ci: confirm cargo bench regressions with a targeted rerun - #38810
Open
antiguru wants to merge 1 commit into
Open
ci: confirm cargo bench regressions with a targeted rerun#38810antiguru wants to merge 1 commit into
antiguru wants to merge 1 commit into
Conversation
The first scheduled run of the cargo-bench step flagged regressions of up to +167% on inputs of 8M and 128M rows. None reproduced: local runs at the same commits flagged different rows each time, and a rerun of one pair of identical binaries moved a third set. Criterion's interval only covers the variance inside one process, while the variance between processes on large inputs is far above the threshold. A benchmark that regresses in the first pass is now measured again, both sides, filtered to exactly the flagged ids through criterion's regex filter, into a separate criterion home so the first pass stays intact. A regression counts only when both passes agree. A rerun that disagrees, or that produced no result, leaves the row as `unconfirmed`, shown in the table with its rerun delta and never failing the step. The cost scales with the number of flagged rows, nothing when there are none. The results table also opens its own Buildkite log group instead of landing inside the last benchmark's collapsed section. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Motivation
The first scheduled run of the cargo-bench step (nightly build 18289) flagged regressions of +21% to +167% across four shards. None of them held up. Local runs at the same two commits flagged a different set of rows each time, and rerunning one pair of identical binaries flagged a third set, always on the 8M and 128M inputs and always with a confidence interval under one percentage point. Criterion's interval covers the variance inside one process. The variance between processes on large inputs is what dominates, and it is far above the 10% threshold.
Description
A benchmark that regresses in the first pass is measured again, both sides, filtered to exactly the flagged benchmark ids through criterion's regex filter, into a separate criterion home so the first pass stays intact. A regression counts only when both passes agree. A rerun that disagrees, or that produced no result, leaves the row as
unconfirmed, shown in the table with its rerun delta and never failing the step. The cost scales with the number of flagged rows and is zero when there are none, so it adds nothing to the common case.The results table now opens its own Buildkite log group instead of landing inside the last benchmark's collapsed section.
Smoke on mz-repr across the Row decode change with
--threshold 0.0: 21 first-pass regressions, 8 confirmed, 13 unconfirmed. One row read +48.2% in the first pass and −0.3% on the rerun.Unit tests cover a confirmed regression, a downgraded one, and a rerun missing the benchmark.
Posted by Claude Code.
🤖 Generated with Claude Code