perf(index): reduce incremental delete from O(index_size) to O(IVF_size) (rebase of #135)#138
Closed
raphaelsty wants to merge 2 commits into
Closed
perf(index): reduce incremental delete from O(index_size) to O(IVF_size) (rebase of #135)#138raphaelsty wants to merge 2 commits into
raphaelsty wants to merge 2 commits into
Conversation
Three optimizations cut incremental update time from 60s to ~3s on large indices (~90K docs, 12GB vector data, 5GB metadata): 1. IVF patch instead of full rebuild (delete.rs): Load existing IVF, filter deleted doc IDs and renumber survivors via binary search, write back. Avoids re-reading all chunk code files and BTreeMap construction. 2. In-place PK re-sequencing (filtering.rs): Replace the temp-table-copy approach (which copied all TEXT/BLOB data 3x) with targeted UPDATE statements that only modify the integer _subset_ column for affected ranges. 3. Skip merged file regeneration (index.rs + colgrep): Add MmapIndex::update_append that writes new chunks and merges the IVF without loading the full MmapIndex (which triggers merged mmap file generation). Merged files regenerate lazily on next search. Also skips FTS5 rebuild on the delete hot path when encoding follows immediately (changed files), deferring to the post-encoding rebuild.
raphaelsty
added a commit
that referenced
this pull request
Jun 17, 2026
…supersedes #135/#138) (#139) * perf(index): reduce incremental delete from O(index_size) to O(IVF_size) Three optimizations cut incremental update time from 60s to ~3s on large indices (~90K docs, 12GB vector data, 5GB metadata): 1. IVF patch instead of full rebuild (delete.rs): Load existing IVF, filter deleted doc IDs and renumber survivors via binary search, write back. Avoids re-reading all chunk code files and BTreeMap construction. 2. In-place PK re-sequencing (filtering.rs): Replace the temp-table-copy approach (which copied all TEXT/BLOB data 3x) with targeted UPDATE statements that only modify the integer _subset_ column for affected ranges. 3. Skip merged file regeneration (index.rs + colgrep): Add MmapIndex::update_append that writes new chunks and merges the IVF without loading the full MmapIndex (which triggers merged mmap file generation). Merged files regenerate lazily on next search. Also skips FTS5 rebuild on the delete hot path when encoding follows immediately (changed files), deferring to the post-encoding rebuild. * fix(index): harden #135 (re-sequencing robustness, FTS/merge clarity, tests) Review fixes on top of the #135 rebase: - filtering::delete range re-sequencing now clamps the deleted-id set to ids that were actually present and in range before computing shifts. A stray negative or out-of-range id previously corrupted every survivor's _subset_ id (and thus the metadata/FTS/IVF alignment); the old ROW_NUMBER path was robust to any input. - delete.rs: document that the in-place IVF patch relies on (and preserves) create's sorted+deduped per-centroid doc-id layout — the same layout deployed/next-plaid-api indexes use, so the patch is format-safe. - update_append doc: clarify it DEFERS merged-file regeneration to the next search (lazy), not eliminates it. - colgrep: correct the inaccurate 'final FTS rebuild after encoding' comment — there is none; keyword search stays correct because it re-verifies FTS hits against each doc's current content; stale survivor rows realign on the next delete-only rebuild. Tests: - filtering: test_delete_resequence_ignores_out_of_range_and_negative_ids (fails without the clamp). - delete: assert the in-place IVF patch keeps each centroid bucket deduped and ivf_lengths consistent. Validated: clippy clean; next-plaid 122, colgrep 557 tests pass. Co-authored-by: Vlad Lasky <12727610+vlasky@users.noreply.github.com> --------- Co-authored-by: Vlad Lasky <vlad.lasky@energyone.com> Co-authored-by: Vlad Lasky <12727610+vlasky@users.noreply.github.com>
Collaborator
Author
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Rebase of #135 (by @vlasky) onto current
main(post #137). No conflicts — the FTS5-in-delete code #135 builds on came from #132 (already in 1.5.5), so it merged cleanly.The change reduces incremental-delete cost from O(index_size) to ~O(IVF_size):
UPDATE ... SET _subset_ = _subset_ - shift) instead of copying the whole METADATA table through a temp table.MmapIndex::update_appendto skip merged-file regeneration during incremental updates (deferred to the next search, which regenerates lazily).delete_files_from_index_no_fts_rebuild).Status / review notes (see PR discussion)
-e) search returns correct results after incremental edits/deletes (deleted content not surfaced, survivors resolve correctly).Co-authored with @vlasky (#135).