *lists, int *dimensions);
void IvfflatUpdateList(Relation index, ListInfo listInfo, BlockNumber insertPage, BlockNumber originalInsertPage, BlockNumber startPage, ForkNumber forkNum);
void IvfflatCommitBuffer(Buffer buf, GenericXLogState *state);
void IvfflatAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **state, ForkNumber forkNum);

View File

@@ -11,36 +11,37 @@
* Find the list that minimizes the distance function
*/
static void
FindInsertPage(Relation rel, Datum *values, BlockNumber *insertPage, ListInfo * listInfo)
FindInsertPage(Relation index, Datum *values, BlockNumber *insertPage, ListInfo * listInfo)
{
Buffer cbuf;
Page cpage;
IvfflatList list;
double distance;
double minDistance = DBL_MAX;
BlockNumber nextblkno = IVFFLAT_HEAD_BLKNO;
FmgrInfo *procinfo;
Oid collation;
OffsetNumber offno;
OffsetNumber maxoffno;
/* Avoid compiler warning */
listInfo->blkno = nextblkno;
listInfo->offno = FirstOffsetNumber;
procinfo = index_getprocinfo(rel, 1, IVFFLAT_DISTANCE_PROC);
collation = rel->rd_indcollation[0];
procinfo = index_getprocinfo(index, 1, IVFFLAT_DISTANCE_PROC);
collation = index->rd_indcollation[0];
/* Search all list pages */
while (BlockNumberIsValid(nextblkno))
{
cbuf = ReadBuffer(rel, nextblkno);
Buffer cbuf;
Page cpage;
OffsetNumber maxoffno;
cbuf = ReadBuffer(index, nextblkno);
LockBuffer(cbuf, BUFFER_LOCK_SHARE);
cpage = BufferGetPage(cbuf);
maxoffno = PageGetMaxOffsetNumber(cpage);
for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
for (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
{
IvfflatList list;
double distance;
list = (IvfflatList) PageGetItem(cpage, PageGetItemId(cpage, offno));
distance = DatumGetFloat8(FunctionCall2Coll(procinfo, collation, values[0], PointerGetDatum(&list->center)));
@@ -63,7 +64,7 @@ FindInsertPage(Relation rel, Datum *values, BlockNumber *insertPage, ListInfo *
* Insert a tuple into the index
*/
static void
InsertTuple(Relation rel, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel)
InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel)
{
IndexTuple itup;
Datum value;
@@ -80,20 +81,20 @@ InsertTuple(Relation rel, Datum *values, bool *isnull, ItemPointer heap_tid, Rel
value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
/* Normalize if needed */
normprocinfo = IvfflatOptionalProcInfo(rel, IVFFLAT_NORM_PROC);
normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
if (normprocinfo != NULL)
{
if (!IvfflatNormValue(normprocinfo, rel->rd_indcollation[0], &value, NULL))
if (!IvfflatNormValue(normprocinfo, index->rd_indcollation[0], &value, NULL))
return;
}
/* Find the insert page - sets the page and list info */
FindInsertPage(rel, values, &insertPage, &listInfo);
FindInsertPage(index, values, &insertPage, &listInfo);
Assert(BlockNumberIsValid(insertPage));
originalInsertPage = insertPage;
/* Form tuple */
itup = index_form_tuple(RelationGetDescr(rel), &value, isnull);
itup = index_form_tuple(RelationGetDescr(index), &value, isnull);
itup->t_tid = *heap_tid;
/* Get tuple size */
@@ -103,10 +104,10 @@ InsertTuple(Relation rel, Datum *values, bool *isnull, ItemPointer heap_tid, Rel
/* Find a page to insert the item */
for (;;)
{
buf = ReadBuffer(rel, insertPage);
buf = ReadBuffer(index, insertPage);
LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
state = GenericXLogStart(rel);
state = GenericXLogStart(index);
page = GenericXLogRegisterBuffer(state, buf, 0);
if (PageGetFreeSpace(page) >= itemsz)
@@ -126,9 +127,9 @@ InsertTuple(Relation rel, Datum *values, bool *isnull, ItemPointer heap_tid, Rel
Page newpage;
/* Add a new page */
LockRelationForExtension(rel, ExclusiveLock);
newbuf = IvfflatNewBuffer(rel, MAIN_FORKNUM);
UnlockRelationForExtension(rel, ExclusiveLock);
LockRelationForExtension(index, ExclusiveLock);
newbuf = IvfflatNewBuffer(index, MAIN_FORKNUM);
UnlockRelationForExtension(index, ExclusiveLock);
/* Init new page */
newpage = GenericXLogRegisterBuffer(state, newbuf, GENERIC_XLOG_FULL_IMAGE);
@@ -141,15 +142,13 @@ InsertTuple(Relation rel, Datum *values, bool *isnull, ItemPointer heap_tid, Rel
IvfflatPageGetOpaque(page)->nextblkno = insertPage;
/* Commit */
MarkBufferDirty(newbuf);
MarkBufferDirty(buf);
GenericXLogFinish(state);
/* Unlock previous buffer */
UnlockReleaseBuffer(buf);
/* Prepare new buffer */
state = GenericXLogStart(rel);
state = GenericXLogStart(index);
buf = newbuf;
page = GenericXLogRegisterBuffer(state, buf, 0);
break;
@@ -158,13 +157,13 @@ InsertTuple(Relation rel, Datum *values, bool *isnull, ItemPointer heap_tid, Rel
/* Add to next offset */
if (PageAddItem(page, (Item) itup, itemsz, InvalidOffsetNumber, false, false) == InvalidOffsetNumber)
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(rel));
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
IvfflatCommitBuffer(buf, state);
/* Update the insert page */
if (insertPage != originalInsertPage)
IvfflatUpdateList(rel, listInfo, insertPage, originalInsertPage, InvalidBlockNumber, MAIN_FORKNUM);
IvfflatUpdateList(index, listInfo, insertPage, originalInsertPage, InvalidBlockNumber, MAIN_FORKNUM);
}
/*

View File

@@ -17,10 +17,6 @@ InitCenters(Relation index, VectorArray samples, VectorArray centers, float *low
FmgrInfo *procinfo;
Oid collation;
int64 j;
double distance;
double sum;
double choice;
Vector *vec;
float *weight = palloc(samples->length * sizeof(float));
int numCenters = centers->maxlen;
int numSamples = samples->length;
@@ -33,17 +29,21 @@ InitCenters(Relation index, VectorArray samples, VectorArray centers, float *low
centers->length++;
for (j = 0; j < numSamples; j++)
weight[j] = DBL_MAX;
weight[j] = FLT_MAX;
for (int i = 0; i < numCenters; i++)
{
double sum;
double choice;
CHECK_FOR_INTERRUPTS();
sum = 0.0;
for (j = 0; j < numSamples; j++)
{
vec = VectorArrayGet(samples, j);
Vector *vec = VectorArrayGet(samples, j);
double distance;
/* Only need to compute distance for new center */
/* TODO Use triangle inequality to reduce distance calculations */
@@ -112,7 +112,6 @@ CompareVectors(const void *a, const void *b)
static void
QuickCenters(Relation index, VectorArray samples, VectorArray centers)
{
Vector *vec;
int dimensions = centers->dim;
Oid collation = index->rd_indcollation[0];
FmgrInfo *normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_KMEANS_NORM_PROC);
@@ -123,7 +122,7 @@ QuickCenters(Relation index, VectorArray samples, VectorArray centers)
qsort(samples->items, samples->length, VECTOR_SIZE(samples->dim), CompareVectors);
for (int i = 0; i < samples->length; i++)
{
vec = VectorArrayGet(samples, i);
Vector *vec = VectorArrayGet(samples, i);
if (i == 0 || CompareVectors(vec, VectorArrayGet(samples, i - 1)) != 0)
{
@@ -136,7 +135,7 @@ QuickCenters(Relation index, VectorArray samples, VectorArray centers)
/* Fill remaining with random data */
while (centers->length < centers->maxlen)
{
vec = VectorArrayGet(centers, centers->length);
Vector *vec = VectorArrayGet(centers, centers->length);
SET_VARSIZE(vec, VECTOR_SIZE(dimensions));
vec->dim = dimensions;
@@ -168,7 +167,6 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
Oid collation;
Vector *vec;
Vector *newCenter;
int iteration;
int64 j;
int64 k;
int dimensions = centers->dim;
@@ -182,14 +180,6 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
float *s;
float *halfcdist;
float *newcdist;
int changes;
double minDistance;
int closestCenter;
double distance;
bool rj;
bool rjreset;
double dxcx;
double dxc;
/* Calculate allocation sizes */
Size samplesSize = VECTOR_ARRAY_SIZE(samples->maxlen, samples->dim);
@@ -247,14 +237,14 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
/* Assign each x to its closest initial center c(x) = argmin d(x,c) */
for (j = 0; j < numSamples; j++)
{
minDistance = DBL_MAX;
closestCenter = 0;
float minDistance = FLT_MAX;
int closestCenter = 0;
/* Find closest center */
for (k = 0; k < numCenters; k++)
{
/* TODO Use Lemma 1 in k-means++ initialization */
distance = lowerBound[j * numCenters + k];
float distance = lowerBound[j * numCenters + k];
if (distance < minDistance)
{
@@ -268,13 +258,14 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
}
/* Give 500 iterations to converge */
for (iteration = 0; iteration < 500; iteration++)
for (int iteration = 0; iteration < 500; iteration++)
{
int changes = 0;
bool rjreset;
/* Can take a while, so ensure we can interrupt */
CHECK_FOR_INTERRUPTS();
changes = 0;
/* Step 1: For all centers, compute distance */
for (j = 0; j < numCenters; j++)
{
@@ -282,7 +273,8 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
for (k = j + 1; k < numCenters; k++)
{
distance = 0.5 * DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(vec), PointerGetDatum(VectorArrayGet(centers, k))));
float distance = 0.5 * DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(vec), PointerGetDatum(VectorArrayGet(centers, k))));
halfcdist[j * numCenters + k] = distance;
halfcdist[k * numCenters + j] = distance;
}
@@ -291,10 +283,12 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
/* For all centers c, compute s(c) */
for (j = 0; j < numCenters; j++)
{
minDistance = DBL_MAX;
float minDistance = FLT_MAX;
for (k = 0; k < numCenters; k++)
{
float distance;
if (j == k)
continue;
@@ -310,6 +304,8 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
for (j = 0; j < numSamples; j++)
{
bool rj;
/* Step 2: Identify all points x such that u(x) <= s(c(x)) */
if (upperBound[j] <= s[closestCenters[j]])
continue;
@@ -318,6 +314,8 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
for (k = 0; k < numCenters; k++)
{
float dxcx;
/* Step 3: For all remaining points x and centers c */
if (k == closestCenters[j])
continue;
@@ -347,7 +345,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
/* Step 3b */
if (dxcx > lowerBound[j * numCenters + k] || dxcx > halfcdist[closestCenters[j] * numCenters + k])
{
dxc = DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(vec), PointerGetDatum(VectorArrayGet(centers, k))));
float dxc = DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(vec), PointerGetDatum(VectorArrayGet(centers, k))));
/* d(x,c) calculated */
lowerBound[j * numCenters + k] = dxc;
@@ -361,7 +359,6 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
changes++;
}
}
}
}
@@ -378,6 +375,8 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
for (j = 0; j < numSamples; j++)
{
int closestCenter;
vec = VectorArrayGet(samples, j);
closestCenter = closestCenters[j];
@@ -426,7 +425,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
{
for (k = 0; k < numCenters; k++)
{
distance = lowerBound[j * numCenters + k] - newcdist[k];
float distance = lowerBound[j * numCenters + k] - newcdist[k];
if (distance < 0)
distance = 0;
@@ -442,7 +441,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
/* Step 7 */
for (j = 0; j < numCenters; j++)
memcpy(VectorArrayGet(centers, j), VectorArrayGet(newCenters, j), VECTOR_SIZE(dimensions));
VectorArraySet(centers, j, VectorArrayGet(newCenters, j));
if (changes == 0 && iteration != 0)
break;
@@ -465,9 +464,6 @@ static void
CheckCenters(Relation index, VectorArray centers)
{
FmgrInfo *normprocinfo;
Oid collation;
Vector *vec;
double norm;
if (centers->length != centers->maxlen)
elog(ERROR, "Not enough centers. Please report a bug.");
@@ -475,7 +471,7 @@ CheckCenters(Relation index, VectorArray centers)
/* Ensure no NaN or infinite values */
for (int i = 0; i < centers->length; i++)
{
vec = VectorArrayGet(centers, i);
Vector *vec = VectorArrayGet(centers, i);
for (int j = 0; j < vec->dim; j++)
{
@@ -501,11 +497,12 @@ CheckCenters(Relation index, VectorArray centers)
normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
if (normprocinfo != NULL)
{
collation = index->rd_indcollation[0];
Oid collation = index->rd_indcollation[0];
for (int i = 0; i < centers->length; i++)
{
norm = DatumGetFloat8(FunctionCall1Coll(normprocinfo, collation, PointerGetDatum(VectorArrayGet(centers, i))));
double norm = DatumGetFloat8(FunctionCall1Coll(normprocinfo, collation, PointerGetDatum(VectorArrayGet(centers, i))));
if (norm == 0)
elog(ERROR, "Zero norm detected. Please report a bug.");
}

View File

@@ -31,36 +31,36 @@ CompareLists(const pairingheap_node *a, const pairingheap_node *b, void *arg)
static void
GetScanLists(IndexScanDesc scan, Datum value)
{
Buffer cbuf;
Page cpage;
IvfflatList list;
OffsetNumber offno;
OffsetNumber maxoffno;
IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
BlockNumber nextblkno = IVFFLAT_HEAD_BLKNO;
int listCount = 0;
IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
double distance;
IvfflatScanList *scanlist;
double maxDistance = DBL_MAX;
/* Search all list pages */
while (BlockNumberIsValid(nextblkno))
{
Buffer cbuf;
Page cpage;
OffsetNumber maxoffno;
cbuf = ReadBuffer(scan->indexRelation, nextblkno);
LockBuffer(cbuf, BUFFER_LOCK_SHARE);
cpage = BufferGetPage(cbuf);
maxoffno = PageGetMaxOffsetNumber(cpage);
for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
for (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
{
list = (IvfflatList) PageGetItem(cpage, PageGetItemId(cpage, offno));
IvfflatList list = (IvfflatList) PageGetItem(cpage, PageGetItemId(cpage, offno));
double distance;
/* Use procinfo from the index instead of scan key for performance */
distance = DatumGetFloat8(FunctionCall2Coll(so->procinfo, so->collation, PointerGetDatum(&list->center), value));
if (listCount < so->probes)
{
IvfflatScanList *scanlist;
scanlist = &so->lists[listCount];
scanlist->startPage = list->startPage;
scanlist->distance = distance;
@@ -75,6 +75,8 @@ GetScanLists(IndexScanDesc scan, Datum value)
}
else if (distance < maxDistance)
{
IvfflatScanList *scanlist;
/* Remove */
scanlist = (IvfflatScanList *) pairingheap_remove_first(so->listQueue);
@@ -101,14 +103,6 @@ static void
GetScanItems(IndexScanDesc scan, Datum value)
{
IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
Buffer buf;
Page page;
IndexTuple itup;
BlockNumber searchPage;
OffsetNumber offno;
OffsetNumber maxoffno;
Datum datum;
bool isnull;
TupleDesc tupdesc = RelationGetDescr(scan->indexRelation);
double tuples = 0;
@@ -128,19 +122,32 @@ GetScanItems(IndexScanDesc scan, Datum value)
/* Search closest probes lists */
while (!pairingheap_is_empty(so->listQueue))
{
searchPage = ((IvfflatScanList *) pairingheap_remove_first(so->listQueue))->startPage;
BlockNumber searchPage = ((IvfflatScanList *) pairingheap_remove_first(so->listQueue))->startPage;
/* Search all entry pages for list */
while (BlockNumberIsValid(searchPage))
{
Buffer buf;
Page page;
OffsetNumber maxoffno;
buf = ReadBufferExtended(scan->indexRelation, MAIN_FORKNUM, searchPage, RBM_NORMAL, bas);
LockBuffer(buf, BUFFER_LOCK_SHARE);
page = BufferGetPage(buf);
maxoffno = PageGetMaxOffsetNumber(page);
for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
for (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
{
itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, offno));
IndexTuple itup;
Datum datum;
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);
/*
@@ -181,26 +188,52 @@ GetScanItems(IndexScanDesc scan, Datum value)
}
/*
* Get dimensions from metapage
* Mark prior tuple as dead
*/
static int
GetDimensions(Relation index)
static void
MarkPriorTupleDead(IndexScanDesc scan)
{
Buffer buf;
IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
Buffer buf = so->buf;
Page page;
IvfflatMetaPage metap;
int dimensions;
OffsetNumber maxoffno;
buf = ReadBuffer(index, IVFFLAT_METAPAGE_BLKNO);
/* 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);
metap = IvfflatPageGetMeta(page);
maxoffno = PageGetMaxOffsetNumber(page);
dimensions = metap->dimensions;
for (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
{
ItemId itemid = PageGetItemId(page, offno);
IndexTuple itup = (IndexTuple) PageGetItem(page, itemid);
UnlockReleaseBuffer(buf);
/*
* 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);
}
return dimensions;
break;
}
}
/* Unlock buffer */
LockBuffer(buf, BUFFER_LOCK_UNLOCK);
}
/*
@@ -212,6 +245,7 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
IndexScanDesc scan;
IvfflatScanOpaque so;
int lists;
int dimensions;
AttrNumber attNums[] = {1};
Oid sortOperators[] = {Float8LessOperator};
Oid sortCollations[] = {InvalidOid};
@@ -219,7 +253,9 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
int probes = ivfflat_probes;
scan = RelationGetIndexScan(index, nkeys, norderbys);
lists = IvfflatGetLists(scan->indexRelation);
/* Get lists and dimensions from metapage */
IvfflatGetMetaPageInfo(index, &lists, &dimensions);
if (probes > lists)
probes = lists;
@@ -227,7 +263,9 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
so = (IvfflatScanOpaque) palloc(offsetof(IvfflatScanOpaqueData, lists) + probes * sizeof(IvfflatScanList));
so->buf = InvalidBuffer;
so->first = true;
ItemPointerSetInvalid(&so->heaptid);
so->probes = probes;
so->dimensions = dimensions;
/* Set support functions */
so->procinfo = index_getprocinfo(index, 1, IVFFLAT_DISTANCE_PROC);
@@ -241,7 +279,7 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
so->tupdesc = CreateTemplateTupleDesc(3, false);
#endif
TupleDescInitEntry(so->tupdesc, (AttrNumber) 1, "distance", FLOAT8OID, -1, 0);
TupleDescInitEntry(so->tupdesc, (AttrNumber) 2, "tid", TIDOID, -1, 0);
TupleDescInitEntry(so->tupdesc, (AttrNumber) 2, "heaptid", TIDOID, -1, 0);
TupleDescInitEntry(so->tupdesc, (AttrNumber) 3, "indexblkno", INT4OID, -1, 0);
/* Prep sort */
@@ -274,6 +312,7 @@ 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)
@@ -309,7 +348,7 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
elog(ERROR, "cannot scan ivfflat index without order");
if (scan->orderByData->sk_flags & SK_ISNULL)
value = PointerGetDatum(InitVector(GetDimensions(scan->indexRelation)));
value = PointerGetDatum(InitVector(so->dimensions));
else
{
value = scan->orderByData->sk_argument;
@@ -331,18 +370,27 @@ 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 tid = (ItemPointer) DatumGetPointer(slot_getattr(so->slot, 2, &so->isnull));
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 = *tid;
scan->xs_heaptid = *heaptid;
#else
scan->xs_ctup.t_self = *tid;
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);

View File

@@ -57,12 +57,12 @@ IvfflatGetLists(Relation index)
* Get proc
*/
FmgrInfo *
IvfflatOptionalProcInfo(Relation rel, uint16 procnum)
IvfflatOptionalProcInfo(Relation index, uint16 procnum)
{
if (!OidIsValid(index_getprocid(rel, 1, procnum)))
if (!OidIsValid(index_getprocid(index, 1, procnum)))
return NULL;
return index_getprocinfo(rel, 1, procnum);
return index_getprocinfo(index, 1, procnum);
}
/*
@@ -136,7 +136,6 @@ IvfflatInitRegisterPage(Relation index, Buffer *buf, Page *page, GenericXLogStat
void
IvfflatCommitBuffer(Buffer buf, GenericXLogState *state)
{
MarkBufferDirty(buf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
}
@@ -160,8 +159,6 @@ IvfflatAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **st
IvfflatInitPage(newbuf, newpage);
/* Commit */
MarkBufferDirty(*buf);
MarkBufferDirty(newbuf);
GenericXLogFinish(*state);
/* Unlock */
@@ -172,6 +169,29 @@ IvfflatAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **st
*buf = newbuf;
}
/*
* Get the metapage info
*/
void
IvfflatGetMetaPageInfo(Relation index, int *lists, int *dimensions)
{
Buffer buf;
Page page;
IvfflatMetaPage metap;
buf = ReadBuffer(index, IVFFLAT_METAPAGE_BLKNO);
LockBuffer(buf, BUFFER_LOCK_SHARE);
page = BufferGetPage(buf);
metap = IvfflatPageGetMeta(page);
*lists = metap->lists;
if (dimensions != NULL)
*dimensions = metap->dimensions;
UnlockReleaseBuffer(buf);
}
/*
* Update the start or insert page of a list
*/

View File

@@ -107,7 +107,6 @@ ivfflatbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
{
/* Delete tuples */
PageIndexMultiDelete(page, deletable, ndeletable);
MarkBufferDirty(buf);
GenericXLogFinish(state);
}
else

View File

@@ -695,6 +695,8 @@ vector_spherical_distance(PG_FUNCTION_ARGS)
{
Vector *a = PG_GETARG_VECTOR_P(0);
Vector *b = PG_GETARG_VECTOR_P(1);
float *ax = a->x;
float *bx = b->x;
float dp = 0.0;
double distance;
@@ -702,7 +704,7 @@ vector_spherical_distance(PG_FUNCTION_ARGS)
/* Auto-vectorized */
for (int i = 0; i < a->dim; i++)
dp += a->x[i] * b->x[i];
dp += ax[i] * bx[i];
distance = (double) dp;

View File

@@ -106,6 +106,12 @@ SELECT cosine_distance('[1,1]', '[1,1]');
0
(1 row)
SELECT cosine_distance('[1,0]', '[0,2]');
cosine_distance
-----------------
1
(1 row)
SELECT cosine_distance('[1,1]', '[-1,-1]');
cosine_distance
-----------------

View File

@@ -25,6 +25,7 @@ SELECT inner_product('[3e38]', '[3e38]');
SELECT cosine_distance('[1,2]', '[2,4]');
SELECT cosine_distance('[1,2]', '[0,0]');
SELECT cosine_distance('[1,1]', '[1,1]');
SELECT cosine_distance('[1,0]', '[0,2]');
SELECT cosine_distance('[1,1]', '[-1,-1]');
SELECT cosine_distance('[1,2]', '[3]');
SELECT cosine_distance('[1,1]', '[1.1,1.1]');

View File

@@ -19,8 +19,6 @@ sub test_index_replay
# Wait for replica to catch up
my $applname = $node_replica->name;
my $server_version_num = $node_primary->safe_psql("postgres", "SHOW server_version_num");
my $caughtup_query = "SELECT pg_current_wal_lsn() <= replay_lsn FROM pg_stat_replication WHERE application_name = '$applname';";
$node_primary->poll_query_until('postgres', $caughtup_query)
or die "Timed out while waiting for replica 1 to catch up";

View File

@@ -94,7 +94,7 @@ for my $i (0 .. $#operators)
# Test approximate results
if ($operator ne "<#>")
{
# TODO fix test
# TODO Fix test (uniform random vectors all have similar inner product)
test_recall(1, 0.71, $operator);
test_recall(10, 0.95, $operator);
}
@@ -115,7 +115,7 @@ for my $i (0 .. $#operators)
# Test approximate results
if ($operator ne "<#>")
{
# TODO fix test
# TODO Fix test (uniform random vectors all have similar inner product)
test_recall(1, 0.71, $operator);
test_recall(10, 0.95, $operator);
}

View File

@@ -19,8 +19,6 @@ sub test_index_replay
# Wait for replica to catch up
my $applname = $node_replica->name;
my $server_version_num = $node_primary->safe_psql("postgres", "SHOW server_version_num");
my $caughtup_query = "SELECT pg_current_wal_lsn() <= replay_lsn FROM pg_stat_replication WHERE application_name = '$applname';";
$node_primary->poll_query_until('postgres', $caughtup_query)
or die "Timed out while waiting for replica 1 to catch up";

View File

@@ -83,31 +83,11 @@ for my $i (0 .. $#operators)
push(@expected, $res);
}
# Build index serially
$node->safe_psql("postgres", qq(
SET max_parallel_maintenance_workers = 0;
CREATE INDEX idx ON tst USING hnsw (v $opclass);
));
# Add index
$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v $opclass);");
# Test approximate results
my $min = $operator eq "<#>" ? 0.80 : 0.99;
test_recall($min, $operator);
$node->safe_psql("postgres", "DROP INDEX idx;");
# Build index in parallel
my ($ret, $stdout, $stderr) = $node->psql("postgres", qq(
SET client_min_messages = DEBUG;
SET min_parallel_table_scan_size = 1;
CREATE INDEX idx ON tst USING hnsw (v $opclass);
));
is($ret, 0, $stderr);
like($stderr, qr/using \d+ parallel workers/);
# Test approximate results
test_recall($min, $operator);
$node->safe_psql("postgres", "DROP INDEX idx;");
}
done_testing();

View File

@@ -89,7 +89,7 @@ foreach (@queries)
test_recall(0.20, $limit, "before vacuum");
test_recall(0.95, 100, "before vacuum");
# TODO test concurrent inserts with vacuum
# TODO Test concurrent inserts with vacuum
$node->safe_psql("postgres", "VACUUM tst;");
test_recall(0.95, $limit, "after vacuum");

View File

@@ -0,0 +1,117 @@
use strict;
use warnings;
use PostgresNode;
use TestLib;
use Test::More;
my $node;
my @queries = ();
my @expected;
my $limit = 20;
sub test_recall
{
my ($probes, $min, $operator) = @_;
my $correct = 0;
my $total = 0;
my $explain = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SET ivfflat.probes = $probes;
EXPLAIN ANALYZE SELECT i FROM tst 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;
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 serial, v vector(3));");
# 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_ip_ops", "vector_cosine_ops");
for my $i (0 .. $#operators)
{
my $operator = $operators[$i];
my $opclass = $opclasses[$i];
# Add index
$node->safe_psql("postgres", "CREATE INDEX idx ON tst USING ivfflat (v $opclass);");
# Use concurrent inserts
$node->pgbench(
"--no-vacuum --client=10 --transactions=1000",
0,
[qr{actually processed}],
[qr{^$}],
"concurrent INSERTs",
{
"017_ivfflat_insert_recall_$opclass" => "INSERT INTO tst (v) SELECT ARRAY[random(), random(), random()] FROM generate_series(1, 10) i;"
}
);
# Get exact results
@expected = ();
foreach (@queries)
{
my $res = $node->safe_psql("postgres", qq(
SET enable_indexscan = off;
SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;
));
push(@expected, $res);
}
# Test approximate results
if ($operator ne "<#>")
{
# TODO Fix test (uniform random vectors all have similar inner product)
test_recall(1, 0.71, $operator);
test_recall(10, 0.95, $operator);
}
# Account for equal distances
test_recall(100, 0.9925, $operator);
$node->safe_psql("postgres", "DROP INDEX idx;");
$node->safe_psql("postgres", "TRUNCATE tst;");
}
done_testing();

View File

@@ -0,0 +1,43 @@
use strict;
use warnings;
use PostgresNode;
use TestLib;
use Test::More;
my $dim = 3;
my $array_sql = join(",", ('random()') x $dim);
# Initialize node
my $node = get_new_node('node');
$node->init;
$node->start;
# Create table and index
$node->safe_psql("postgres", "CREATE EXTENSION vector;");
$node->safe_psql("postgres", "CREATE TABLE tst (i serial, v vector($dim));");
$node->safe_psql("postgres",
"INSERT INTO tst (v) SELECT ARRAY[$array_sql] FROM generate_series(1, 10000) i;"
);
$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops);");
# Delete data
$node->safe_psql("postgres", "DELETE FROM tst WHERE i % 100 != 0;");
my $exp = $node->safe_psql("postgres", qq(
SET enable_indexscan = off;
SELECT i FROM tst ORDER BY v <-> '[0,0,0]';
));
# Run twice to make sure correct tuples marked as dead
for (1 .. 2)
{
my $res = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SET ivfflat.probes = 100;
SELECT i FROM tst ORDER BY v <-> '[0,0,0]';
));
is($res, $exp);
}
done_testing();

View File

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