perf(fts): bound classic conjunction scoring in dense windows - #9033
Merged
Conversation
This was referenced Sep 7, 2026
![lance-gatekeeper[bot]](https://avatars.githubusercontent.com/in/4433141?s=60&v=4){kind=small}
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
lance-gatekeeper
Bot
added
K-approved
BubbleCal
added a commit
that referenced
this pull request
Sep 8, 2026
Before a competitive floor exists, AND queries still decoded frequencies and prepared bounds while looking for intersections. This PR keeps that work lazy, particularly when the final intersection has fewer than `k` matches. - Split DocID and frequency decoding for full posting blocks and tails, retaining the existing encoded format and combined decoding entry points. - Advance and intersect DocIDs before decoding frequencies; initialize score bounds when a positive competitive floor makes them useful. - Avoid loading document lengths for partitions whose dictionaries cannot satisfy all required tokens. - Use SIMD catch-up for the existing classic 4/5-clause path, with unsigned comparisons covering the full `u32` DocID range and a binary-search fallback without AVX2. Tests cover empty/underfilled intersections, delayed frequency and bound activation, block/tail transitions, required-token misses, and scalar/SIMD boundary behavior. Code and its regression tests stay together in this PR. No isolated `main` A/B measurement is claimed for this layer. The later bulk-intersection measurements belong to [#9030](#9030). ## Validation - Local inverted-index suite: 530 passed. - Required all-workspace Clippy: passed. - CI: pending. Stack, 1/4: **this PR** → [#9032](#9032) → [#9033](#9033) → [#9030](#9030). Base: `main`. Merge in stack order; retarget/restack the remaining PRs after their parent merges.
BubbleCal
force-pushed
the
yang/fts-stack-certified-and
branch
from
September 8, 2026 06:25
0311000 to
f694a58
Compare
Certify dense posting ranges before progressive scoring, retain conservative suffix bounds and exact query-order scores, and reuse zero-floor windows and lazy norm lookup. Preserve sparse, phrase, and compound-cursor behavior. Share test fixtures without removing assertions or parameter coverage.
BubbleCal
added a commit
that referenced
this pull request
Sep 8, 2026
Fully prewarmed FTS queries still pay asynchronous document-frequency aggregation overhead, and segment searches clone an already prepared canonical scorer. This PR reuses the resident metadata and scorer. - Aggregate corpus statistics and posting lengths synchronously only when every required value is already loaded. The probe never starts I/O; incomplete residency uses the existing asynchronous path. - Reuse the canonical prepared `Arc<MemBM25Scorer>` across modern segments. Legacy and externally supplied scorers retain their existing fallback behavior. - Check statistic shapes and integer overflow while merging global counts. - Provide `LANCE_FTS_SYNC_DF=off` and `LANCE_FTS_REUSE_PREPARED_SCORER=off` switches. The accompanying tests cover cold/partial/full residency, multiple segments, missing terms, overflow, injected scorers, fuzzy preparation, and exact result parity. Shared fixtures remove repetition while retaining the relevant execution modes and assertions. This layer does not change the query API or persistent format. No isolated `main` A/B performance result is claimed. ## Validation - Local inverted-index suite: 542 passed. - Required all-workspace Clippy and formatting: passed. - PR metadata/gatekeeper checks: pending. Full hosted test workflows only run for `main`/`release/**` bases; this intermediate layer has been tested and linted locally. The full hosted matrix will run after its base is retargeted to `main` following the parent merge. Stack, 2/4: [#9031](#9031) → **this PR** → [#9033](#9033) → [#9030](#9030). Base: [#9031](#9031); this diff contains only resident-statistic and prepared-scorer changes with their tests. Merge in stack order; retarget/restack the remaining PRs after their parent merges.
BubbleCal
force-pushed
the
yang/fts-stack-certified-and
branch
from
September 8, 2026 16:05
f694a58 to
50c5d9b
Compare
lance-gatekeeper
Bot
removed
K-approved
Contributor
There was a problem hiding this comment.
The rebase onto main leaves the score-first implementation unchanged. Its strict eligibility, conservative suffix bounds, canonical score reconstruction, and focused parity/transition coverage continue to support correctness.
The non-blocking risk remains performance: no isolated latency/CPU comparison establishes the net benefit or possible short-tail overhead. Keep the scope narrow until representative classic 4/5-term workloads confirm the trade-off; if the incremental gain is small, progressive scoring after full DocID alignment is the simpler alternative.
lance-gatekeeper
Bot
added
K-approved
LuQQiu
approved these changes
Sep 9, 2026
BubbleCal
added a commit
that referenced
this pull request
Sep 10, 2026
Wide root Match AND queries still spend substantial time in per-document iterator advancement. This PR enables the existing bulk scoring route for four or more modern block-256, non-grouped clauses, and adds bounded pairwise SIMD/scalar DocID intersection before the competitive floor becomes positive. Posting offsets are recovered only for final survivors; positive-floor scoring retains the existing bulk implementation. Ordinary root Match AND queries directly use `Wand.search`. Compound `WandCursor` continues to use the classic path covered by [#9033](#9033). The separate Boolean MUST-child floor-propagation change in #9015 is not included. ## Measurements The comparison is **incremental**, from `85ec700bb4aacbc4e87f826f93fe7bc78fe53e41` to measured candidate `02c5f5a73fb29a47a0ff33293936e38c8f90c470`. The current stacked tip, `0181df1136be876a04b327f24a0f1a4709263b58`, has identical production code; subsequent changes only deduplicate tests and exclude benchmark files. The parent stack layer likewise retains the measured baseline's production code. These measurements establish the final bulk-routing/intersection increment, not the combined stack's speedup versus `main`. | Scenario / metric | Baseline | This PR | Benefit | | --- | ---: | ---: | ---: | | 4 terms, 0 matches / latency | 1.237869 ms | 0.335107 ms | 3.694x speedup | | 5 terms, 5 matches / latency | 1.552503 ms | 0.448852 ms | 3.459x speedup | | 8 terms, 0 matches / latency | 3.984348 ms | 0.841667 ms | 4.734x speedup | | 16 terms, 5 matches / latency | 8.309449 ms | 1.882632 ms | 4.414x speedup | | 4 terms, contiguous 32,768 matches / latency | 2.627815 ms | 1.309050 ms | 2.007x speedup | | 16 terms, contiguous 32,768 matches / latency | 13.304164 ms | 3.344204 ms | 3.978x speedup | | 16 terms, sparse correlated 32,768 matches / latency | 13.401886 ms | 3.383210 ms | 3.961x speedup | Lower latency is better; benefit is baseline/current. Same AWS `c7i.8xlarge`, Xeon Platinum 8488C, Ubuntu 24.04, Rust 1.97.0/clang 18, repository `release-with-debug` with thin LTO and configured Haswell/AVX2 target. Both versions use Auto mode, CPU 2, one query in flight, `LANCE_CPU_THREADS=1`, and `RAYON_NUM_THREADS=1`. The complete comparison contains 42 frozen cases: 65,536-document scaling/additional-distribution corpora and a 131,072-document control corpus. All use `k=10`, V3/block256, no positions/filter, and fully resident prewarm. High-hit scaling holds the intersection and each term's DF at 32,768 across term counts. Four independent processes per version/case run in ABBAABBA order, each with 2 s warmup, 5 s measurement, and 30 samples. Values are the median of four run medians. Independent scoring oracles and exact row/score-bit signatures passed. Eight of 15 short-query controls show observed median regressions of approximately 0.6–2.2%. Some baseline runs have wider between-process spread; these are descriptive repeated measurements, not significance claims. Results cover warm index-query CPU work, not a 2.86B-row production or end-to-end dataset workload. Combined performance versus `main` remains unverified. Benchmark drivers and raw evidence are archived separately; no benchmark files are included in this stack. ## Validation - The equivalent combined implementation passed 762 inverted-index tests on ARM and x86. - Six dataset integration cases passed: N4/N16 × 0/5/high matches, two fragments, V3/block256, independent BM25 row/score-bit oracle, verified MatchQuery plans, and a strict 10th/11th score gap. - Thirteen existing WAND/certificate/tie tests passed, including bounded completion and exact overflow replay. - Per-layer inverted-index suite: 762 passed; required all-workspace Clippy and formatting passed. - Dataset integration rerun: 6 root Match AND cases and 13 WAND/certificate/tie cases passed. CI: pending. Stack, 4/4: [#9031](#9031) → [#9032](#9032) → [#9033](#9033) → **this PR**. Base: [#9033](#9033). Merge in stack order; retarget/restack the remaining PRs after their parent merges.
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
2 participants
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.
Classic 4/5-term AND evaluation can decode and score every follower even when a partial exact score already proves the candidate cannot compete. Applying score-first evaluation universally is costly when posting overlap is low. This PR enables it only inside a metadata-certified dense range.
f32scores and inclusive tie behavior.The classic path remains used by compound
WandCursorand explicit classic searches after #9030 enables standalone wide-query bulk routing.Most of this diff is regression coverage: approximately 327 production-section additions and 1,334 test additions. Shared fixtures retain the distinct dense/sparse, malformed-metadata, short-tail, high-DocID, score-bit, tie, norm-fallback, seek/shallow, and floor-transition cases. No isolated
mainA/B result is claimed for this layer.Validation
main/release/**bases; this intermediate layer has been tested and linted locally. The full hosted matrix will run after its base is retargeted tomainfollowing the parent merge.Stack, 3/4: #9031 → #9032 → this PR → #9030. Base: #9032. Uses the lazy posting-cursor support introduced in #9031.
Merge in stack order; retarget/restack the remaining PRs after their parent merges.