Compare commits

..

207 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
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
Andrew Kane
da3b2fab46 Updated readme [skip ci] 2023-08-31 22:20:13 -07:00
Andrew Kane
884026a23c Updated changelog [skip ci] 2023-08-29 10:13:05 -07:00
Andrew Kane
4d352e6c30 Updated changelog [skip ci] 2023-08-29 10:11:53 -07:00
Andrew Kane
a8e257e1f1 Added comments [skip ci] 2023-08-28 22:02:48 -07:00
Andrew Kane
3913432303 Version bump to 0.5.0 [skip ci] 2023-08-28 17:01:16 -07:00
Andrew Kane
baeebec244 Updated readme [skip ci] 2023-08-28 15:34:07 -07:00
Andrew Kane
d578a4cccc Updated readme [skip ci] 2023-08-28 15:30:28 -07:00
Andrew Kane
1e18e19496 Updated readme [skip ci] 2023-08-28 15:25:11 -07:00
Andrew Kane
39f973dab2 Updated readme [skip ci] 2023-08-28 15:00:58 -07:00
Andrew Kane
453fa15f28 Updated readme [skip ci] 2023-08-28 09:57:46 -07:00
Jonathan S. Katz
e50a79108f Set default HNSW "ef_construction" to 64 (#230) 2023-08-26 16:05:08 -07:00
Andrew Kane
cfaa2ecd7f Improved HNSW graph repair - #239 2023-08-26 14:06:03 -07:00
Andrew Kane
bb1e5ed98f Improved code [skip ci] 2023-08-25 21:05:18 -07:00
Andrew Kane
8801832946 Fixed check in SelectNeighbors 2023-08-25 20:46:05 -07:00
Andrew Kane
552c64d492 Improved HNSW code 2023-08-25 20:39:07 -07:00
Andrew Kane
86c29b3bf0 Improved param code 2023-08-23 21:10:00 -07:00
Andrew Kane
e406b7f5ea Added comments [skip ci] 2023-08-23 21:03:07 -07:00
Andrew Kane
6d88a9e1d2 Updated HNSW_PAGE_ID [skip ci] 2023-08-21 22:59:53 -07:00
Andrew Kane
1e851c12c0 Updated comment [skip ci] 2023-08-21 22:55:24 -07:00
Andrew Kane
8ed3cc5f0b Improved macro [skip ci] 2023-08-21 22:52:58 -07:00
Andrew Kane
bace0891bd Updated comments [skip ci] 2023-08-21 22:51:24 -07:00
Andrew Kane
4600979504 Updated comments [skip ci] 2023-08-21 22:38:09 -07:00
Andrew Kane
69addf05d0 Updated comments [skip ci] 2023-08-21 22:21:53 -07:00
Andrew Kane
4a5ae8a8df Updated comment [skip ci] 2023-08-21 21:59:47 -07:00
Andrew Kane
ce7da66ca2 Updated comment [skip ci] 2023-08-21 20:57:54 -07:00
Andrew Kane
c6be1c5e13 Fixed test 2023-08-21 20:45:56 -07:00
Andrew Kane
9b3d1a32ff Updated comment [skip ci] 2023-08-21 16:27:03 -07:00
Andrew Kane
8420734350 Updated comments [skip ci] 2023-08-21 16:26:32 -07:00
Andrew Kane
8b03267267 Improved locking for HNSW vacuum [skip ci] 2023-08-21 16:24:55 -07:00
Andrew Kane
641ddf5413 Improved locking for HNSW vacuum 2023-08-21 16:12:29 -07:00
Andrew Kane
782a1051e3 Improved locking for HNSW vacuum 2023-08-21 16:06:32 -07:00
Andrew Kane
ca2be5be6e Updated comment [skip ci] 2023-08-21 16:00:44 -07:00
Andrew Kane
0e1de45463 Improved locking code [skip ci] 2023-08-21 15:42:59 -07:00
Andrew Kane
3f3b3ca8e3 Made function static [skip ci] 2023-08-21 03:27:42 -07:00
Andrew Kane
d4fe67e8ee Simplified locking for entry point 2023-08-21 03:22:23 -07:00
Andrew Kane
02f4e0ec8b Revert "Added version to reduce stale reads and writes and prepare for optimistic locking"
This reverts commit ef1209eaf4.
2023-08-21 02:47:27 -07:00
Andrew Kane
1301706d30 Increased concurrency in HNSW insert test 2023-08-21 02:46:16 -07:00
Andrew Kane
0d58683011 Improved HNSW insert test 2023-08-21 02:44:20 -07:00
Andrew Kane
90a042e5aa Wait for selects to complete 2023-08-21 02:24:53 -07:00
Andrew Kane
bbd57dfebf Moved wait [skip ci] 2023-08-21 01:02:05 -07:00
Andrew Kane
40a013a622 Wait for inserts to complete when vacuuming HNSW 2023-08-21 00:36:32 -07:00
Andrew Kane
ef1209eaf4 Added version to reduce stale reads and writes and prepare for optimistic locking 2023-08-20 17:08:20 -07:00
Andrew Kane
687263ccd4 DRY HNSW vacuum code 2023-08-20 14:52:31 -07:00
Andrew Kane
a62c045c93 Update metapage if needed for HNSW vacuum 2023-08-19 12:19:42 -07:00
Andrew Kane
651e4343c1 Made entryPoint argument for RepairGraphElement 2023-08-19 11:47:13 -07:00
Andrew Kane
2ca2ef94e6 Use rmdir [skip ci] 2023-08-19 10:15:50 -07:00
Andrew Kane
6ba95691d3 Added headers for Windows 2023-08-19 10:03:15 -07:00
Andrew Kane
e9de248e4f Skip i386 tests for windows branches [skip ci] 2023-08-19 10:01:55 -07:00
Andrew Kane
fe79d435c0 Use HEADERS for simplicity [skip ci] 2023-08-19 09:58:25 -07:00
Florents Tselai
b34525fbc2 Make pgvector headers available to others. (#233) 2023-08-19 09:56:02 -07:00
Andrew Kane
206c42e170 Cleaned up header 2023-08-19 01:22:28 -07:00
Andrew Kane
4f8d824280 Use variable for uninstall [skip ci] 2023-08-19 01:21:05 -07:00
Andrew Kane
ca847d02b0 Fixed highest point in HNSW vacuum [skip ci] 2023-08-18 22:23:19 -07:00
Andrew Kane
ca726052ae Fixed flaky test [skip ci] 2023-08-18 20:54:31 -07:00
Andrew Kane
2b25127f3d Fixed flaky test [skip ci] 2023-08-18 20:53:47 -07:00
Andrew Kane
fecb8c06c9 Simplified code 2023-08-18 20:49:00 -07:00
Andrew Kane
a03f6ae4bc Added prefix to function name [skip ci] 2023-08-18 00:54:09 -07:00
Andrew Kane
ed20d98777 Check if item pointer is valid [skip ci] 2023-08-16 17:29:59 -07:00
Andrew Kane
b72a22b3c0 Fixed duplicate connections when repairing graph 2023-08-16 17:07:19 -07:00
Andrew Kane
8c5c0f791e Improved HNSW insert code [skip ci] 2023-08-16 16:31:11 -07:00
Andrew Kane
f9f5ae61df Added test for select with no data for ivfflat 2023-08-15 23:44:18 -07:00
Andrew Kane
2b138d6cb5 Fixed select with no data 2023-08-15 23:25:28 -07:00
Andrew Kane
e8f36aee92 Improved HNSW vacuum code [skip ci] 2023-08-15 22:36:47 -07:00
Andrew Kane
508a8a9ac7 Updated comments [skip ci] 2023-08-15 20:32:15 -07:00
Andrew Kane
d0b0298cda Improved HNSW vacuum code [skip ci] 2023-08-15 20:02:21 -07:00
Andrew Kane
c3bafc76e8 Only update entry point on inserts if level is greater 2023-08-15 12:31:06 -07:00
Andrew Kane
86a062e504 Set entry level in CreateMetaPage [skip ci] 2023-08-15 11:47:05 -07:00
Andrew Kane
b421e76f29 Updated comment [skip ci] 2023-08-13 17:23:52 -07:00
Andrew Kane
ef480ff0b9 Added comment [skip ci] 2023-08-13 17:22:25 -07:00
Andrew Kane
e1d6654063 Revert "Improved HNSW vacuum performance"
This reverts commit c5b2f3ac8b.
2023-08-13 17:19:22 -07:00
Andrew Kane
c5b2f3ac8b Improved HNSW vacuum performance 2023-08-13 16:50:36 -07:00
Andrew Kane
0f238b1fa9 Update neighbors when vacuuming 2023-08-13 16:17:30 -07:00
Andrew Kane
6fc7d06313 Improved HNSW vacuuming 2023-08-13 15:53:07 -07:00
Andrew Kane
304e262a41 Fixed initialization 2023-08-13 15:50:05 -07:00
Andrew Kane
23cb79c1dc Fixed recall for HNSW after vacuuming 2023-08-13 15:37:47 -07:00
Andrew Kane
dc8d467ba5 Improved HNSW vacuum recall test 2023-08-13 13:46:06 -07:00
Andrew Kane
29a0a00731 Improved test [skip ci] 2023-08-13 13:42:51 -07:00
Andrew Kane
eda07ccf7c Improved test 2023-08-13 13:35:43 -07:00
Andrew Kane
c694dbb776 Added test for HNSW vacuum recall 2023-08-13 13:09:51 -07:00
Andrew Kane
3a857ef9f9 Updated readme [skip ci] 2023-08-13 12:16:36 -07:00
Andrew Kane
b60e2483a9 Improved HNSW vacuum stats [skip ci] 2023-08-12 14:09:13 -07:00
Andrew Kane
4295ee3b3a Updated todos [skip ci] 2023-08-12 13:47:20 -07:00
Andrew Kane
8b1ea42d42 Improved HNSW vacuum test 2023-08-12 12:41:27 -07:00
Andrew Kane
3c291fea41 Added todo [skip ci] 2023-08-12 12:41:02 -07:00
Andrew Kane
3d568ef6a1 Added comments [skip ci] 2023-08-11 21:12:18 -07:00
Andrew Kane
7db3b08ef4 Simplified tests [skip ci] 2023-08-11 19:41:55 -07:00
Andrew Kane
4196c88dbd Added opclass to tests [skip ci] 2023-08-11 19:40:04 -07:00
Andrew Kane
3e7ee6ea04 Simplified opclasses in TAP tests [skip ci] 2023-08-11 19:20:03 -07:00
Andrew Kane
5616c1d7a7 Simplified TAP tests [skip ci] 2023-08-11 19:16:08 -07:00
Andrew Kane
0778a507de Use consistent style in TAP tests, part 2 [skip ci] 2023-08-11 19:13:22 -07:00
Andrew Kane
aabbeac30c Use consistent style in TAP tests [skip ci] 2023-08-11 19:08:15 -07:00
Andrew Kane
746dce72ef Use concurrent inserts for insert recall test 2023-08-11 18:41:00 -07:00
Andrew Kane
098813c483 Improved concurrent inserts test for HNSW [skip ci] 2023-08-11 18:33:01 -07:00
Andrew Kane
9d689afd7e Updated test name [skip ci] 2023-08-11 17:17:28 -07:00
Andrew Kane
03930e15a4 Updated comment [skip ci] 2023-08-11 17:09:09 -07:00
Andrew Kane
8a807c01f6 Test fallback path for duplicates [skip ci] 2023-08-11 17:07:38 -07:00
Andrew Kane
85ebb7714e Updated test name [skip ci] 2023-08-11 17:07:17 -07:00
Andrew Kane
34e99593ab Added comment [skip ci] 2023-08-11 17:02:50 -07:00
Andrew Kane
fc45231ef0 Improving naming [skip ci] 2023-08-11 14:56:46 -07:00
Andrew Kane
0191a90f9f Updated comment [skip ci] 2023-08-11 14:37:56 -07:00
Andrew Kane
82e83bd3b1 Added comment [skip ci] 2023-08-11 14:29:24 -07:00
Andrew Kane
d8a14c658a Fixed flaky test [skip ci] 2023-08-11 13:38:42 -07:00
Andrew Kane
ae5de26893 Improved logic for updating HNSW insert page 2023-08-11 13:24:56 -07:00
Andrew Kane
daef83a112 Added comment [skip ci] 2023-08-11 13:03:38 -07:00
Andrew Kane
c9d82c6bdc Revert "Improved insert logic [skip ci]"
This reverts commit 573a336f53.
2023-08-11 12:52:58 -07:00
Andrew Kane
573a336f53 Improved insert logic [skip ci] 2023-08-11 12:39:22 -07:00
Andrew Kane
e7a913f361 Updated comment [skip ci] 2023-08-11 10:11:01 -07:00
Andrew Kane
33438c3cf9 Use neighbor page for insert page [skip ci] 2023-08-11 09:51:41 -07:00
Andrew Kane
8ec1821f1e Added comment [skip ci] 2023-08-11 09:45:08 -07:00
Andrew Kane
b288de719d Consider previously deleted tuples for insert page [skip ci] 2023-08-11 09:41:31 -07:00
Andrew Kane
c13f713f29 Improved test [skip ci] 2023-08-11 09:36:29 -07:00
Andrew Kane
6d15435003 Updated comment [skip ci] 2023-08-11 09:26:54 -07:00
Andrew Kane
12d15a9c41 Improved inserts for HNSW 2023-08-11 09:23:56 -07:00
Andrew Kane
c64288377b Updated min ef_construction to 4 [skip ci] 2023-08-10 21:11:10 -07:00
Andrew Kane
451e36cee7 Added check for ef_construction [skip ci] 2023-08-10 21:09:38 -07:00
Andrew Kane
3ff2e34d7f Updated min M to 2 [skip ci] 2023-08-10 20:57:50 -07:00
Andrew Kane
94a684c9e9 Added test for duplicates 2023-08-10 20:52:58 -07:00
Andrew Kane
1a0d7bccc7 Updated min ef_search to 1 [skip ci] 2023-08-10 20:47:15 -07:00
Andrew Kane
ded8bda72a Fixed flaky test [skip ci] 2023-08-10 18:59:11 -07:00
Andrew Kane
9b967d592f Improved concurrent inserts with empty entry point, part 2 2023-08-10 18:31:58 -07:00
Andrew Kane
ed513e62c1 Improved code for skipping element [skip ci] 2023-08-10 15:29:08 -07:00
Andrew Kane
27ccb5fa7a Improved code [skip ci] 2023-08-10 15:23:26 -07:00
Andrew Kane
92e25e7733 Fixed warning 2023-08-10 15:10:07 -07:00
Andrew Kane
a18bea24b8 Improved concurrent inserts with low number of elements 2023-08-10 15:07:28 -07:00
Andrew Kane
483173460b Improved concurrent inserts with empty entry point 2023-08-10 14:35:24 -07:00
Andrew Kane
da8a914106 Moved FindDuplicate [skip ci] 2023-08-10 13:44:21 -07:00
Andrew Kane
4d6da72b08 Simplified code 2023-08-10 09:56:17 -07:00
Andrew Kane
443c7a5dab Updated variable name [skip ci] 2023-08-10 09:02:01 -07:00
Andrew Kane
9287fe8bcc Updated comment [skip ci] 2023-08-10 08:59:38 -07:00
Andrew Kane
309fa94c05 Improved performance for duplicates 2023-08-10 08:52:45 -07:00
Andrew Kane
6f15dd266c Improved construction code 2023-08-10 08:38:31 -07:00
Andrew Kane
df68eb4570 Look for duplicates last, since may need to fallback 2023-08-10 00:00:27 -07:00
Andrew Kane
7c0d94c99c Improved concurrent inserts 2023-08-09 23:51:35 -07:00
Andrew Kane
d63d430af8 Fixed flaky test [skip ci] 2023-08-09 20:06:28 -07:00
Andrew Kane
5d62e4d080 Moved SelectNeighbors after duplicate check 2023-08-09 19:19:41 -07:00
Andrew Kane
dab8f25d1c Fixed overflow with vector_norm 2023-08-09 16:33:54 -07:00
57 changed files with 1793 additions and 978 deletions

View File

@@ -77,5 +77,24 @@ jobs:
nmake /NOLOGO /F Makefile.win clean && ^
nmake /NOLOGO /F Makefile.win uninstall
shell: cmd
- if: ${{ failure() }}
run: cat regression.diffs
i386:
if: ${{ !startsWith(github.ref_name, 'mac') && !startsWith(github.ref_name, 'windows') }}
runs-on: ubuntu-latest
container:
image: debian:11
options: --platform linux/386
steps:
- run: apt-get update && apt-get install -y build-essential git libipc-run-perl postgresql-13 postgresql-server-dev-13 sudo
- run: service postgresql start
- run: |
git clone https://github.com/${{ github.repository }}.git pgvector
cd pgvector
git fetch origin ${{ github.ref }}
git reset --hard FETCH_HEAD
make
make install
chown -R postgres .
sudo -u postgres make installcheck
sudo -u postgres make prove_installcheck
env:
PG_CFLAGS: -Wall -Wextra -Werror -Wno-unused-parameter -Wno-sign-compare

View File

@@ -1,14 +1,18 @@
## 0.5.0 (unreleased)
## 0.5.1 (2023-10-10)
- Improved performance of HNSW index builds
- Added check for MVCC-compliant snapshot for index scans
## 0.5.0 (2023-08-28)
- Added HNSW index type
- Added support for parallel index builds
- Added support for parallel index builds for IVFFlat
- Added `l1_distance` function
- Added `normalize_l2` function
- Added element-wise multiplication for vectors
- Added `sum` aggregate
- Improved performance of distance functions
- Fixed out of range results for cosine distance
- Fixed results for NULL and NaN distances
- Fixed results for NULL and NaN distances for IVFFlat
## 0.4.4 (2023-06-12)

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": "0.4.4",
"version": "0.5.1",
"maintainer": [
"Andrew Kane <andrew@ankane.org>"
],
@@ -20,7 +20,7 @@
"vector": {
"file": "sql/vector.sql",
"docfile": "README.md",
"version": "0.4.4",
"version": "0.5.1",
"abstract": "Open-source vector similarity search for Postgres"
}
},

View File

@@ -1,13 +1,14 @@
EXTENSION = vector
EXTVERSION = 0.4.4
EXTVERSION = 0.5.1
MODULE_big = vector
DATA = $(wildcard sql/*--*.sql)
OBJS = src/hnsw.o src/hnswbuild.o src/hnswinsert.o src/hnswscan.o src/hnswutils.o src/hnswvacuum.o src/ivfbuild.o src/ivfflat.o src/ivfinsert.o src/ivfkmeans.o src/ivfscan.o src/ivfutils.o src/ivfvacuum.o src/vector.o
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

@@ -1,10 +1,11 @@
EXTENSION = vector
EXTVERSION = 0.4.4
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
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
@@ -54,6 +55,8 @@ install:
copy $(SHLIB) "$(PKGLIBDIR)"
copy $(EXTENSION).control "$(SHAREDIR)\extension"
copy sql\$(EXTENSION)--*.sql "$(SHAREDIR)\extension"
mkdir "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
copy $(HEADERS) "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
installcheck:
"$(BINDIR)\pg_regress" --bindir="$(BINDIR)" $(REGRESS_OPTS) $(REGRESS)
@@ -61,7 +64,9 @@ installcheck:
uninstall:
del /f "$(PKGLIBDIR)\$(SHLIB)"
del /f "$(SHAREDIR)\extension\$(EXTENSION).control"
del /f "$(SHAREDIR)\extension\vector--*.sql"
del /f "$(SHAREDIR)\extension\$(EXTENSION)--*.sql"
del /f "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)\*.h"
rmdir "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
clean:
del /f $(SHLIB) $(EXTENSION).lib $(EXTENSION).exp

135
README.md
View File

@@ -2,7 +2,7 @@
Open-source vector similarity search for Postgres
Store all of your application data in one place. Supports:
Store your vectors with the rest of your data. Supports:
- exact and approximate nearest neighbor search
- L2 distance, inner product, and cosine distance
@@ -18,7 +18,7 @@ Compile and install the extension (supports Postgres 11+)
```sh
cd /tmp
git clone --branch v0.4.4 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
@@ -157,7 +157,16 @@ SELECT category_id, AVG(embedding) FROM items GROUP BY category_id;
By default, pgvector performs exact nearest neighbor search, which provides perfect recall.
You can add an index to use approximate nearest neighbor search, which trades some recall for performance. Unlike typical indexes, you will see different results for queries after adding an approximate index.
You can add an index to use approximate nearest neighbor search, which trades some recall for speed. Unlike typical indexes, you will see different results for queries after adding an approximate index.
Supported index types are:
- [IVFFlat](#ivfflat)
- [HNSW](#hnsw) - added in 0.5.0
## IVFFlat
An IVFFlat index divides vectors into lists, and then searches a subset of those lists that are closest to the query vector. It has faster build times and uses less memory than HNSW, but has lower query performance (in terms of speed-recall tradeoff).
Three keys to achieving good recall are:
@@ -206,7 +215,63 @@ SELECT ...
COMMIT;
```
### Indexing Progress
## HNSW
An HNSW index creates a multilayer graph. It has slower build times and uses more memory than IVFFlat, but has better query performance (in terms of speed-recall tradeoff). Theres no training step like IVFFlat, so the index can be created without any data in the table.
Add an index for each distance function you want to use.
L2 distance
```sql
CREATE INDEX ON items USING hnsw (embedding vector_l2_ops);
```
Inner product
```sql
CREATE INDEX ON items USING hnsw (embedding vector_ip_ops);
```
Cosine distance
```sql
CREATE INDEX ON items USING hnsw (embedding vector_cosine_ops);
```
Vectors with up to 2,000 dimensions can be indexed.
### Index Options
Specify HNSW parameters
- `m` - the max number of connections per layer (16 by default)
- `ef_construction` - the size of the dynamic candidate list for constructing the graph (64 by default)
```sql
CREATE INDEX ON items USING hnsw (embedding vector_l2_ops) WITH (m = 16, ef_construction = 64);
```
### Query Options
Specify the size of the dynamic candidate list for search (40 by default)
```sql
SET hnsw.ef_search = 100;
```
A higher value provides better recall at the cost of speed.
Use `SET LOCAL` inside a transaction to set it for a single query
```sql
BEGIN;
SET LOCAL hnsw.ef_search = 100;
SELECT ...
COMMIT;
```
## Indexing Progress
Check [indexing progress](https://www.postgresql.org/docs/current/progress-reporting.html#CREATE-INDEX-PROGRESS-REPORTING) with Postgres 12+
@@ -217,8 +282,8 @@ SELECT phase, tuples_done, tuples_total FROM pg_stat_progress_create_index;
The phases are:
1. `initializing`
2. `performing k-means`
3. `sorting tuples`
2. `performing k-means` - IVFFlat only
3. `assigning tuples` - IVFFlat only
4. `loading tuples`
Note: `tuples_done` and `tuples_total` are only populated during the `loading tuples` phase
@@ -283,7 +348,7 @@ SELECT * FROM items ORDER BY embedding <#> '[3,1,2]' LIMIT 5;
### Approximate Search
To speed up queries with an index, increase the number of inverted lists (at the expense of recall).
To speed up queries with an IVFFlat index, increase the number of inverted lists (at the expense of recall).
```sql
CREATE INDEX ON items USING ivfflat (embedding vector_l2_ops) WITH (lists = 1000);
@@ -359,7 +424,7 @@ or choose to store vectors inline:
ALTER TABLE items ALTER COLUMN embedding SET STORAGE PLAIN;
```
#### Why are there less results for a query after adding an index?
#### Why are there less results for a query after adding an IVFFlat index?
The index was likely created with too little data for the number of lists. Drop the index until the table has more data.
@@ -375,33 +440,32 @@ Each vector takes `4 * dimensions + 8` bytes of storage. Each element is a singl
### Vector Operators
Operator | Description
--- | ---
\+ | element-wise addition
\- | element-wise subtraction
\* | element-wise multiplication [unreleased]
<-> | Euclidean distance
<#> | negative inner product
<=> | cosine distance
Operator | Description | Added
--- | --- | ---
\+ | element-wise addition |
\- | element-wise subtraction |
\* | element-wise multiplication | 0.5.0
<-> | Euclidean distance |
<#> | negative inner product |
<=> | cosine distance |
### Vector Functions
Function | Description
--- | ---
cosine_distance(vector, vector) → double precision | cosine distance
inner_product(vector, vector) → double precision | inner product
l2_distance(vector, vector) → double precision | Euclidean distance
l1_distance(vector, vector) → double precision | taxicab distance [unreleased]
normalize_l2(vector) → vector | normalize with Euclidean norm [unreleased]
vector_dims(vector) → integer | number of dimensions
vector_norm(vector) → double precision | Euclidean norm
Function | Description | Added
--- | --- | ---
cosine_distance(vector, vector) → double precision | cosine distance |
inner_product(vector, vector) → double precision | inner product |
l2_distance(vector, vector) → double precision | Euclidean distance |
l1_distance(vector, vector) → double precision | taxicab distance | 0.5.0
vector_dims(vector) → integer | number of dimensions |
vector_norm(vector) → double precision | Euclidean norm |
### Aggregate Functions
Function | Description
--- | ---
avg(vector) → vector | arithmetic mean
sum(vector) → vector | sum [unreleased]
Function | Description | Added
--- | --- | ---
avg(vector) → vector | average |
sum(vector) → vector | sum | 0.5.0
## Installation Notes
@@ -445,7 +509,7 @@ Then use `nmake` to build:
```cmd
set "PGROOT=C:\Program Files\PostgreSQL\15"
git clone --branch v0.4.4 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
@@ -466,9 +530,8 @@ This adds pgvector to the [Postgres image](https://hub.docker.com/_/postgres) (r
You can also build the image manually:
```sh
git clone --branch v0.4.4 https://github.com/pgvector/pgvector.git
git clone --branch v0.5.1 https://github.com/pgvector/pgvector.git
cd pgvector
git cherry-pick 237a6df
docker build --build-arg PG_MAJOR=15 -t myuser/pgvector .
```
@@ -532,12 +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

@@ -4,9 +4,6 @@
CREATE FUNCTION l1_distance(vector, vector) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION normalize_l2(vector) RETURNS vector
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION vector_mul(vector, vector) RETURNS vector
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;

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

@@ -49,9 +49,6 @@ CREATE FUNCTION vector_dims(vector) RETURNS integer
CREATE FUNCTION vector_norm(vector) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION normalize_l2(vector) RETURNS vector
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION vector_add(vector, vector) RETURNS vector
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;

View File

@@ -37,7 +37,7 @@ HnswInit(void)
);
DefineCustomIntVariable("hnsw.ef_search", "Sets the size of the dynamic candidate list for search",
"Valid range is 10..1000.", &hnsw_ef_search,
"Valid range is 1..1000.", &hnsw_ef_search,
HNSW_DEFAULT_EF_SEARCH, HNSW_MIN_EF_SEARCH, HNSW_MAX_EF_SEARCH, PGC_USERSET, 0, NULL, NULL, NULL);
}
@@ -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

@@ -22,33 +22,44 @@
#define HNSW_VERSION 1
#define HNSW_MAGIC_NUMBER 0xA953A953
#define HNSW_PAGE_ID 0xFF85
#define HNSW_PAGE_ID 0xFF90
/* Preserved page numbers */
#define HNSW_METAPAGE_BLKNO 0
#define HNSW_HEAD_BLKNO 1 /* first element page */
/* Must correspond to page numbers since page lock is used */
#define HNSW_UPDATE_LOCK 0
#define HNSW_SCAN_LOCK 1
/* HNSW parameters */
#define HNSW_DEFAULT_M 16
#define HNSW_MIN_M 4
#define HNSW_MIN_M 2
#define HNSW_MAX_M 100
#define HNSW_DEFAULT_EF_CONSTRUCTION 40
#define HNSW_MIN_EF_CONSTRUCTION 10
#define HNSW_DEFAULT_EF_CONSTRUCTION 64
#define HNSW_MIN_EF_CONSTRUCTION 4
#define HNSW_MAX_EF_CONSTRUCTION 1000
#define HNSW_DEFAULT_EF_SEARCH 40
#define HNSW_MIN_EF_SEARCH 10
#define HNSW_MIN_EF_SEARCH 1
#define HNSW_MAX_EF_SEARCH 1000
/* Tuple types */
#define HNSW_ELEMENT_TUPLE_TYPE 1
#define HNSW_NEIGHBOR_TUPLE_TYPE 2
/* Make graph robust against non-HOT updates */
#define HNSW_HEAPTIDS 10
#define HNSW_UPDATE_ENTRY_GREATER 1
#define HNSW_UPDATE_ENTRY_ALWAYS 2
/* Build phases */
/* 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))
@@ -68,10 +79,13 @@
#define HnswIsElementTuple(tup) ((tup)->type == HNSW_ELEMENT_TUPLE_TYPE)
#define HnswIsNeighborTuple(tup) ((tup)->type == HNSW_NEIGHBOR_TUPLE_TYPE)
#define HnswGetLayerM(m, layer) (layer == 0 ? m * 2 : m)
/* 2 * M connections for ground layer */
#define HnswGetLayerM(m, layer) (layer == 0 ? (m) * 2 : (m))
/* Optimal ML from paper */
#define HnswGetMl(m) (1 / log(m))
/* Ensure fits in uint8 */
/* Ensure fits on page and in uint8 */
#define HnswGetMaxLevel(m) Min(((BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - offsetof(HnswNeighborTupleData, indextids) - sizeof(ItemIdData)) / (sizeof(ItemPointerData)) / m) - 2, 255)
/* Variables */
@@ -84,12 +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;
@@ -98,21 +113,16 @@ typedef struct HnswCandidate
{
HnswElement element;
float distance;
bool closer;
} HnswCandidate;
typedef struct HnswNeighborArray
{
int length;
bool closerSet;
HnswCandidate *items;
} HnswNeighborArray;
typedef struct HnswUpdate
{
HnswCandidate hc;
int level;
int index;
} HnswUpdate;
typedef struct HnswPairingHeapNode
{
pairingheap_node ph_node;
@@ -195,7 +205,7 @@ typedef struct HnswElementTupleData
ItemPointerData heaptids[HNSW_HEAPTIDS];
ItemPointerData neighbortid;
uint16 unused2;
Vector vec;
char value[FLEXIBLE_ARRAY_MEMBER];
} HnswElementTupleData;
typedef HnswElementTupleData * HnswElementTuple;
@@ -213,7 +223,6 @@ typedef HnswNeighborTupleData * HnswNeighborTuple;
typedef struct HnswScanOpaqueData
{
bool first;
Buffer buf;
List *w;
MemoryContext tmpCtx;
@@ -254,26 +263,33 @@ 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, BlockNumber *skipPage, OffsetNumber *skipOffno);
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 HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo *procinfo, Oid collation, int m, int efConstruction, List **updates, bool vacuuming);
HnswCandidate *HnswEntryCandidate(HnswElement em, Datum q, Relation rel, FmgrInfo *procinfo, Oid collation, bool loadvec);
void HnswUpdateMetaPage(Relation index, bool updateEntry, HnswElement entryPoint, BlockNumber insertPage, ForkNumber forkNum);
HnswElement HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno);
void HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo *procinfo, Oid collation, int m, int efConstruction, bool existing);
HnswElement HnswFindDuplicate(HnswElement e);
HnswCandidate *HnswEntryCandidate(HnswElement em, Datum q, Relation rel, FmgrInfo *procinfo, Oid collation, bool loadVec);
void HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, BlockNumber insertPage, ForkNumber forkNum);
void HnswSetNeighborTuple(HnswNeighborTuple ntup, HnswElement e, int m);
void HnswAddHeapTid(HnswElement element, ItemPointer heaptid);
void HnswInitNeighbors(HnswElement element, int m);
bool HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel);
void HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement e, int m, bool checkExisting);
void HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec);
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, int m);
/* Index access methods */
IndexBuildResult *hnswbuild(Relation heap, Relation index, IndexInfo *indexInfo);

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
@@ -60,6 +61,7 @@ CreateMetaPage(HnswBuildState * buildstate)
metap->efConstruction = buildstate->efConstruction;
metap->entryBlkno = InvalidBlockNumber;
metap->entryOffno = InvalidOffsetNumber;
metap->entryLevel = -1;
metap->insertPage = InvalidBlockNumber;
((PageHeader) page)->pd_lower =
((char *) metap + sizeof(HnswMetaPageData)) - (char *) page;
@@ -80,7 +82,6 @@ HnswBuildAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **
HnswPageGetOpaque(*page)->nextblkno = BufferGetBlockNumber(newbuf);
/* Commit */
MarkBufferDirty(*buf);
GenericXLogFinish(*state);
UnlockReleaseBuffer(*buf);
@@ -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);
@@ -178,11 +178,10 @@ CreateElementPages(HnswBuildState * buildstate)
insertPage = BufferGetBlockNumber(buf);
/* Commit */
MarkBufferDirty(buf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
HnswUpdateMetaPage(index, true, buildstate->entryPoint, insertPage, forkNum);
HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_ALWAYS, buildstate->entryPoint, insertPage, forkNum);
pfree(etup);
pfree(ntup);
@@ -226,7 +225,6 @@ CreateNeighborPages(HnswBuildState * buildstate)
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
/* Commit */
MarkBufferDirty(buf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
}
@@ -275,20 +273,33 @@ 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 */
*dup = HnswInsertElement(element, entryPoint, NULL, procinfo, collation, m, efConstruction, NULL, false);
HnswInsertElement(element, entryPoint, NULL, procinfo, collation, m, efConstruction, false);
/* Look for duplicate */
*dup = HnswFindDuplicate(element);
/* Update neighbors if needed */
if (*dup == NULL)
{
for (int lc = element->level; lc >= 0; lc--)
{
int lm = HnswGetLayerM(m, lc);
HnswNeighborArray *neighbors = &element->neighbors[lc];
for (int i = 0; i < neighbors->length; i++)
HnswUpdateConnection(element, &neighbors->items[i], lm, lc, NULL, NULL, procinfo, collation);
}
}
/* Update entry point if needed */
if (*dup == NULL && (entryPoint == NULL || element->level > entryPoint->level))
@@ -346,7 +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);
@@ -354,9 +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)
@@ -364,9 +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);
}
/*
@@ -381,6 +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;
}
@@ -400,13 +417,12 @@ 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);
if (buildstate->efConstruction < 2 * buildstate->m)
elog(ERROR, "ef_construction must be greater than or equal to 2 * m");
buildstate->reltuples = 0;
buildstate->indtuples = 0;
@@ -423,6 +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

@@ -34,7 +34,7 @@ GetInsertPage(Relation index)
* Check for a free offset
*/
static bool
HnswFreeOffset(Relation index, Buffer buf, Page page, HnswElement element, Size ntupSize, Buffer *nbuf, Page *npage, OffsetNumber *freeOffno, OffsetNumber *freeNeighborOffno, BlockNumber *firstFreePage)
HnswFreeOffset(Relation index, Buffer buf, Page page, HnswElement element, Size ntupSize, Buffer *nbuf, Page *npage, OffsetNumber *freeOffno, OffsetNumber *freeNeighborOffno, BlockNumber *newInsertPage)
{
OffsetNumber offno;
OffsetNumber maxoffno = PageGetMaxOffsetNumber(page);
@@ -49,14 +49,15 @@ HnswFreeOffset(Relation index, Buffer buf, Page page, HnswElement element, Size
if (etup->deleted)
{
BlockNumber elementPage = BufferGetBlockNumber(buf);
BlockNumber neighborPage = ItemPointerGetBlockNumber(&etup->neighbortid);
OffsetNumber neighborOffno = ItemPointerGetOffsetNumber(&etup->neighbortid);
ItemId itemid;
if (!BlockNumberIsValid(*firstFreePage))
*firstFreePage = neighborPage;
if (!BlockNumberIsValid(*newInsertPage))
*newInsertPage = elementPage;
if (neighborPage == BufferGetBlockNumber(buf))
if (neighborPage == elementPage)
{
*nbuf = buf;
*npage = page;
@@ -110,7 +111,7 @@ HnswInsertAppendPage(Relation index, Buffer *nbuf, Page *npage, GenericXLogState
* Add to element and neighbor pages
*/
static void
WriteNewElementPages(Relation index, HnswElement e, int m)
WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPage, BlockNumber *updatedInsertPage)
{
Buffer buf;
Page page;
@@ -118,21 +119,23 @@ WriteNewElementPages(Relation index, HnswElement e, int m)
Size etupSize;
Size ntupSize;
Size combinedSize;
Size maxSize;
Size minCombinedSize;
HnswElementTuple etup;
BlockNumber insertPage = GetInsertPage(index);
BlockNumber originalInsertPage = insertPage;
int dimensions = e->vec->dim;
BlockNumber currentPage = insertPage;
HnswNeighborTuple ntup;
Buffer nbuf;
Page npage;
OffsetNumber freeOffno = InvalidOffsetNumber;
OffsetNumber freeNeighborOffno = InvalidOffsetNumber;
BlockNumber firstFreePage = InvalidBlockNumber;
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 = HNSW_MAX_SIZE;
minCombinedSize = etupSize + HNSW_NEIGHBOR_TUPLE_SIZE(0, m) + sizeof(ItemIdData);
/* Prepare element tuple */
etup = palloc0(etupSize);
@@ -142,16 +145,22 @@ WriteNewElementPages(Relation index, HnswElement e, int m)
ntup = palloc0(ntupSize);
HnswSetNeighborTuple(ntup, e, m);
/* Find a page to insert the item */
/* Find a page (or two if needed) to insert the tuples */
for (;;)
{
buf = ReadBuffer(index, insertPage);
buf = ReadBuffer(index, currentPage);
LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
state = GenericXLogStart(index);
page = GenericXLogRegisterBuffer(state, buf, 0);
/* Space for both */
/* Keep track of first page where element at level 0 can fit */
if (!BlockNumberIsValid(newInsertPage) && PageGetFreeSpace(page) >= minCombinedSize)
newInsertPage = currentPage;
/* First, try the fastest path */
/* Space for both tuples on the current page */
/* This can split existing tuples in rare cases */
if (PageGetFreeSpace(page) >= combinedSize)
{
nbuf = buf;
@@ -159,15 +168,8 @@ WriteNewElementPages(Relation index, HnswElement e, int m)
break;
}
/* Space for element but not neighbors and last page */
if (PageGetFreeSpace(page) >= etupSize && !BlockNumberIsValid(HnswPageGetOpaque(page)->nextblkno))
{
HnswInsertAppendPage(index, &nbuf, &npage, state, page);
break;
}
/* Space from deleted item */
if (HnswFreeOffset(index, buf, page, e, ntupSize, &nbuf, &npage, &freeOffno, &freeNeighborOffno, &firstFreePage))
/* Next, try space from a deleted element */
if (HnswFreeOffset(index, buf, page, e, ntupSize, &nbuf, &npage, &freeOffno, &freeNeighborOffno, &newInsertPage))
{
if (nbuf != buf)
npage = GenericXLogRegisterBuffer(state, nbuf, 0);
@@ -175,9 +177,17 @@ WriteNewElementPages(Relation index, HnswElement e, int m)
break;
}
insertPage = HnswPageGetOpaque(page)->nextblkno;
/* Finally, try space for element only if last page */
/* Skip if both tuples can fit on the same page */
if (combinedSize > maxSize && PageGetFreeSpace(page) >= etupSize && !BlockNumberIsValid(HnswPageGetOpaque(page)->nextblkno))
{
HnswInsertAppendPage(index, &nbuf, &npage, state, page);
break;
}
if (BlockNumberIsValid(insertPage))
currentPage = HnswPageGetOpaque(page)->nextblkno;
if (BlockNumberIsValid(currentPage))
{
/* Move to next page */
GenericXLogAbort(state);
@@ -191,8 +201,6 @@ WriteNewElementPages(Relation index, HnswElement e, int m)
HnswInsertAppendPage(index, &newbuf, &newpage, state, page);
/* Commit */
MarkBufferDirty(newbuf);
MarkBufferDirty(buf);
GenericXLogFinish(state);
/* Unlock previous buffer */
@@ -219,7 +227,10 @@ WriteNewElementPages(Relation index, HnswElement e, int m)
e->blkno = BufferGetBlockNumber(buf);
e->neighborPage = BufferGetBlockNumber(nbuf);
insertPage = e->neighborPage;
/* Added tuple to new page if newInsertPage is not set */
/* So can set to neighbor page instead of element page */
if (!BlockNumberIsValid(newInsertPage))
newInsertPage = e->neighborPage;
if (OffsetNumberIsValid(freeOffno))
{
@@ -256,84 +267,131 @@ WriteNewElementPages(Relation index, HnswElement e, int m)
}
/* Commit */
MarkBufferDirty(buf);
if (nbuf != buf)
MarkBufferDirty(nbuf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
if (nbuf != buf)
UnlockReleaseBuffer(nbuf);
/* Update the insert page */
if (insertPage != originalInsertPage && (!OffsetNumberIsValid(freeOffno) || firstFreePage == insertPage))
HnswUpdateMetaPage(index, false, NULL, insertPage, MAIN_FORKNUM);
if (BlockNumberIsValid(newInsertPage) && newInsertPage != insertPage)
*updatedInsertPage = newInsertPage;
}
/*
* Calculate index for update
* Check if connection already exists
*/
static int
HnswGetIndex(HnswUpdate * update, int m)
static bool
ConnectionExists(HnswElement e, HnswNeighborTuple ntup, int startIdx, int lm)
{
return (update->hc.element->level - update->level) * m + update->index;
for (int i = 0; i < lm; i++)
{
ItemPointer indextid = &ntup->indextids[startIdx + i];
if (!ItemPointerIsValid(indextid))
break;
if (ItemPointerGetBlockNumber(indextid) == e->blkno && ItemPointerGetOffsetNumber(indextid) == e->offno)
return true;
}
return false;
}
/*
* Update neighbors
*/
static void
UpdateNeighborPages(Relation index, HnswElement e, int m, List *updates)
void
HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement e, int m, bool checkExisting)
{
ListCell *lc;
/* Could update multiple at once for same element */
/* but should only happen a low percent of time, so keep simple for now */
foreach(lc, updates)
for (int lc = e->level; lc >= 0; lc--)
{
Buffer buf;
Page page;
GenericXLogState *state;
HnswUpdate *update = lfirst(lc);
ItemId itemid;
HnswNeighborTuple ntup;
Size ntupSize;
int idx;
OffsetNumber offno = update->hc.element->neighborOffno;
int lm = HnswGetLayerM(m, lc);
HnswNeighborArray *neighbors = &e->neighbors[lc];
/* Register page */
buf = ReadBuffer(index, update->hc.element->neighborPage);
LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
state = GenericXLogStart(index);
page = GenericXLogRegisterBuffer(state, buf, 0);
/* Get tuple */
itemid = PageGetItemId(page, offno);
ntup = (HnswNeighborTuple) PageGetItem(page, itemid);
ntupSize = ItemIdGetLength(itemid);
/* Calculate index */
idx = HnswGetIndex(update, m);
/* Make robust to issues */
if (idx < ntup->count)
for (int i = 0; i < neighbors->length; i++)
{
ItemPointer indextid = &ntup->indextids[idx];
HnswCandidate *hc = &neighbors->items[i];
Buffer buf;
Page page;
GenericXLogState *state;
ItemId itemid;
HnswNeighborTuple ntup;
Size ntupSize;
int idx = -1;
int startIdx;
OffsetNumber offno = hc->element->neighborOffno;
/* Update neighbor */
ItemPointerSet(indextid, e->blkno, e->offno);
/* Get latest neighbors since they may have changed */
/* Do not lock yet since selecting neighbors can take time */
HnswLoadNeighbors(hc->element, index, m);
/* Overwrite tuple */
if (!PageIndexTupleOverwrite(page, offno, (Item) ntup, ntupSize))
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
/*
* Could improve performance for vacuuming by checking neighbors
* against list of elements being deleted to find index. It's
* important to exclude already deleted elements for this since
* they can be replaced at any time.
*/
/* Commit */
MarkBufferDirty(buf);
GenericXLogFinish(state);
/* Select neighbors */
HnswUpdateConnection(e, hc, lm, lc, &idx, index, procinfo, collation);
/* New element was not selected as a neighbor */
if (idx == -1)
continue;
/* Register page */
buf = ReadBuffer(index, hc->element->neighborPage);
LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
state = GenericXLogStart(index);
page = GenericXLogRegisterBuffer(state, buf, 0);
/* Get tuple */
itemid = PageGetItemId(page, offno);
ntup = (HnswNeighborTuple) PageGetItem(page, itemid);
ntupSize = ItemIdGetLength(itemid);
/* Calculate index for update */
startIdx = (hc->element->level - lc) * m;
/* Check for existing connection */
if (checkExisting && ConnectionExists(e, ntup, startIdx, lm))
idx = -1;
else if (idx == -2)
{
/* Find free offset if still exists */
/* TODO Retry updating connections if not */
for (int j = 0; j < lm; j++)
{
if (!ItemPointerIsValid(&ntup->indextids[startIdx + j]))
{
idx = startIdx + j;
break;
}
}
}
else
idx += startIdx;
/* Make robust to issues */
if (idx >= 0 && idx < ntup->count)
{
ItemPointer indextid = &ntup->indextids[idx];
/* Update neighbor */
ItemPointerSet(indextid, e->blkno, e->offno);
/* Overwrite tuple */
if (!PageIndexTupleOverwrite(page, offno, (Item) ntup, ntupSize))
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
/* Commit */
GenericXLogFinish(state);
}
else
GenericXLogAbort(state);
UnlockReleaseBuffer(buf);
}
else
GenericXLogAbort(state);
UnlockReleaseBuffer(buf);
}
}
@@ -346,7 +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;
@@ -380,7 +438,6 @@ 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);
@@ -391,8 +448,10 @@ HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup)
* Write changes to disk
*/
static void
WriteElement(Relation index, HnswElement element, int m, List *updates, HnswElement dup, HnswElement entryPoint)
WriteElement(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement element, int m, int efConstruction, HnswElement dup, HnswElement entryPoint)
{
BlockNumber newInsertPage = InvalidBlockNumber;
/* Try to add to existing page */
if (dup != NULL)
{
@@ -400,13 +459,19 @@ WriteElement(Relation index, HnswElement element, int m, List *updates, HnswElem
return;
}
/* If fails, take this path */
WriteNewElementPages(index, element, m);
UpdateNeighborPages(index, element, m, updates);
/* Write element and neighbor tuples */
WriteNewElementPages(index, element, m, GetInsertPage(index), &newInsertPage);
/* Update insert page if needed */
if (BlockNumberIsValid(newInsertPage))
HnswUpdateMetaPage(index, 0, NULL, newInsertPage, MAIN_FORKNUM);
/* Update neighbors */
HnswUpdateNeighborPages(index, procinfo, collation, element, m, false);
/* Update metapage if needed */
if (entryPoint == NULL || element->level > entryPoint->level)
HnswUpdateMetaPage(index, true, element, InvalidBlockNumber, MAIN_FORKNUM);
HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_GREATER, element, InvalidBlockNumber, MAIN_FORKNUM);
}
/*
@@ -419,13 +484,12 @@ 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];
List *updates = NIL;
HnswElement dup;
LOCKMODE lockmode = ShareLock;
/* Detoast once for all calls */
value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
@@ -438,18 +502,45 @@ 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
* buffer lock (or reads when vacuuming).
*/
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->value = value;
/* Prevent concurrent inserts when likely updating entry point */
if (entryPoint == NULL || element->level > entryPoint->level)
{
/* Release shared lock */
UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
/* Get exclusive lock */
lockmode = ExclusiveLock;
LockPage(index, HNSW_UPDATE_LOCK, lockmode);
/* Get latest entry point after lock is acquired */
entryPoint = HnswGetEntryPoint(index);
}
/* Insert element in graph */
dup = HnswInsertElement(element, entryPoint, index, procinfo, collation, m, efConstruction, &updates, false);
HnswInsertElement(element, entryPoint, index, procinfo, collation, m, efConstruction, false);
/* Look for duplicate */
dup = HnswFindDuplicate(element);
/* Write to disk */
WriteElement(index, element, m, updates, dup, entryPoint);
WriteElement(index, procinfo, collation, element, m, efConstruction, dup, entryPoint);
/* Release lock */
UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
return true;
}

View File

@@ -4,34 +4,39 @@
#include "hnsw.h"
#include "pgstat.h"
#include "storage/bufmgr.h"
#include "storage/lmgr.h"
#include "utils/memutils.h"
/*
* Algorithm 5 from paper
*/
static void
static List *
GetScanItems(IndexScanDesc scan, Datum q)
{
HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
Relation index = scan->indexRelation;
FmgrInfo *procinfo = so->procinfo;
Oid collation = so->collation;
List *ep = NIL;
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;
return NIL;
ep = lappend(ep, HnswEntryCandidate(entryPoint, q, index, procinfo, collation, false));
ep = list_make1(HnswEntryCandidate(entryPoint, q, index, procinfo, collation, false));
for (int lc = entryPoint->level; lc >= 1; lc--)
{
w = HnswSearchLayer(q, ep, 1, lc, index, procinfo, collation, false, NULL, NULL);
w = HnswSearchLayer(q, ep, 1, lc, index, procinfo, collation, m, false, NULL);
ep = w;
}
so->w = HnswSearchLayer(q, ep, hnsw_ef_search, 0, index, procinfo, collation, false, NULL, NULL);
return HnswSearchLayer(q, ep, hnsw_ef_search, 0, index, procinfo, collation, m, false, NULL);
}
/*
@@ -57,6 +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
*/
@@ -69,7 +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",
@@ -129,62 +160,52 @@ 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;
/* 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");
/* Value should not be compressed or toasted */
Assert(!VARATT_IS_COMPRESSED(DatumGetPointer(value)));
Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));
/* Get scan value */
value = GetScanValue(scan);
/* Fine if normalization fails */
if (so->normprocinfo != NULL)
HnswNormValue(so->normprocinfo, so->collation, &value, NULL);
}
/*
* Get a shared lock. This allows vacuum to ensure no in-flight scans
* before marking tuples as deleted.
*/
LockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
so->w = GetScanItems(scan, value);
/* Release shared lock */
UnlockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
GetScanItems(scan, value);
so->first = false;
}
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
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;
}
@@ -201,10 +222,6 @@ hnswendscan(IndexScanDesc scan)
{
HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
/* Release pin */
if (BufferIsValid(so->buf))
ReleaseBuffer(so->buf);
MemoryContextDelete(so->tmpCtx);
pfree(so);

View File

@@ -4,10 +4,11 @@
#include "hnsw.h"
#include "storage/bufmgr.h"
#include "utils/datum.h"
#include "vector.h"
/*
* Get the number of connection in the index
* Get the max number of connections in an upper layer for each element in the index
*/
int
HnswGetM(Relation index)
@@ -38,12 +39,12 @@ 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 +118,6 @@ HnswInitRegisterPage(Relation index, Buffer *buf, Page *page, GenericXLogState *
void
HnswCommitBuffer(Buffer buf, GenericXLogState *state)
{
MarkBufferDirty(buf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
}
@@ -140,9 +140,21 @@ 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
*/
@@ -174,11 +186,10 @@ 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);
if (element->loaded)
pfree(DatumGetPointer(element->value));
pfree(element);
}
@@ -197,59 +208,67 @@ HnswAddHeapTid(HnswElement element, ItemPointer heaptid)
/*
* Allocate an element from block and offset numbers
*/
static HnswElement
InitElementFromBlock(BlockNumber blkno, OffsetNumber offno)
HnswElement
HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno)
{
HnswElement element = palloc(sizeof(HnswElementData));
element->blkno = blkno;
element->offno = offno;
element->neighbors = NULL;
element->vec = NULL;
element->loaded = false;
return element;
}
/*
* Get the metapage info
*/
void
HnswGetMetaPageInfo(Relation index, int *m, HnswElement * entryPoint)
{
Buffer buf;
Page page;
HnswMetaPage metap;
buf = ReadBuffer(index, HNSW_METAPAGE_BLKNO);
LockBuffer(buf, BUFFER_LOCK_SHARE);
page = BufferGetPage(buf);
metap = HnswPageGetMeta(page);
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)
{
Buffer buf;
Page page;
HnswMetaPage metap;
HnswElement entryPoint = NULL;
HnswElement entryPoint;
buf = ReadBuffer(index, HNSW_METAPAGE_BLKNO);
LockBuffer(buf, BUFFER_LOCK_SHARE);
page = BufferGetPage(buf);
metap = HnswPageGetMeta(page);
if (BlockNumberIsValid(metap->entryBlkno))
entryPoint = InitElementFromBlock(metap->entryBlkno, metap->entryOffno);
UnlockReleaseBuffer(buf);
HnswGetMetaPageInfo(index, NULL, &entryPoint);
return entryPoint;
}
/*
* Update the metapage
* Update the metapage info
*/
void
HnswUpdateMetaPage(Relation index, bool updateEntry, HnswElement entryPoint, BlockNumber insertPage, ForkNumber forkNum)
static void
HnswUpdateMetaPageInfo(Page page, int updateEntry, HnswElement entryPoint, BlockNumber insertPage)
{
Buffer buf;
Page page;
GenericXLogState *state;
HnswMetaPage metap;
buf = ReadBufferExtended(index, forkNum, HNSW_METAPAGE_BLKNO, RBM_NORMAL, NULL);
LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
state = GenericXLogStart(index);
page = GenericXLogRegisterBuffer(state, buf, 0);
metap = HnswPageGetMeta(page);
HnswMetaPage metap = HnswPageGetMeta(page);
if (updateEntry)
{
@@ -259,7 +278,7 @@ HnswUpdateMetaPage(Relation index, bool updateEntry, HnswElement entryPoint, Blo
metap->entryOffno = InvalidOffsetNumber;
metap->entryLevel = -1;
}
else
else if (entryPoint->level > metap->entryLevel || updateEntry == HNSW_UPDATE_ENTRY_ALWAYS)
{
metap->entryBlkno = entryPoint->blkno;
metap->entryOffno = entryPoint->offno;
@@ -269,6 +288,24 @@ HnswUpdateMetaPage(Relation index, bool updateEntry, HnswElement entryPoint, Blo
if (BlockNumberIsValid(insertPage))
metap->insertPage = insertPage;
}
/*
* Update the metapage
*/
void
HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, BlockNumber insertPage, ForkNumber forkNum)
{
Buffer buf;
Page page;
GenericXLogState *state;
buf = ReadBufferExtended(index, forkNum, HNSW_METAPAGE_BLKNO, RBM_NORMAL, NULL);
LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
state = GenericXLogStart(index);
page = GenericXLogRegisterBuffer(state, buf, 0);
HnswUpdateMetaPageInfo(page, updateEntry, entryPoint, insertPage);
HnswCommitBuffer(buf, state);
}
@@ -289,7 +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));
}
/*
@@ -329,10 +366,9 @@ 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));
@@ -356,7 +392,7 @@ LoadNeighborsFromPage(HnswElement element, Relation index, Page page)
if (!ItemPointerIsValid(indextid))
continue;
e = InitElementFromBlock(ItemPointerGetBlockNumber(indextid), ItemPointerGetOffsetNumber(indextid));
e = HnswInitElementFromBlock(ItemPointerGetBlockNumber(indextid), ItemPointerGetOffsetNumber(indextid));
/* Calculate level based on offset */
level = element->level - i / m;
@@ -372,8 +408,8 @@ LoadNeighborsFromPage(HnswElement element, Relation index, Page page)
/*
* Load neighbors
*/
static void
LoadNeighbors(HnswElement element, Relation index)
void
HnswLoadNeighbors(HnswElement element, Relation index, int m)
{
Buffer buf;
Page page;
@@ -382,11 +418,44 @@ LoadNeighbors(HnswElement element, Relation index)
LockBuffer(buf, BUFFER_LOCK_SHARE);
page = BufferGetPage(buf);
LoadNeighborsFromPage(element, index, page);
LoadNeighborsFromPage(element, index, page, m);
UnlockReleaseBuffer(buf);
}
/*
* Load an element from a tuple
*/
void
HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec)
{
element->level = etup->level;
element->deleted = etup->deleted;
element->neighborPage = ItemPointerGetBlockNumber(&etup->neighbortid);
element->neighborOffno = ItemPointerGetOffsetNumber(&etup->neighbortid);
element->heaptids = NIL;
if (loadHeaptids)
{
for (int i = 0; i < HNSW_HEAPTIDS; i++)
{
/* Can stop at first invalid */
if (!ItemPointerIsValid(&etup->heaptids[i]))
break;
HnswAddHeapTid(element, &etup->heaptids[i]);
}
}
if (loadVec)
{
Datum value = PointerGetDatum(&etup->value);
element->value = datumCopy(value, false, -1);
element->loaded = true;
}
}
/*
* Load an element and optionally get its distance from q
*/
@@ -407,29 +476,11 @@ HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index,
Assert(HnswIsElementTuple(etup));
/* Load element */
element->heaptids = NIL;
for (int i = 0; i < HNSW_HEAPTIDS; i++)
{
/* Can stop at first invalid */
if (!ItemPointerIsValid(&etup->heaptids[i]))
break;
HnswAddHeapTid(element, &etup->heaptids[i]);
}
element->level = etup->level;
element->neighborPage = ItemPointerGetBlockNumber(&etup->neighbortid);
element->neighborOffno = ItemPointerGetOffsetNumber(&etup->neighbortid);
element->deleted = etup->deleted;
if (loadVec)
{
element->vec = palloc(VECTOR_SIZE(etup->vec.dim));
memcpy(element->vec, &etup->vec, VECTOR_SIZE(etup->vec.dim));
}
HnswLoadElementFromTuple(element, etup, true, loadVec);
/* 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);
}
@@ -440,14 +491,14 @@ 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));
}
/*
* Create a candidate for the entry point
*/
HnswCandidate *
HnswEntryCandidate(HnswElement entryPoint, Datum q, Relation index, FmgrInfo *procinfo, Oid collation, bool loadvec)
HnswEntryCandidate(HnswElement entryPoint, Datum q, Relation index, FmgrInfo *procinfo, Oid collation, bool loadVec)
{
HnswCandidate *hc = palloc(sizeof(HnswCandidate));
@@ -455,7 +506,7 @@ HnswEntryCandidate(HnswElement entryPoint, Datum q, Relation index, FmgrInfo *pr
if (index == NULL)
hc->distance = GetCandidateDistance(hc, q, procinfo, collation);
else
HnswLoadElement(hc->element, &hc->distance, &q, index, procinfo, collation, loadvec);
HnswLoadElement(hc->element, &hc->distance, &q, index, procinfo, collation, loadVec);
return hc;
}
@@ -522,7 +573,7 @@ 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, BlockNumber *skipPage, OffsetNumber *skipOffno)
HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement)
{
ListCell *lc2;
@@ -558,7 +609,13 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
pairingheap_add(C, &(CreatePairingHeapNode(hc)->ph_node));
pairingheap_add(W, &(CreatePairingHeapNode(hc)->ph_node));
wlen++;
/*
* Do not count elements being deleted towards ef when vacuuming. It
* would be ideal to do this for inserts as well, but this could
* affect insert performance.
*/
if (skipElement == NULL || list_length(hc->element->heaptids) != 0)
wlen++;
}
while (!pairingheap_is_empty(C))
@@ -571,7 +628,7 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
break;
if (c->element->neighbors == NULL)
LoadNeighbors(c->element, index);
HnswLoadNeighbors(c->element, index, m);
/* Get the neighborhood at layer lc */
neighborhood = &c->element->neighbors[lc];
@@ -594,17 +651,7 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
else
HnswLoadElement(e->element, &eDistance, &q, index, procinfo, collation, inserting);
/* Skip if fully deleted */
if (e->element->deleted)
continue;
/* Skip for inserts if deleting */
if (inserting && list_length(e->element->heaptids) == 0)
continue;
/* Skip self for vacuuming update */
if (skipPage != NULL && e->element->neighborPage == *skipPage && e->element->neighborOffno == *skipOffno)
continue;
Assert(!e->element->deleted);
/* Make robust to issues */
if (e->element->level < lc)
@@ -620,11 +667,20 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
pairingheap_add(C, &(CreatePairingHeapNode(ec)->ph_node));
pairingheap_add(W, &(CreatePairingHeapNode(ec)->ph_node));
wlen++;
/* No need to decrement wlen */
if (wlen > ef)
pairingheap_remove_first(W);
/*
* Do not count elements being deleted towards ef when
* vacuuming. It would be ideal to do this for inserts as
* well, but this could affect insert performance.
*/
if (skipElement == NULL || list_length(e->element->heaptids) != 0)
{
wlen++;
/* No need to decrement wlen */
if (wlen > ef)
pairingheap_remove_first(W);
}
}
}
}
@@ -641,6 +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
*/
@@ -670,7 +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));
}
/*
@@ -697,33 +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)
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);
@@ -743,17 +871,17 @@ SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswC
/*
* Find duplicate element
*/
static HnswElement
HnswFindDuplicate(HnswElement e, List *neighbors)
HnswElement
HnswFindDuplicate(HnswElement e)
{
ListCell *lc;
HnswNeighborArray *neighbors = &e->neighbors[0];
foreach(lc, neighbors)
for (int i = 0; i < neighbors->length; i++)
{
HnswCandidate *neighbor = lfirst(lc);
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 */
@@ -777,205 +905,172 @@ 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;
}
/*
* Create update
*/
static HnswUpdate *
CreateUpdate(HnswCandidate * hc, int level, int index)
{
HnswUpdate *update = palloc(sizeof(HnswUpdate));
update->hc = *hc;
update->level = level;
update->index = index;
return update;
}
/*
* Update connections
*/
static void
UpdateConnections(HnswElement element, List *neighbors, int m, int lc, List **updates, Relation index, FmgrInfo *procinfo, Oid collation)
void
HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int *updateIdx, Relation index, FmgrInfo *procinfo, Oid collation)
{
ListCell *lc2;
HnswNeighborArray *currentNeighbors = &hc->element->neighbors[lc];
foreach(lc2, neighbors)
HnswCandidate hc2;
hc2.element = element;
hc2.distance = hc->distance;
if (currentNeighbors->length < m)
{
HnswCandidate *hc = (HnswCandidate *) lfirst(lc2);
HnswNeighborArray *currentNeighbors = &hc->element->neighbors[lc];
currentNeighbors->items[currentNeighbors->length++] = hc2;
HnswCandidate hc2;
/* Track update */
if (updateIdx != NULL)
*updateIdx = -2;
}
else
{
/* Shrink connections */
HnswCandidate *pruned = NULL;
hc2.element = element;
hc2.distance = hc->distance;
if (currentNeighbors->length < m)
/* Load elements on insert */
if (index != NULL)
{
currentNeighbors->items[currentNeighbors->length++] = hc2;
Datum q = hc->element->value;
/* Track updates */
if (updates != NULL)
*updates = lappend(*updates, CreateUpdate(hc, lc, currentNeighbors->length - 1));
}
else
{
/* Shrink connections */
HnswCandidate *pruned = NULL;
List *c = NIL;
/* Load elements on insert */
if (index != NULL)
{
Datum q = PointerGetDatum(hc->element->vec);
for (int i = 0; i < currentNeighbors->length; i++)
{
HnswCandidate *hc3 = &currentNeighbors->items[i];
if (hc3->element->vec == NULL)
HnswLoadElement(hc3->element, &hc3->distance, &q, index, procinfo, collation, true);
else
hc3->distance = GetCandidateDistance(hc3, q, procinfo, collation);
/* Prune element if being deleted */
if (list_length(hc3->element->heaptids) == 0)
{
pruned = &currentNeighbors->items[i];
break;
}
}
}
if (pruned == NULL)
{
/* Add and sort 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);
/* Should not happen */
if (pruned == NULL)
continue;
}
/* Find and replace the pruned element */
for (int i = 0; i < currentNeighbors->length; i++)
{
if (currentNeighbors->items[i].element == pruned->element)
HnswCandidate *hc3 = &currentNeighbors->items[i];
if (!hc3->element->loaded)
HnswLoadElement(hc3->element, &hc3->distance, &q, index, procinfo, collation, true);
else
hc3->distance = GetCandidateDistance(hc3, q, procinfo, collation);
/* Prune element if being deleted */
if (list_length(hc3->element->heaptids) == 0)
{
currentNeighbors->items[i] = hc2;
/* Track updates */
if (updates != NULL)
*updates = lappend(*updates, CreateUpdate(hc, lc, i));
pruned = &currentNeighbors->items[i];
break;
}
}
}
if (pruned == NULL)
{
List *c = NIL;
/* Add candidates */
for (int i = 0; i < currentNeighbors->length; i++)
c = lappend(c, &currentNeighbors->items[i]);
c = lappend(c, &hc2);
SelectNeighbors(c, m, lc, procinfo, collation, hc->element, &hc2, &pruned, true);
/* Should not happen */
if (pruned == NULL)
return;
}
/* Find and replace the pruned element */
for (int i = 0; i < currentNeighbors->length; i++)
{
if (currentNeighbors->items[i].element == pruned->element)
{
currentNeighbors->items[i] = hc2;
/* Track update */
if (updateIdx != NULL)
*updateIdx = i;
break;
}
}
}
}
/*
* Remove elements being deleted or skipped
*/
static List *
RemoveElements(List *w, HnswElement skipElement)
{
ListCell *lc2;
List *w2 = NIL;
foreach(lc2, w)
{
HnswCandidate *hc = (HnswCandidate *) lfirst(lc2);
/* Skip self for vacuuming update */
if (skipElement != NULL && hc->element->blkno == skipElement->blkno && hc->element->offno == skipElement->offno)
continue;
if (list_length(hc->element->heaptids) != 0)
w2 = lappend(w2, hc);
}
return w2;
}
/*
* Algorithm 1 from paper
*/
HnswElement
HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo *procinfo, Oid collation, int m, int efConstruction, List **updates, bool vacuuming)
void
HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo *procinfo, Oid collation, int m, int efConstruction, bool existing)
{
List *ep = NIL;
List *ep;
List *w;
int level = element->level;
int entryLevel;
List **newNeighbors = palloc(sizeof(List *) * (level + 1));
Datum q = PointerGetDatum(element->vec);
HnswElement dup;
BlockNumber *skipPage = vacuuming ? &element->neighborPage : NULL;
OffsetNumber *skipOffno = vacuuming ? &element->neighborOffno : NULL;
bool removeEntryPoint;
HnswCandidate *entryCandidate;
Datum q = element->value;
HnswElement skipElement = existing ? element : NULL;
/* No neighbors if no entry point */
if (entryPoint == NULL)
return;
/* Get entry point and level */
if (entryPoint != NULL)
{
entryCandidate = HnswEntryCandidate(entryPoint, q, index, procinfo, collation, true);
ep = lappend(ep, entryCandidate);
entryLevel = entryPoint->level;
removeEntryPoint = vacuuming && list_length(entryPoint->heaptids) == 0;
}
else
{
entryLevel = -1;
removeEntryPoint = false;
}
ep = list_make1(HnswEntryCandidate(entryPoint, q, index, procinfo, collation, true));
entryLevel = entryPoint->level;
/* 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, skipPage, skipOffno);
w = HnswSearchLayer(q, ep, 1, lc, index, procinfo, collation, m, true, skipElement);
ep = w;
}
if (level > entryLevel)
level = entryLevel;
/* Add one for existing element */
if (existing)
efConstruction++;
/* 2nd phase */
for (int lc = level; lc >= 0; lc--)
{
int lm = HnswGetLayerM(m, lc);
List *neighbors;
List *lw;
w = HnswSearchLayer(q, ep, efConstruction, lc, index, procinfo, collation, true, skipPage, skipOffno);
w = HnswSearchLayer(q, ep, efConstruction, lc, index, procinfo, collation, m, true, skipElement);
/* Remove entry point if it's being deleted */
if (removeEntryPoint)
w = list_delete_ptr(w, entryCandidate);
/* Elements being deleted or skipped can help with search */
/* but should be removed before selecting neighbors */
if (index != NULL)
lw = RemoveElements(w, skipElement);
else
lw = w;
/*
* 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);
newNeighbors[lc] = SelectNeighbors(w, lm, lc, procinfo, collation, NULL);
ep = w;
}
/* Look for duplicate */
if (level >= 0 && !vacuuming)
{
dup = HnswFindDuplicate(element, newNeighbors[0]);
if (dup != NULL)
return dup;
}
/* Update connections */
for (int lc = level; lc >= 0; lc--)
{
int lm = HnswGetLayerM(m, lc);
AddConnections(element, newNeighbors[lc], lm, lc);
if (!vacuuming)
UpdateConnections(element, newNeighbors[lc], lm, lc, updates, index, procinfo, collation);
}
return NULL;
}

View File

@@ -5,10 +5,11 @@
#include "commands/vacuum.h"
#include "hnsw.h"
#include "storage/bufmgr.h"
#include "storage/lmgr.h"
#include "utils/memutils.h"
/*
* Check if deleted list contains an index tid
* Check if deleted list contains an index TID
*/
static bool
DeletedContains(HTAB *deleted, ItemPointer indextid)
@@ -32,6 +33,7 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
Relation index = vacuumstate->index;
BufferAccessStrategy bas = vacuumstate->bas;
HnswElement entryPoint = HnswGetEntryPoint(vacuumstate->index);
IndexBulkDeleteResult *stats = vacuumstate->stats;
/* Store separately since highestPoint.level is uint8 */
int highestLevel = -1;
@@ -77,17 +79,21 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
break;
if (vacuumstate->callback(&etup->heaptids[i], vacuumstate->callback_state))
{
itemUpdated = true;
stats->tuples_removed++;
}
else
{
/* Move to front of list */
etup->heaptids[idx++] = etup->heaptids[i];
stats->num_index_tuples++;
}
}
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++)
@@ -109,7 +115,7 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
(void) hash_search(vacuumstate->deleted, &ip, HASH_ENTER, NULL);
}
else if (etup->level > highestLevel && !(blkno == entryPoint->blkno && offno == entryPoint->offno))
else if (etup->level > highestLevel && !(entryPoint != NULL && blkno == entryPoint->blkno && offno == entryPoint->offno))
{
/* Keep track of highest non-entry point */
highestPoint->blkno = blkno;
@@ -122,10 +128,7 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
blkno = HnswPageGetOpaque(page)->nextblkno;
if (updated)
{
MarkBufferDirty(buf);
GenericXLogFinish(state);
}
else
GenericXLogAbort(state);
@@ -169,6 +172,11 @@ NeedsUpdated(HnswVacuumState * vacuumstate, HnswElement element)
}
}
/* Also update if layer 0 is not full */
/* This could indicate too many candidates being deleted during insert */
if (!needsUpdated)
needsUpdated = !ItemPointerIsValid(&ntup->indextids[ntup->count - 1]);
UnlockReleaseBuffer(buf);
return needsUpdated;
@@ -178,7 +186,7 @@ NeedsUpdated(HnswVacuumState * vacuumstate, HnswElement element)
* Repair graph for a single element
*/
static void
RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element)
RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswElement entryPoint)
{
Relation index = vacuumstate->index;
Buffer buf;
@@ -188,42 +196,20 @@ RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element)
int efConstruction = vacuumstate->efConstruction;
FmgrInfo *procinfo = vacuumstate->procinfo;
Oid collation = vacuumstate->collation;
HnswElement entryPoint;
BufferAccessStrategy bas = vacuumstate->bas;
HnswNeighborTuple ntup = vacuumstate->ntup;
Size ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(element->level, m);
/* Check if any neighbors point to deleted values */
if (!NeedsUpdated(vacuumstate, element))
/* Skip if element is entry point */
if (entryPoint != NULL && element->blkno == entryPoint->blkno && element->offno == entryPoint->offno)
return;
/* Refresh entry point for each element */
entryPoint = HnswGetEntryPoint(index);
/* Special case for entry point */
if (element->blkno == entryPoint->blkno && element->offno == entryPoint->offno)
{
if (BlockNumberIsValid(vacuumstate->highestPoint.blkno))
{
/* Already updated */
if (vacuumstate->highestPoint.blkno == element->blkno && vacuumstate->highestPoint.offno == element->offno)
return;
entryPoint = &vacuumstate->highestPoint;
/* Reset neighbors from previous update */
entryPoint->neighbors = NULL;
}
else
entryPoint = NULL;
}
/* Init fields */
HnswInitNeighbors(element, m);
element->heaptids = NIL;
/* Add element to graph, skipping itself */
HnswInsertElement(element, entryPoint, index, procinfo, collation, m, efConstruction, NULL, true);
HnswInsertElement(element, entryPoint, index, procinfo, collation, m, efConstruction, true);
/* Update neighbor tuple */
/* Do this before getting page to minimize locking */
@@ -240,9 +226,11 @@ RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element)
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
/* Commit */
MarkBufferDirty(buf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
/* Update neighbors */
HnswUpdateNeighborPages(index, procinfo, collation, element, m, true);
}
/*
@@ -256,16 +244,35 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
HnswElement entryPoint;
MemoryContext oldCtx = MemoryContextSwitchTo(vacuumstate->tmpCtx);
/* Repair graph for highest non-entry point */
/* This may not be the highest with new inserts, but should be fine */
if (BlockNumberIsValid(highestPoint->blkno))
if (!BlockNumberIsValid(highestPoint->blkno))
highestPoint = NULL;
/*
* Repair graph for highest non-entry point. Highest point may be outdated
* due to inserts that happen during and after RemoveHeapTids.
*/
if (highestPoint != NULL)
{
/* Get a shared lock */
LockPage(index, HNSW_UPDATE_LOCK, ShareLock);
/* Load element */
HnswLoadElement(highestPoint, NULL, NULL, index, vacuumstate->procinfo, vacuumstate->collation, true);
RepairGraphElement(vacuumstate, highestPoint);
/* Repair if needed */
if (NeedsUpdated(vacuumstate, highestPoint))
RepairGraphElement(vacuumstate, highestPoint, HnswGetEntryPoint(index));
/* Release lock */
UnlockPage(index, HNSW_UPDATE_LOCK, ShareLock);
}
/* See if entry point needs updated */
/* Prevent concurrent inserts when possibly updating entry point */
LockPage(index, HNSW_UPDATE_LOCK, ExclusiveLock);
/* Get latest entry point */
entryPoint = HnswGetEntryPoint(index);
if (entryPoint != NULL)
{
ItemPointerData epData;
@@ -273,15 +280,37 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
ItemPointerSet(&epData, entryPoint->blkno, entryPoint->offno);
if (DeletedContains(vacuumstate->deleted, &epData))
HnswUpdateMetaPage(index, true, highestPoint, InvalidBlockNumber, MAIN_FORKNUM);
{
/*
* Replace the entry point with the highest point. If highest
* point is outdated and empty, the entry point will be empty
* until an element is repaired.
*/
HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_ALWAYS, highestPoint, InvalidBlockNumber, MAIN_FORKNUM);
}
else
{
/* Highest point will be used to repair */
/*
* Repair the entry point with the highest point. If highest point
* is outdated, this can remove connections at higher levels in
* the graph until they are repaired, but this should be fine.
*/
HnswLoadElement(entryPoint, NULL, NULL, index, vacuumstate->procinfo, vacuumstate->collation, true);
RepairGraphElement(vacuumstate, entryPoint);
if (NeedsUpdated(vacuumstate, entryPoint))
{
/* Reset neighbors from previous update */
if (highestPoint != NULL)
highestPoint->neighbors = NULL;
RepairGraphElement(vacuumstate, entryPoint, highestPoint);
}
}
}
/* Release lock */
UnlockPage(index, HNSW_UPDATE_LOCK, ExclusiveLock);
/* Reset memory context */
MemoryContextSwitchTo(oldCtx);
MemoryContextReset(vacuumstate->tmpCtx);
@@ -297,6 +326,14 @@ RepairGraph(HnswVacuumState * vacuumstate)
BufferAccessStrategy bas = vacuumstate->bas;
BlockNumber blkno = HNSW_HEAD_BLKNO;
/*
* 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);
/* Repair entry point first */
RepairGraphEntryPoint(vacuumstate);
while (BlockNumberIsValid(blkno))
@@ -333,14 +370,8 @@ RepairGraph(HnswVacuumState * vacuumstate)
continue;
/* Create an element */
element = palloc(sizeof(HnswElementData));
element->neighborPage = ItemPointerGetBlockNumber(&etup->neighbortid);
element->neighborOffno = ItemPointerGetOffsetNumber(&etup->neighbortid);
element->level = etup->level;
element->blkno = blkno;
element->offno = offno;
element->vec = palloc(VECTOR_SIZE(etup->vec.dim));
memcpy(element->vec, &etup->vec, VECTOR_SIZE(etup->vec.dim));
element = HnswInitElementFromBlock(blkno, offno);
HnswLoadElementFromTuple(element, etup, false, true);
elements = lappend(elements, element);
}
@@ -351,7 +382,48 @@ RepairGraph(HnswVacuumState * vacuumstate)
/* Update neighbor pages */
foreach(lc2, elements)
RepairGraphElement(vacuumstate, (HnswElement) lfirst(lc2));
{
HnswElement element = (HnswElement) lfirst(lc2);
HnswElement entryPoint;
LOCKMODE lockmode = ShareLock;
/* Check if any neighbors point to deleted values */
if (!NeedsUpdated(vacuumstate, element))
continue;
/* Get a shared lock */
LockPage(index, HNSW_UPDATE_LOCK, lockmode);
/* Refresh entry point for each element */
entryPoint = HnswGetEntryPoint(index);
/* Prevent concurrent inserts when likely updating entry point */
if (entryPoint == NULL || element->level > entryPoint->level)
{
/* Release shared lock */
UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
/* Get exclusive lock */
lockmode = ExclusiveLock;
LockPage(index, HNSW_UPDATE_LOCK, lockmode);
/* Get latest entry point after lock is acquired */
entryPoint = HnswGetEntryPoint(index);
}
/* Repair connections */
RepairGraphElement(vacuumstate, element, entryPoint);
/*
* Update metapage if needed. Should only happen if entry point
* was replaced and highest point was outdated.
*/
if (entryPoint == NULL || element->level > entryPoint->level)
HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_GREATER, element, InvalidBlockNumber, MAIN_FORKNUM);
/* Release lock */
UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
}
/* Reset memory context */
MemoryContextSwitchTo(oldCtx);
@@ -369,7 +441,14 @@ MarkDeleted(HnswVacuumState * vacuumstate)
BlockNumber insertPage = InvalidBlockNumber;
Relation index = vacuumstate->index;
BufferAccessStrategy bas = vacuumstate->bas;
IndexBulkDeleteResult *stats = vacuumstate->stats;
/*
* 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);
while (BlockNumberIsValid(blkno))
{
@@ -402,6 +481,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
HnswNeighborTuple ntup;
Size etupSize;
Size ntupSize;
Datum value;
Buffer nbuf;
Page npage;
BlockNumber neighborPage;
@@ -413,20 +493,23 @@ MarkDeleted(HnswVacuumState * vacuumstate)
/* Skip deleted tuples */
if (etup->deleted)
continue;
/* Skip live tuples */
if (ItemPointerIsValid(&etup->heaptids[0]))
{
stats->num_index_tuples++;
/* Set to first free page */
if (!BlockNumberIsValid(insertPage))
insertPage = blkno;
continue;
}
/* Update stats */
stats->tuples_removed++;
/* Skip live tuples */
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 */
@@ -449,7 +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++)
@@ -464,9 +547,6 @@ 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);
@@ -486,7 +566,8 @@ MarkDeleted(HnswVacuumState * vacuumstate)
UnlockReleaseBuffer(buf);
}
HnswUpdateMetaPage(index, false, NULL, insertPage, MAIN_FORKNUM);
/* Update insert page last, after everything has been marked as deleted */
HnswUpdateMetaPage(index, 0, NULL, insertPage, MAIN_FORKNUM);
}
/*
@@ -505,7 +586,6 @@ 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);
@@ -515,6 +595,9 @@ 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

@@ -24,7 +24,7 @@ IvfflatInit(void)
{
ivfflat_relopt_kind = add_reloption_kind();
add_int_reloption(ivfflat_relopt_kind, "lists", "Number of inverted lists",
IVFFLAT_DEFAULT_LISTS, 1, IVFFLAT_MAX_LISTS
IVFFLAT_DEFAULT_LISTS, IVFFLAT_MIN_LISTS, IVFFLAT_MAX_LISTS
#if PG_VERSION_NUM >= 130000
,AccessExclusiveLock
#endif
@@ -32,7 +32,7 @@ IvfflatInit(void)
DefineCustomIntVariable("ivfflat.probes", "Sets the number of probes",
"Valid range is 1..lists.", &ivfflat_probes,
1, 1, IVFFLAT_MAX_LISTS, PGC_USERSET, 0, NULL, NULL, NULL);
IVFFLAT_DEFAULT_PROBES, IVFFLAT_MIN_LISTS, IVFFLAT_MAX_LISTS, PGC_USERSET, 0, NULL, NULL, NULL);
}
/*
@@ -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

@@ -40,8 +40,11 @@
#define IVFFLAT_METAPAGE_BLKNO 0
#define IVFFLAT_HEAD_BLKNO 1 /* first list page */
/* IVFFlat parameters */
#define IVFFLAT_DEFAULT_LISTS 100
#define IVFFLAT_MIN_LISTS 1
#define IVFFLAT_MAX_LISTS 32768
#define IVFFLAT_DEFAULT_PROBES 1
/* Build phases */
/* PROGRESS_CREATEIDX_SUBPHASE_INITIALIZE is 1 */
@@ -241,8 +244,8 @@ typedef struct IvfflatScanList
typedef struct IvfflatScanOpaqueData
{
int probes;
int dimensions;
bool first;
Buffer buf;
/* Sorting */
Tuplesortstate *sortstate;
@@ -272,9 +275,10 @@ 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;
@@ -80,33 +81,33 @@ 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 */
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 (;;)
{
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 +122,28 @@ 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);
itup = (IndexTuple) PageGetItem(page, itemid);
datum = index_getattr(itup, 1, tupdesc, &isnull);
/*
@@ -154,8 +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);
@@ -180,29 +181,6 @@ GetScanItems(IndexScanDesc scan, Datum value)
tuplesort_performsort(so->sortstate);
}
/*
* Get dimensions from metapage
*/
static int
GetDimensions(Relation index)
{
Buffer buf;
Page page;
IvfflatMetaPage metap;
int dimensions;
buf = ReadBuffer(index, IVFFLAT_METAPAGE_BLKNO);
LockBuffer(buf, BUFFER_LOCK_SHARE);
page = BufferGetPage(buf);
metap = IvfflatPageGetMeta(page);
dimensions = metap->dimensions;
UnlockReleaseBuffer(buf);
return dimensions;
}
/*
* Prepare for an index scan
*/
@@ -212,6 +190,7 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
IndexScanDesc scan;
IvfflatScanOpaque so;
int lists;
int dimensions;
AttrNumber attNums[] = {1};
Oid sortOperators[] = {Float8LessOperator};
Oid sortCollations[] = {InvalidOid};
@@ -219,15 +198,17 @@ 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;
so = (IvfflatScanOpaque) palloc(offsetof(IvfflatScanOpaqueData, lists) + probes * sizeof(IvfflatScanList));
so->buf = InvalidBuffer;
so->first = true;
so->probes = probes;
so->dimensions = dimensions;
/* Set support functions */
so->procinfo = index_getprocinfo(index, 1, IVFFLAT_DISTANCE_PROC);
@@ -236,13 +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, "tid", TIDOID, -1, 0);
TupleDescInitEntry(so->tupdesc, (AttrNumber) 3, "indexblkno", INT4OID, -1, 0);
TupleDescInitEntry(so->tupdesc, (AttrNumber) 2, "heaptid", TIDOID, -1, 0);
/* Prep sort */
so->sortstate = tuplesort_begin_heap(so->tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, work_mem, NULL, false);
@@ -308,8 +288,13 @@ 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(GetDimensions(scan->indexRelation)));
value = PointerGetDatum(InitVector(so->dimensions));
else
{
value = scan->orderByData->sk_argument;
@@ -334,26 +319,14 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
if (tuplesort_gettupleslot(so->sortstate, true, false, so->slot, NULL))
{
ItemPointer tid = (ItemPointer) DatumGetPointer(slot_getattr(so->slot, 2, &so->isnull));
BlockNumber indexblkno = DatumGetInt32(slot_getattr(so->slot, 3, &so->isnull));
ItemPointer heaptid = (ItemPointer) DatumGetPointer(slot_getattr(so->slot, 2, &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
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;
}
@@ -369,10 +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);

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

@@ -15,6 +15,10 @@
#include "utils/numeric.h"
#include "vector.h"
#if PG_VERSION_NUM >= 160000
#include "varatt.h"
#endif
#if PG_VERSION_NUM >= 120000
#include "common/shortest_dec.h"
#include "utils/float.h"
@@ -35,6 +39,7 @@ PG_MODULE_MAGIC;
/*
* Initialize index options and variables
*/
PGDLLEXPORT void _PG_init(void);
void
_PG_init(void)
{
@@ -100,6 +105,23 @@ CheckElement(float value)
errmsg("infinite value not allowed in vector")));
}
/*
* Allocate and initialize a new vector
*/
Vector *
InitVector(int dim)
{
Vector *result;
int size;
size = VECTOR_SIZE(dim);
result = (Vector *) palloc0(size);
SET_VARSIZE(result, size);
result->dim = dim;
return result;
}
/*
* Check for whitespace, since array_isspace() is static
*/
@@ -673,6 +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;
@@ -680,7 +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;
@@ -733,55 +757,16 @@ vector_dims(PG_FUNCTION_ARGS)
PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_norm);
Datum
vector_norm(PG_FUNCTION_ARGS)
{
Vector *a = PG_GETARG_VECTOR_P(0);
float *ax = a->x;
float norm = 0.0;
/* Auto-vectorized */
for (int i = 0; i < a->dim; i++)
norm += ax[i] * ax[i];
PG_RETURN_FLOAT8(sqrt((double) norm));
}
/*
* Normalize a vector with the L2 norm
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(normalize_l2);
Datum
normalize_l2(PG_FUNCTION_ARGS)
{
Vector *a = PG_GETARG_VECTOR_P(0);
float *ax = a->x;
double norm = 0.0;
Vector *result;
float *rx;
result = InitVector(a->dim);
rx = result->x;
/* Auto-vectorized */
for (int i = 0; i < a->dim; i++)
norm += (double) ax[i] * (double) ax[i];
norm = sqrt(norm);
if (norm > 0)
{
/* Auto-vectorized */
for (int i = 0, imax = a->dim; i < imax; i++)
rx[i] = ax[i] / norm;
/* Check for overflow */
for (int i = 0, imax = a->dim; i < imax; i++)
{
if (isinf(rx[i]))
float_overflow_error();
}
}
PG_RETURN_POINTER(result);
PG_RETURN_FLOAT8(sqrt(norm));
}
/*

View File

@@ -1,12 +1,6 @@
#ifndef VECTOR_H
#define VECTOR_H
#include "postgres.h"
#if PG_VERSION_NUM >= 160000
#include "varatt.h"
#endif
#define VECTOR_MAX_DIM 16000
#define VECTOR_SIZE(_dim) (offsetof(Vector, x) + sizeof(float)*(_dim))
@@ -14,9 +8,6 @@
#define PG_GETARG_VECTOR_P(x) DatumGetVector(PG_GETARG_DATUM(x))
#define PG_RETURN_VECTOR_P(x) PG_RETURN_POINTER(x)
/* Exported functions */
PGDLLEXPORT void _PG_init(void);
typedef struct Vector
{
int32 vl_len_; /* varlena header (do not touch directly!) */
@@ -25,24 +16,8 @@ typedef struct Vector
float x[FLEXIBLE_ARRAY_MEMBER];
} Vector;
Vector *InitVector(int dim);
void PrintVector(char *msg, Vector * vector);
int vector_cmp_internal(Vector * a, Vector * b);
/*
* Allocate and initialize a new vector
*/
static inline Vector *
InitVector(int dim)
{
Vector *result;
int size;
size = VECTOR_SIZE(dim);
result = (Vector *) palloc0(size);
SET_VARSIZE(result, size);
result->dim = dim;
return result;
}
#endif

View File

@@ -48,34 +48,10 @@ SELECT vector_norm('[0,1]');
1
(1 row)
SELECT normalize_l2('[3,4]');
normalize_l2
--------------
[0.6,0.8]
(1 row)
SELECT normalize_l2('[3,0]');
normalize_l2
--------------
[1,0]
(1 row)
SELECT normalize_l2('[0,0.1]');
normalize_l2
--------------
[0,1]
(1 row)
SELECT normalize_l2('[0,0]');
normalize_l2
--------------
[0,0]
(1 row)
SELECT normalize_l2('[3e38]');
normalize_l2
--------------
[1]
SELECT vector_norm('[3e37,4e37]')::real;
vector_norm
-------------
5e+37
(1 row)
SELECT l2_distance('[0,0]', '[3,4]');
@@ -130,6 +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

@@ -27,4 +27,10 @@ SELECT COUNT(*) FROM t;
5
(1 row)
TRUNCATE t;
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
val
-----
(0 rows)
DROP TABLE t;

View File

@@ -1,25 +1,26 @@
SET enable_seqscan = off;
CREATE TABLE t (val vector(3));
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 3);
ERROR: value 3 out of bounds for option "m"
DETAIL: Valid values are between "4" and "100".
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 1);
ERROR: value 1 out of bounds for option "m"
DETAIL: Valid values are between "2" and "100".
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 101);
ERROR: value 101 out of bounds for option "m"
DETAIL: Valid values are between "4" and "100".
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (ef_construction = 9);
ERROR: value 9 out of bounds for option "ef_construction"
DETAIL: Valid values are between "10" and "1000".
DETAIL: Valid values are between "2" and "100".
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (ef_construction = 3);
ERROR: value 3 out of bounds for option "ef_construction"
DETAIL: Valid values are between "4" and "1000".
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (ef_construction = 1001);
ERROR: value 1001 out of bounds for option "ef_construction"
DETAIL: Valid values are between "10" and "1000".
DETAIL: Valid values are between "4" and "1000".
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 16, ef_construction = 31);
ERROR: ef_construction must be greater than or equal to 2 * m
SHOW hnsw.ef_search;
hnsw.ef_search
----------------
40
(1 row)
SET hnsw.ef_search = 9;
ERROR: 9 is outside the valid range for parameter "hnsw.ef_search" (10 .. 1000)
SET hnsw.ef_search = 0;
ERROR: 0 is outside the valid range for parameter "hnsw.ef_search" (1 .. 1000)
SET hnsw.ef_search = 1001;
ERROR: 1001 is outside the valid range for parameter "hnsw.ef_search" (10 .. 1000)
ERROR: 1001 is outside the valid range for parameter "hnsw.ef_search" (1 .. 1000)
DROP TABLE t;

View File

@@ -1,7 +1,7 @@
SET enable_seqscan = off;
CREATE TABLE t (val vector(3));
INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
CREATE INDEX ON t USING ivfflat (val) WITH (lists = 1);
CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 1);
INSERT INTO t (val) VALUES ('[1,2,4]');
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
val
@@ -27,4 +27,13 @@ SELECT COUNT(*) FROM t;
5
(1 row)
TRUNCATE t;
NOTICE: ivfflat index created with little data
DETAIL: This will cause low recall.
HINT: Drop the index until the table has more data.
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
val
-----
(0 rows)
DROP TABLE t;

View File

@@ -1,9 +1,8 @@
SET enable_seqscan = off;
CREATE TABLE t (val vector(3));
CREATE INDEX ON t USING ivfflat (val) WITH (lists = 0);
CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 0);
ERROR: value 0 out of bounds for option "lists"
DETAIL: Valid values are between "1" and "32768".
CREATE INDEX ON t USING ivfflat (val) WITH (lists = 32769);
CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 32769);
ERROR: value 32769 out of bounds for option "lists"
DETAIL: Valid values are between "1" and "32768".
SHOW ivfflat.probes;

View File

@@ -1,7 +1,7 @@
SET enable_seqscan = off;
CREATE UNLOGGED TABLE t (val vector(3));
INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
CREATE INDEX ON t USING ivfflat (val) WITH (lists = 1);
CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 1);
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
val
---------

View File

@@ -11,12 +11,7 @@ SELECT vector_dims('[1,2,3]');
SELECT round(vector_norm('[1,1]')::numeric, 5);
SELECT vector_norm('[3,4]');
SELECT vector_norm('[0,1]');
SELECT normalize_l2('[3,4]');
SELECT normalize_l2('[3,0]');
SELECT normalize_l2('[0,0.1]');
SELECT normalize_l2('[0,0]');
SELECT normalize_l2('[3e38]');
SELECT vector_norm('[3e37,4e37]')::real;
SELECT l2_distance('[0,0]', '[3,4]');
SELECT l2_distance('[0,0]', '[0,1]');
@@ -30,6 +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

@@ -10,4 +10,7 @@ SELECT * FROM t ORDER BY val <-> '[3,3,3]';
SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
SELECT COUNT(*) FROM t;
TRUNCATE t;
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
DROP TABLE t;

View File

@@ -1,14 +1,13 @@
SET enable_seqscan = off;
CREATE TABLE t (val vector(3));
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 3);
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 1);
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 101);
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (ef_construction = 9);
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (ef_construction = 3);
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (ef_construction = 1001);
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 16, ef_construction = 31);
SHOW hnsw.ef_search;
SET hnsw.ef_search = 9;
SET hnsw.ef_search = 0;
SET hnsw.ef_search = 1001;
DROP TABLE t;

View File

@@ -2,7 +2,7 @@ SET enable_seqscan = off;
CREATE TABLE t (val vector(3));
INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
CREATE INDEX ON t USING ivfflat (val) WITH (lists = 1);
CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 1);
INSERT INTO t (val) VALUES ('[1,2,4]');
@@ -10,4 +10,7 @@ SELECT * FROM t ORDER BY val <-> '[3,3,3]';
SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
SELECT COUNT(*) FROM t;
TRUNCATE t;
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
DROP TABLE t;

View File

@@ -1,8 +1,6 @@
SET enable_seqscan = off;
CREATE TABLE t (val vector(3));
CREATE INDEX ON t USING ivfflat (val) WITH (lists = 0);
CREATE INDEX ON t USING ivfflat (val) WITH (lists = 32769);
CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 0);
CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 32769);
SHOW ivfflat.probes;

View File

@@ -2,7 +2,7 @@ SET enable_seqscan = off;
CREATE UNLOGGED TABLE t (val vector(3));
INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
CREATE INDEX ON t USING ivfflat (val) WITH (lists = 1);
CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 1);
SELECT * FROM t ORDER BY val <-> '[3,3,3]';

View File

@@ -19,14 +19,13 @@ 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";
my @r = ();
for (1 .. $dim) {
for (1 .. $dim)
{
push(@r, rand());
}
my $sql = join(",", @r);
@@ -52,11 +51,13 @@ my $array_sql = join(",", ('random()') x $dim);
# Initialize primary node
$node_primary = get_new_node('primary');
$node_primary->init(allows_streaming => 1);
if ($dim > 32) {
if ($dim > 32)
{
# TODO use wal_keep_segments for Postgres < 13
$node_primary->append_conf('postgresql.conf', qq(wal_keep_size = 1GB));
}
if ($dim > 1500) {
if ($dim > 1500)
{
$node_primary->append_conf('postgresql.conf', qq(maintenance_work_mem = 128MB));
}
$node_primary->start;
@@ -67,8 +68,7 @@ $node_primary->backup($backup_name);
# Create streaming replica linking to primary
$node_replica = get_new_node('replica');
$node_replica->init_from_backup($node_primary, $backup_name,
has_streaming => 1);
$node_replica->init_from_backup($node_primary, $backup_name, has_streaming => 1);
$node_replica->start;
# Create ivfflat index on primary
@@ -77,7 +77,7 @@ $node_primary->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector($dim));
$node_primary->safe_psql("postgres",
"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
);
$node_primary->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v);");
$node_primary->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops);");
# Test that queries give same result
test_index_replay('initial');

View File

@@ -7,7 +7,8 @@ use Test::More;
my $dim = 3;
my @r = ();
for (1 .. $dim) {
for (1 .. $dim)
{
my $v = int(rand(1000)) + 1;
push(@r, "i % $v");
}
@@ -24,7 +25,7 @@ $node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector($dim));");
$node->safe_psql("postgres",
"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
);
$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v);");
$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops);");
# Get size
my $size = $node->safe_psql("postgres", "SELECT pg_total_relation_size('tst_v_idx');");

View File

@@ -22,7 +22,8 @@ sub test_recall
));
like($explain, qr/Index Scan using idx on tst/);
for my $i (0 .. $#queries) {
for my $i (0 .. $#queries)
{
my $actual = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SET ivfflat.probes = $probes;
@@ -33,8 +34,10 @@ sub test_recall
my @expected_ids = split("\n", $expected[$i]);
foreach (@expected_ids) {
if (exists($actual_set{$_})) {
foreach (@expected_ids)
{
if (exists($actual_set{$_}))
{
$correct++;
}
$total++;
@@ -57,7 +60,8 @@ $node->safe_psql("postgres",
);
# Generate queries
for (1..20) {
for (1 .. 20)
{
my $r1 = rand();
my $r2 = rand();
my $r3 = rand();
@@ -66,26 +70,21 @@ for (1..20) {
# Check each index type
my @operators = ("<->", "<#>", "<=>");
my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_ops");
foreach (@operators) {
my $operator = $_;
for my $i (0 .. $#operators)
{
my $operator = $operators[$i];
my $opclass = $opclasses[$i];
# Get exact results
@expected = ();
foreach (@queries) {
foreach (@queries)
{
my $res = $node->safe_psql("postgres", "SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;");
push(@expected, $res);
}
my $opclass;
if ($operator eq "<->") {
$opclass = "vector_l2_ops";
} elsif ($operator eq "<#>") {
$opclass = "vector_ip_ops";
} else {
$opclass = "vector_cosine_ops";
}
# Build index serially
$node->safe_psql("postgres", qq(
SET max_parallel_maintenance_workers = 0;
@@ -93,13 +92,14 @@ foreach (@operators) {
));
# Test approximate results
if ($operator ne "<#>") {
# TODO fix test
test_recall(1, 0.75, $operator);
if ($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, 0.995, $operator);
test_recall(100, 0.9925, $operator);
$node->safe_psql("postgres", "DROP INDEX idx;");
@@ -113,13 +113,14 @@ foreach (@operators) {
like($stderr, qr/using \d+ parallel workers/);
# Test approximate results
if ($operator ne "<#>") {
# TODO fix test
test_recall(1, 0.75, $operator);
if ($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, 0.995, $operator);
test_recall(100, 0.9925, $operator);
$node->safe_psql("postgres", "DROP INDEX idx;");
}

View File

@@ -20,7 +20,7 @@ sub test_centers
{
my ($lists, $min) = @_;
my ($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING ivfflat (v) WITH (lists = $lists);");
my ($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops) WITH (lists = $lists);");
is($ret, 0, $stderr);
}

View File

@@ -18,24 +18,21 @@ $node->safe_psql("postgres",
# Check each index type
my @operators = ("<->", "<#>", "<=>");
foreach (@operators) {
my $operator = $_;
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
my $opclass;
if ($operator eq "<->") {
$opclass = "vector_l2_ops";
} elsif ($operator eq "<#>") {
$opclass = "vector_ip_ops";
} else {
$opclass = "vector_cosine_ops";
}
$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v $opclass);");
# Test 100% recall
for (1..20) {
my $i = int(rand() * 100000);
my $query = $node->safe_psql("postgres", "SELECT v FROM tst WHERE i = $i;");
for (1 .. 20)
{
my $id = int(rand() * 100000);
my $query = $node->safe_psql("postgres", "SELECT v FROM tst WHERE i = $id;");
my $res = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SELECT v FROM tst ORDER BY v <-> '$query' LIMIT 1;

View File

@@ -16,8 +16,8 @@ $node->safe_psql("postgres",
"INSERT INTO tst SELECT ARRAY[random(), random(), random()] FROM generate_series(1, 100000) i;"
);
$node->safe_psql("postgres", "CREATE INDEX lists50 ON tst USING ivfflat (v) WITH (lists = 50);");
$node->safe_psql("postgres", "CREATE INDEX lists100 ON tst USING ivfflat (v) WITH (lists = 100);");
$node->safe_psql("postgres", "CREATE INDEX lists50 ON tst USING ivfflat (v vector_l2_ops) WITH (lists = 50);");
$node->safe_psql("postgres", "CREATE INDEX lists100 ON tst USING ivfflat (v vector_l2_ops) WITH (lists = 100);");
# Test prefers more lists
my $res = $node->safe_psql("postgres", "EXPLAIN SELECT v FROM tst ORDER BY v <-> '[0.5,0.5,0.5]' LIMIT 10;");
@@ -26,7 +26,7 @@ unlike($res, qr/lists50/);
# Test errors with too much memory
my ($ret, $stdout, $stderr) = $node->psql("postgres",
"CREATE INDEX lists10000 ON tst USING ivfflat (v) WITH (lists = 10000);"
"CREATE INDEX lists10000 ON tst USING ivfflat (v vector_l2_ops) WITH (lists = 10000);"
);
like($stderr, qr/memory required is/);

View File

@@ -19,7 +19,7 @@ $node->safe_psql("postgres", "CREATE TABLE tst (v vector($dim));");
$node->safe_psql("postgres",
"INSERT INTO tst SELECT ARRAY[$array_sql] FROM generate_series(1, 10000) i;"
);
$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v);");
$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops);");
$node->pgbench(
"--no-vacuum --client=5 --transactions=100",
@@ -28,7 +28,7 @@ $node->pgbench(
[qr{^$}],
"concurrent INSERTs",
{
"007_inserts" => "INSERT INTO tst SELECT ARRAY[$array_sql] FROM generate_series(1, 10) i;"
"007_ivfflat_inserts" => "INSERT INTO tst SELECT ARRAY[$array_sql] FROM generate_series(1, 10) i;"
}
);

View File

@@ -30,7 +30,8 @@ sub test_aggregate
# Test matches real for avg
# Cannot test sum since sum(real) varies between calls
if ($agg eq 'avg') {
if ($agg eq 'avg')
{
my $r1 = $node->safe_psql("postgres", "SELECT $agg(r1)::float4 FROM tst;");
my $r2 = $node->safe_psql("postgres", "SELECT $agg(r2)::float4 FROM tst;");
my $r3 = $node->safe_psql("postgres", "SELECT $agg(r3)::float4 FROM tst;");

View File

@@ -19,14 +19,13 @@ 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";
my @r = ();
for (1 .. $dim) {
for (1 .. $dim)
{
push(@r, rand());
}
my $sql = join(",", @r);
@@ -52,11 +51,13 @@ my $array_sql = join(",", ('random()') x $dim);
# Initialize primary node
$node_primary = get_new_node('primary');
$node_primary->init(allows_streaming => 1);
if ($dim > 32) {
if ($dim > 32)
{
# TODO use wal_keep_segments for Postgres < 13
$node_primary->append_conf('postgresql.conf', qq(wal_keep_size = 1GB));
}
if ($dim > 1500) {
if ($dim > 1500)
{
$node_primary->append_conf('postgresql.conf', qq(maintenance_work_mem = 128MB));
}
$node_primary->start;
@@ -67,8 +68,7 @@ $node_primary->backup($backup_name);
# Create streaming replica linking to primary
$node_replica = get_new_node('replica');
$node_replica->init_from_backup($node_primary, $backup_name,
has_streaming => 1);
$node_replica->init_from_backup($node_primary, $backup_name, has_streaming => 1);
$node_replica->start;
# Create hnsw index on primary

View File

@@ -7,7 +7,8 @@ use Test::More;
my $dim = 3;
my @r = ();
for (1 .. $dim) {
for (1 .. $dim)
{
my $v = int(rand(1000)) + 1;
push(@r, "i % $v");
}
@@ -22,7 +23,7 @@ $node->start;
$node->safe_psql("postgres", "CREATE EXTENSION vector;");
$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector($dim));");
$node->safe_psql("postgres",
"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series(1, 10000) i;"
"INSERT INTO tst SELECT i, ARRAY[$array_sql] FROM generate_series(1, 10000) i;"
);
$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v vector_l2_ops);");
@@ -33,11 +34,21 @@ my $size = $node->safe_psql("postgres", "SELECT pg_total_relation_size('tst_v_id
$node->safe_psql("postgres", "DELETE FROM tst;");
$node->safe_psql("postgres", "VACUUM tst;");
$node->safe_psql("postgres",
"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series(1, 10000) i;"
"INSERT INTO tst SELECT i, ARRAY[$array_sql] FROM generate_series(1, 10000) i;"
);
# Check size
# May increase some due to different levels
my $new_size = $node->safe_psql("postgres", "SELECT pg_total_relation_size('tst_v_idx');");
cmp_ok($new_size, "<=", $size * 1.01, "size does not increase too much");
cmp_ok($new_size, "<=", $size * 1.02, "size does not increase too much");
# Delete all but one
$node->safe_psql("postgres", "DELETE FROM tst WHERE i != 123;");
$node->safe_psql("postgres", "VACUUM tst;");
my $res = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SELECT i FROM tst ORDER BY v <-> '[0,0,0]' LIMIT 10;
));
is($res, 123);
done_testing();

View File

@@ -21,7 +21,8 @@ sub test_recall
));
like($explain, qr/Index Scan/);
for my $i (0 .. $#queries) {
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;
@@ -31,8 +32,10 @@ sub test_recall
my @expected_ids = split("\n", $expected[$i]);
foreach (@expected_ids) {
if (exists($actual_set{$_})) {
foreach (@expected_ids)
{
if (exists($actual_set{$_}))
{
$correct++;
}
$total++;
@@ -55,7 +58,8 @@ $node->safe_psql("postgres",
);
# Generate queries
for (1..20) {
for (1 .. 20)
{
my $r1 = rand();
my $r2 = rand();
my $r3 = rand();
@@ -64,33 +68,26 @@ for (1..20) {
# Check each index type
my @operators = ("<->", "<#>", "<=>");
my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_ops");
foreach (@operators) {
my $operator = $_;
for my $i (0 .. $#operators)
{
my $operator = $operators[$i];
my $opclass = $opclasses[$i];
# Get exact results
@expected = ();
foreach (@queries) {
foreach (@queries)
{
my $res = $node->safe_psql("postgres", "SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;");
push(@expected, $res);
}
# Add index
my $opclass;
if ($operator eq "<->") {
$opclass = "vector_l2_ops";
} elsif ($operator eq "<#>") {
$opclass = "vector_ip_ops";
} else {
$opclass = "vector_cosine_ops";
}
$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v $opclass);");
if ($operator eq "<#>") {
test_recall(0.80, $operator);
} else {
test_recall(0.99, $operator);
}
my $min = $operator eq "<#>" ? 0.80 : 0.99;
test_recall($min, $operator);
}
done_testing();

View File

@@ -21,7 +21,8 @@ sub test_recall
));
like($explain, qr/Index Scan/);
for my $i (0 .. $#queries) {
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;
@@ -31,8 +32,10 @@ sub test_recall
my @expected_ids = split("\n", $expected[$i]);
foreach (@expected_ids) {
if (exists($actual_set{$_})) {
foreach (@expected_ids)
{
if (exists($actual_set{$_}))
{
$correct++;
}
$total++;
@@ -49,10 +52,11 @@ $node->start;
# Create table
$node->safe_psql("postgres", "CREATE EXTENSION vector;");
$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector(3));");
$node->safe_psql("postgres", "CREATE TABLE tst (i serial, v vector(3));");
# Generate queries
for (1..20) {
for (1 .. 20)
{
my $r1 = rand();
my $r2 = rand();
my $r3 = rand();
@@ -61,28 +65,32 @@ for (1..20) {
# Check each index type
my @operators = ("<->", "<#>", "<=>");
my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_ops");
foreach (@operators) {
my $operator = $_;
for my $i (0 .. $#operators)
{
my $operator = $operators[$i];
my $opclass = $opclasses[$i];
# Add index
my $opclass;
if ($operator eq "<->") {
$opclass = "vector_l2_ops";
} elsif ($operator eq "<#>") {
$opclass = "vector_ip_ops";
} else {
$opclass = "vector_cosine_ops";
}
$node->safe_psql("postgres", "CREATE INDEX idx ON tst USING hnsw (v $opclass);");
$node->safe_psql("postgres",
"INSERT INTO tst SELECT i, ARRAY[random(), random(), random()] FROM generate_series(1, 10000) i;"
# Use concurrent inserts
$node->pgbench(
"--no-vacuum --client=10 --transactions=1000",
0,
[qr{actually processed}],
[qr{^$}],
"concurrent INSERTs",
{
"013_hnsw_insert_recall_$opclass" => "INSERT INTO tst (v) VALUES (ARRAY[random(), random(), random()]);"
}
);
# Get exact results
@expected = ();
foreach (@queries) {
foreach (@queries)
{
my $res = $node->safe_psql("postgres", qq(
SET enable_indexscan = off;
SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;
@@ -90,11 +98,8 @@ foreach (@operators) {
push(@expected, $res);
}
if ($operator eq "<#>") {
test_recall(0.80, $operator);
} else {
test_recall(0.99, $operator);
}
my $min = $operator eq "<#>" ? 0.80 : 0.99;
test_recall($min, $operator);
$node->safe_psql("postgres", "DROP INDEX idx;");
$node->safe_psql("postgres", "TRUNCATE tst;");

View File

@@ -17,22 +17,8 @@ $node->start;
# Create table and index
$node->safe_psql("postgres", "CREATE EXTENSION vector;");
$node->safe_psql("postgres", "CREATE TABLE tst (v vector($dim));");
$node->safe_psql("postgres",
"INSERT INTO tst SELECT ARRAY[$array_sql] FROM generate_series(1, 100) i;"
);
$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v vector_l2_ops);");
$node->pgbench(
"--no-vacuum --client=5 --transactions=100",
0,
[qr{actually processed}],
[qr{^$}],
"concurrent INSERTs",
{
"007_inserts" => "INSERT INTO tst SELECT ARRAY[$array_sql] FROM generate_series(1, 10) i;"
}
);
sub idx_scan
{
# Stats do not update instantaneously
@@ -41,18 +27,48 @@ sub idx_scan
$node->safe_psql("postgres", "SELECT idx_scan FROM pg_stat_user_indexes WHERE indexrelid = 'tst_v_idx'::regclass;");
}
my $expected = 100 + 5 * 100 * 10;
for my $i (1 .. 20)
{
$node->pgbench(
"--no-vacuum --client=10 --transactions=1",
0,
[qr{actually processed}],
[qr{^$}],
"concurrent INSERTs",
{
"014_hnsw_inserts_$i" => "INSERT INTO tst VALUES (ARRAY[$array_sql]);"
}
);
my $count = $node->safe_psql("postgres", "SELECT COUNT(*) FROM tst;");
is($count, $expected);
is(idx_scan(), 0);
my $count = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SELECT COUNT(*) FROM (SELECT v FROM tst ORDER BY v <-> (SELECT v FROM tst LIMIT 1)) t;
));
is($count, 10);
$count = $node->safe_psql("postgres", qq(
$node->safe_psql("postgres", "TRUNCATE tst;");
}
$node->pgbench(
"--no-vacuum --client=20 --transactions=5",
0,
[qr{actually processed}],
[qr{^$}],
"concurrent INSERTs",
{
"014_hnsw_inserts" => "INSERT INTO tst SELECT ARRAY[$array_sql] FROM generate_series(1, 10) i;"
}
);
my $count = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SET hnsw.ef_search = 400;
SET hnsw.ef_search = 1000;
SELECT COUNT(*) FROM (SELECT v FROM tst ORDER BY v <-> (SELECT v FROM tst LIMIT 1)) t;
));
is($count, 400);
is(idx_scan(), 1);
# Elements may lose all incoming connections with the HNSW algorithm
# Vacuuming can fix this if one of the elements neighbors is deleted
cmp_ok($count, ">=", 997);
is(idx_scan(), 21);
done_testing();

View File

@@ -0,0 +1,58 @@
use strict;
use warnings;
use PostgresNode;
use TestLib;
use Test::More;
# Initialize node
my $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 (v vector(3));");
sub insert_vectors
{
for my $i (1 .. 20)
{
$node->safe_psql("postgres", "INSERT INTO tst VALUES ('[1,1,1]');");
}
}
sub test_duplicates
{
my $res = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SET hnsw.ef_search = 1;
SELECT COUNT(*) FROM (SELECT * FROM tst ORDER BY v <-> '[1,1,1]') t;
));
is($res, 10);
}
# Test duplicates with build
insert_vectors();
$node->safe_psql("postgres", "CREATE INDEX idx ON tst USING hnsw (v vector_l2_ops);");
test_duplicates();
# Reset
$node->safe_psql("postgres", "TRUNCATE tst;");
# Test duplicates with inserts
insert_vectors();
test_duplicates();
# Test fallback path for inserts
$node->pgbench(
"--no-vacuum --client=5 --transactions=100",
0,
[qr{actually processed}],
[qr{^$}],
"concurrent INSERTs",
{
"015_hnsw_duplicates" => "INSERT INTO tst VALUES ('[1,1,1]');"
}
);
done_testing();

View File

@@ -0,0 +1,97 @@
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, $ef_search, $test_name) = @_;
my $correct = 0;
my $total = 0;
my $explain = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SET hnsw.ef_search = $ef_search;
EXPLAIN ANALYZE SELECT i FROM tst ORDER BY v <-> '$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;
SET hnsw.ef_search = $ef_search;
SELECT i FROM tst ORDER BY v <-> '$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, $test_name);
}
# 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 vector(3));");
$node->safe_psql("postgres", "ALTER TABLE tst SET (autovacuum_enabled = false);");
$node->safe_psql("postgres",
"INSERT INTO tst SELECT i, ARRAY[random(), random(), random()] FROM generate_series(1, 10000) i;"
);
# Add index
$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v vector_l2_ops) WITH (m = 4, ef_construction = 8);");
# Delete data
$node->safe_psql("postgres", "DELETE FROM tst WHERE i > 2500;");
# Generate queries
for (1 .. 20)
{
my $r1 = rand();
my $r2 = rand();
my $r3 = rand();
push(@queries, "[$r1,$r2,$r3]");
}
# Get exact results
@expected = ();
foreach (@queries)
{
my $res = $node->safe_psql("postgres", qq(
SET enable_indexscan = off;
SELECT i FROM tst ORDER BY v <-> '$_' LIMIT $limit;
));
push(@expected, $res);
}
test_recall(0.20, $limit, "before vacuum");
test_recall(0.95, 100, "before vacuum");
# TODO Test concurrent inserts with vacuum
$node->safe_psql("postgres", "VACUUM tst;");
test_recall(0.95, $limit, "after vacuum");
done_testing();

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();

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 access method'
default_version = '0.4.4'
comment = 'vector data type and ivfflat and hnsw access methods'
default_version = '0.5.1'
module_pathname = '$libdir/vector'
relocatable = true