perf(database): create the checkpoint programs directory once per save - #478
Open
dexhunter wants to merge 2 commits into
Open
perf(database): create the checkpoint programs directory once per save#478dexhunter wants to merge 2 commits into
dexhunter wants to merge 2 commits into
Conversation
_save_program called os.makedirs(programs_dir, exist_ok=True) for every program written, so a checkpoint at the default population_size=1000 created the same directory 1,000 times. Create it once in save() and pass it down. Also lift the log_prompts and prompts_by_program lookups out of the per-program loop, and write each record with a single f.write(json.dumps(...)) rather than json.dump(..., f). The bytes written are unchanged: checkpoints from an identical seeded 200-program population compare equal by SHA-256 before and after, both with recorded prompts and without. One save() call at the shipped defaults goes from 241.1829 ms to 182.8334 ms of CPU, a 24.0% reduction, measured as the median of 5 repetitions on freshly generated populations.
Keep the programs directory absent when saving an empty database, matching the pre-optimization checkpoint structure.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
ProgramDatabase.save()previously recreated the sameprograms/directory for every program in a checkpoint. This change creates it once when the database is non-empty, hoists loop-invariant prompt lookups, and writes each encoded record to the buffered text stream in one call.The fallback path in
_save_program()is unchanged for its single-program caller. Empty checkpoints also keep their previous layout: they containmetadata.jsonwithout aprograms/directory.Measurement
CPU time for one
ProgramDatabase.save()call on a fresh 1,000-program population using the shipped defaults (num_islands=5,log_prompts=True):mainat411fb59That is a 24.6% reduction. The benchmark uses process CPU time under an exclusive host lock, and every accepted candidate must produce byte-identical checkpoints for empty, one-program, populated, and path-collision cases.
Autoresearch record
Supplementary autoresearch with Weco contributed the optimization direction. The review-required trajectory contains 11 distinct accepted scalar records, including the measured baseline and an exact replay of this PR diff:
https://dashboard.weco.ai/share/xNYBuZK0Xb0aglM7c-SRbnO3ZHLbDO6j
The fastest search result measured 169.1532 ms, but it used a more intrusive serialization path and was not selected. The PR keeps the smaller repository-compatible change.
Validation
OPENAI_API_KEY=test-key-for-unit-tests uv run --extra dev python -m unittest discover tests: 430 tests passed.git diff --check: passed.