Add a pluggable CostModel for the compressor#8753
Draft
connortsui20 wants to merge 3 commits into
Draft
Conversation
connortsui20
force-pushed
the
ct/compressor-cost-model
branch
from
314bf6e to
48c8634
Compare
connortsui20
changed the title
connortsui20
changed the title
CostModel for the compressor
connortsui20
force-pushed
the
ct/compressor-cost-model
branch
from
48c8634 to
1376bf5
Compare
connortsui20
added
the
action/benchmark
github-actions
Bot
removed
the
action/benchmark
Contributor
Polar Signals Profiling ResultsLatest Run
Powered by Polar Signals Cloud |
Contributor
Benchmarks: PolarSignals ProfilingVortex (geomean): 1.122x ❌ How to read Verdict and Engines
datafusion / vortex-file-compressed (1.122x ❌, 0↑ 7↓)
No file size changes detected. |
Contributor
Benchmarks: FineWeb NVMeVerdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (1.032x ➖, 0↑ 2↓)
datafusion / vortex-compact (0.985x ➖, 0↑ 0↓)
datafusion / parquet (0.966x ➖, 1↑ 0↓)
duckdb / vortex-file-compressed (1.018x ➖, 0↑ 1↓)
duckdb / vortex-compact (1.001x ➖, 0↑ 0↓)
duckdb / parquet (0.988x ➖, 0↑ 0↓)
File Size Changes (1 files changed, -0.1% overall, 0↑ 1↓)
Totals:
|
Contributor
Benchmarks: TPC-H SF=1 on NVMEVerdict: No clear signal (environment too noisy confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (1.040x ➖, 0↑ 0↓)
datafusion / vortex-compact (1.025x ➖, 0↑ 0↓)
datafusion / parquet (1.053x ➖, 0↑ 4↓)
datafusion / arrow (1.060x ➖, 0↑ 5↓)
duckdb / vortex-file-compressed (1.039x ➖, 0↑ 0↓)
duckdb / vortex-compact (1.038x ➖, 0↑ 2↓)
duckdb / parquet (1.054x ➖, 0↑ 3↓)
duckdb / duckdb (1.034x ➖, 0↑ 1↓)
File Size Changes (9 files changed, +0.1% overall, 7↑ 2↓)
Totals:
|
Contributor
Benchmarks: TPC-DS SF=1 on NVMEVerdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (0.985x ➖, 0↑ 1↓)
datafusion / vortex-compact (0.972x ➖, 3↑ 0↓)
datafusion / parquet (0.993x ➖, 3↑ 1↓)
duckdb / vortex-file-compressed (1.001x ➖, 0↑ 1↓)
duckdb / vortex-compact (1.000x ➖, 0↑ 1↓)
duckdb / parquet (0.997x ➖, 0↑ 1↓)
duckdb / duckdb (0.982x ➖, 1↑ 0↓)
File Size Changes (7 files changed, +0.0% overall, 4↑ 3↓)
Totals:
|
Contributor
Benchmarks: FineWeb S3Verdict: No clear signal (environment too noisy confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (1.191x ➖, 0↑ 3↓)
datafusion / vortex-compact (0.986x ➖, 0↑ 0↓)
datafusion / parquet (0.951x ➖, 1↑ 0↓)
duckdb / vortex-file-compressed (1.038x ➖, 0↑ 1↓)
duckdb / vortex-compact (1.024x ➖, 0↑ 0↓)
duckdb / parquet (1.000x ➖, 0↑ 0↓)
|
Contributor
Benchmarks: Statistical and Population GeneticsVerdict: No clear signal (low confidence) How to read Verdict and Engines
duckdb / vortex-file-compressed (0.998x ➖, 0↑ 0↓)
duckdb / vortex-compact (0.997x ➖, 0↑ 0↓)
duckdb / parquet (0.997x ➖, 0↑ 0↓)
File Size Changes (1 files changed, +0.0% overall, 1↑ 0↓)
Totals:
|
Contributor
Benchmarks: Clickbench Sorted on NVMEVerdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (0.991x ➖, 1↑ 1↓)
datafusion / parquet (1.001x ➖, 0↑ 0↓)
duckdb / vortex-file-compressed (1.175x ❌, 0↑ 7↓)
duckdb / parquet (1.024x ➖, 0↑ 0↓)
duckdb / duckdb (1.059x ➖, 0↑ 2↓)
File Size Changes (201 files changed, -0.0% overall, 95↑ 106↓)
Totals:
|
Contributor
Benchmarks: TPC-H SF=10 on NVMEVerdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (1.055x ➖, 0↑ 6↓)
datafusion / vortex-compact (0.988x ➖, 0↑ 1↓)
datafusion / parquet (1.013x ➖, 0↑ 2↓)
datafusion / arrow (1.095x ➖, 0↑ 10↓)
duckdb / vortex-file-compressed (1.109x ❌, 0↑ 10↓)
duckdb / vortex-compact (1.066x ➖, 0↑ 6↓)
duckdb / parquet (1.038x ➖, 0↑ 5↓)
duckdb / duckdb (1.047x ➖, 0↑ 4↓)
File Size Changes (27 files changed, +0.1% overall, 19↑ 8↓)
Totals:
|
Contributor
Benchmarks: Clickbench on NVMEVerdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (1.046x ➖, 1↑ 10↓)
datafusion / parquet (1.022x ➖, 2↑ 5↓)
duckdb / vortex-file-compressed (1.093x ➖, 0↑ 19↓)
duckdb / parquet (1.031x ➖, 0↑ 8↓)
duckdb / duckdb (1.015x ➖, 0↑ 3↓)
File Size Changes (101 files changed, +0.0% overall, 50↑ 51↓)
Totals:
|
Contributor
Benchmarks: TPC-H SF=1 on S3Verdict: No clear signal (environment too noisy confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (1.103x ➖, 0↑ 2↓)
datafusion / vortex-compact (1.204x ➖, 0↑ 7↓)
datafusion / parquet (1.125x ➖, 0↑ 4↓)
duckdb / vortex-file-compressed (1.053x ➖, 0↑ 0↓)
duckdb / vortex-compact (1.053x ➖, 0↑ 0↓)
duckdb / parquet (1.007x ➖, 1↑ 0↓)
|
Contributor
Benchmarks: Random AccessVortex (geomean): 1.002x ➖ How to read Verdict and Engines
unknown / unknown (1.006x ➖, 1↑ 0↓)
|
Contributor
Benchmarks: Appian on NVMEVerdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (1.047x ➖, 0↑ 0↓)
datafusion / parquet (1.060x ➖, 0↑ 3↓)
duckdb / vortex-file-compressed (0.981x ➖, 0↑ 0↓)
duckdb / parquet (1.006x ➖, 0↑ 0↓)
duckdb / duckdb (1.024x ➖, 0↑ 0↓)
File Size Changes (4 files changed, -0.0% overall, 2↑ 2↓)
Totals:
|
Contributor
Benchmarks: TPC-H SF=10 on S3Verdict: No clear signal (environment too noisy confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (1.061x ➖, 0↑ 0↓)
datafusion / vortex-compact (0.863x ➖, 2↑ 2↓)
datafusion / parquet (1.039x ➖, 0↑ 1↓)
duckdb / vortex-file-compressed (0.999x ➖, 0↑ 0↓)
duckdb / vortex-compact (0.988x ➖, 0↑ 0↓)
duckdb / parquet (0.990x ➖, 0↑ 0↓)
|
Contributor
Benchmarks: CompressionVortex (geomean): 1.002x ➖ How to read Verdict and Engines
unknown / unknown (0.998x ➖, 2↑ 0↓)
|
connortsui20
pushed a commit
that referenced
this pull request
Jul 14, 2026
<!-- ccr-slack-attribution --> _Requested by **Connor Tsui** · [Slack thread](https://spiraldb.slack.com/archives/C09QYV51VL6/p1784056443357189?thread_ts=1784055867.604669&cid=C09QYV51VL6)_ Pins OnPair's dictionary-training seed to `42` so compression is deterministic run-to-run. OnPair's default config left `seed: None`, which seeds the dictionary trainer's RNG from OS entropy — so the trained dictionary, and therefore compressed file sizes, varied between runs. That noise showed up in the file-size benchmark comments (e.g. #8753). This sets `DEFAULT_DICT12_CONFIG` (the only OnPair `Config` in the repo) to `seed: Some(42)`, making the production path, benches, and tests reproducible. Published `onpair` 0.0.4 already exposes the `seed` field, so no upstream release or version bump is needed. Co-authored-by: Claude <noreply@anthropic.com>
Separate candidate evaluation from selection policy in the cascading compressor (#7697). Schemes produce model-independent candidate evidence via Scheme::evaluate (replacing expected_compression_ratio); the compressor combines each CandidateEstimate with selection context into a borrowed Candidate; and the configured CostModel prices candidates, with selection becoming argmin cost against a canonical_cost baseline. The new vortex_compressor::cost module contains a finite, totally ordered Cost, the CostModel trait, the model-facing Candidate, and the default SizeCost, which preserves the historical ratio ordering, canonical acceptance, encoding trees, and array sizes. The evaluation types live in scheme/evaluation.rs and are re-exported from vortex_compressor::scheme. Deferred callbacks no longer observe the current winner or a ratio threshold, so Delta and Sequence always finish their evaluation before the active model decides whether their candidates are competitive. The winner_compress trace span gains a model-defined estimated_cost and drops the estimated-ratio field. This commit contains the core abstraction and the two scheme callbacks (Delta, Sequence) whose selection logic changed. The mechanical call-site migrations to the new evaluate API and the compressor test updates follow in separate commits. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Aqw9S8Q1wSpNjaRhdFQ1uf Signed-off-by: Connor Tsui <connor.tsui20@gmail.com> Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01ENnFv2a1HrjTg4j6syybux
Mechanical follow-up to the CostModel extraction: update every btrblocks
scheme, the built-in dictionary schemes, and the tensor L2 denorm
encoding to implement Scheme::evaluate instead of
expected_compression_ratio.
Each call site is a direct translation of the reshaped evaluation enum
and carries no selection-policy change:
- CompressionEstimate::Verdict(EstimateVerdict::Skip) -> SchemeEvaluation::Skip
- CompressionEstimate::Verdict(EstimateVerdict::AlwaysUse) -> SchemeEvaluation::AlwaysUse
- CompressionEstimate::Verdict(EstimateVerdict::Ratio(r))
-> SchemeEvaluation::Candidate(CandidateEstimate::from_compression_ratio(r))
- CompressionEstimate::Deferred(DeferredEstimate::*) -> SchemeEvaluation::Deferred(DeferredEvaluation::*)
along with the corresponding import renames. The two schemes whose
deferred callbacks changed behavior (Delta and Sequence) are in the
preceding commit.
Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ENnFv2a1HrjTg4j6syybux
Adapt the cascading-compressor tests to the evaluate/SchemeEvaluation API and the new cost-model selection: candidates are now priced by the configured CostModel (default SizeCost) rather than compared by ratio in the scheme layer, and deferred callbacks no longer receive a best_so_far threshold. Signed-off-by: Connor Tsui <connor.tsui20@gmail.com> Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01ENnFv2a1HrjTg4j6syybux
connortsui20
force-pushed
the
ct/compressor-cost-model
branch
from
1376bf5 to
cb878e4
Compare
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.
Tracking Issue: #8701
Supersedes #8698, rebased onto the reorganized compressor modules (#8745).
Rationale for this change
The compressor's selection objective — maximize estimated compression ratio — is hardcoded into the selection loop, so it cannot be varied per workload (#7697). This PR separates candidate evaluation from selection policy: schemes produce model-independent candidate evidence, the configured
CostModelprices eachCandidate, and the compressor picks the lowest-cost candidate.The default
SizeCostpreserves today's behavior exactly: same winners, same canonical-acceptance threshold, same encoding trees and sizes (golden snapshots unchanged).What changes are included in this PR?
Details
Scheme::evaluatereplacesexpected_compression_ratio. It returns aSchemeEvaluation: skip, forceAlwaysUse, produce an opaqueCandidateEstimate, or request aDeferredEvaluation(sampling or a callback).vortex_compressor::costmodule holds the finite, totally orderedCost, theCostModeltrait, the defaultSizeCost, and the borrowedCandidateview (scheme, estimate, input, optional compressed sample, cascade ancestry) handed toCostModel::cost. Selection retains and compares onlyCost; compression ratio is transient evidence consumed bySizeCost.winner_compresstrace span gains a model-definedestimated_costand drops the estimated-ratio field; achieved ratio remains post-compression observability.canonical_cost, and sampled evidence is exposed correctly.Decisions that deliberately remain outside the model boundary: constant handling,
AlwaysUse, the final byte-acceptance gate, and extension-versus-storage selection. Not included here: model-driven pruning, scheme priors, builder/session plumbing, or a production time-based model.What APIs are changed? Are there any user-facing changes?
Breaking for scheme implementors (
changelog/break):Scheme::expected_compression_ratio→Scheme::evaluate.CompressionEstimate,EstimateVerdict,DeferredEstimate, andEstimateFnare replaced bySchemeEvaluation,ResolvedEvaluation,DeferredEvaluation,DeferredEvaluationFn, andCandidateEstimate, all re-exported fromvortex_compressor::scheme. Deferred callbacks lose their best-so-far argument.New API:
cost::{Cost, CostModel, SizeCost, Candidate}andCascadingCompressor::with_cost_model.