perf: key-only DELETE row decode (B1) - #370
Merged
Merged
Conversation
added 4 commits
September 3, 2026 20:18
… flush rewrite) Stap-0 profiling showed batch DELETE (ExecuteBatchSQL wraps every batch in a storage transaction) was NOT using the tombstone path: the rollback-safe deferral only counted _pendingLogicalDeletes, so db.Flush() still ran the #366 full-file CompactPendingDeletes rewrite (~0.5-0.7s; DELETE stuck at ~12-16K ops/s). - IStorage.BufferTombstoneForCommit (default no-op) + Storage implementation: buffered offsets per file, applied as in-place negative-prefix markers by ApplyBufferedTombstones() inside FlushBufferedAppendsAndOverwrites() (the commit path) AFTER buffered appends are on disk; rollback discards the buffer via ClearBufferedAppends(). - DeleteRecordsCore + the contiguous FW bulk path now buffer the deleted offsets when IsInTransaction instead of incrementing _pendingLogicalDeletes, so the flush-time full-file rewrite is off the batch-DELETE path entirely. - Regression: batch DELETE survives reopen even WITHOUT an explicit Flush after ExecuteBatchSQL (commit already tombstoned the rows). Measured (same machine, Release, comparative harness DELETE 10K of ~100K rows): SQL 0.82s/12K -> 0.24s/41K ops/s, Direct 0.63s/16K -> 0.17s/58K ops/s; --pk legacy 0.16s/62K, fixed-width 0.13s/78K ops/s. Full suite EXIT=0.
- IStorage.TombstoneRecords (default loops TombstoneRecord); Storage batches the in-place negative-prefix markers over one cached read handle and evicts each affected page-cache page once instead of once per row. - ApplyBufferedTombstones (commit path) and Table.TombstoneDeletedPositions (direct non-transactional deletes) route through the batch API; offsets that are not physical records (EOF/already-marked) are skipped safely.
…parse (B3) - TryParseDeleteForBatch now also returns the canonical WHERE column + raw literal; ExecuteBatchSQL groups (where, column, literal) instead of plain WHERE strings. - Table.DeleteMultipleKeys mirrors DeleteMultiple (contiguous-FW gate -> PK fast path -> hash fast path -> generic fallback) but resolves keys directly from the pre-parsed column/literal; the 'col = literal' string and the second TryParseSimpleWhereClause pass are built only when a generic fallback actually runs. - Non-canonical statements keep the string path (mixed tables rebuild on the rare path).
…index columns The structured batch DELETE path (DeleteMultipleKeys) now decodes a minimal row: BuildDeleteKeyColumns computes the needed columns (PK + every loaded hash-index column), and DeserializeDeleteKeyRow walks the legacy variable-length record skipping the unneeded columns' payload parsing entirely, building a 1-3 entry dictionary instead of a full row. DeleteRecordsCore performs the identical PK/hash lookups on that subset. Fixed-width layouts and corrupt rows fall back to the full DeserializeRowFromSpan. Measured (Release, comparative harness DELETE 10K of ~100K rows, median of 3): SQL ~39-42K and Direct ~57K ops/s — neutral within run noise on this small-row workload; the win shows on wider rows / long unindexed TEXT payloads where payload parsing and boxing are skipped. Full suite + 4 CI-filter suites EXIT=0.
|
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.
Stack
perf/b3c4-dml(PR perf: batch tombstone writes (C4) + structured batch DELETE (B3) #369) ←perf/commit-time-tombstones(perf: apply DELETE tombstones at COMMIT for transactional deletes (no flush rewrite) #368) ←perf/tombstone-delete(perf: durable Columnar DELETE in-place via tombstone markers (no flush rewrite) #367) ← master.Inhoud
B1 (
d37fad51) — key-only DELETE row-decode: de structured batch-DELETE-path (DeleteMultipleKeys) bouwt viaBuildDeleteKeyColumnsde benodigde kolomindexen (PK + alle geladen hash-indexkolommen) en decodeert metDeserializeDeleteKeyRowalleen díe kolommen uit het legacy variabele-lengte record (payload-parsing van overige kolommen wordt overgeslagen, minimal 1-3-entry dictionary i.p.v. full row).DeleteRecordsCoredoet daarna dezelfde PK/hash-lookups op die subset. Fixed-width layouts en corrupte rijen vallen terug op de volledigeDeserializeRowFromSpan.Meting (Release, comparative DELETE 10K op ~100K rijen, median van 3)
Neutraal binnen run-ruis op deze small-row workload — de decode was niet de bottleneck (kleine rijen, gepoolde full-row dict). De winst zit in workloads met brede rijen/lange on-geïndexeerde TEXT-payloads (minder UTF8-decode, boxing en dict-writes). Veilig: full suite + 4 CI-filter-suites EXIT=0.
Conclusie van de meetreeks
De grote sprong zat in #368 (commit-time tombstones: SQL DELETE 0,82s → 0,24s). C4/B3/B1 zijn robuustheids-/allocatiewinsten die op deze benchmark binnen de ruis vallen. De volgende structurele hefboom is A1 (contiguous variabele-lengte single-pass UPDATE/DELETE — de docs-tabel heeft fysiek oplopende posities bij oplopende keys, maar alleen fixed-width gebruikt nu de één-range-read) en daarna bulk-B-tree/index-onderhoud.