Harden corpus scraper + first pair-hunting run - #20
Merged
Conversation
Running the pairs scrape surfaced that GitHub tightened code-search query parsing: a bare "--array" / "-l" / "-W" now leads with a dash and is read as a NOT operator, so every slurm and pbs query returned 422 (the corpus scraper was silently broken for both). Quote the directive phrases to fix. Also make get_with_retry survive the transport failures GitHub throws under its abuse-detection throttling, instead of crashing the run: - honor Retry-After on 403/429 secondary rate limits; - retry on dropped connections and on 408/5xx from the search backend; - guard each phase-1 query so one failure can't abort the sweep. And clear the pairs output dir on start so stale repo dirs from a prior run can't misrepresent the current results. First pairs run (committed manifest): 3 candidate pair-repos, all volcano<->kueue. NVIDIA/knavigator is a genuine equivalent pair (the same benchmark job rendered per scheduler) and is usable as conversion ground truth; the others co-locate volcano+kueue examples that still need review. Repo-level co-location finds candidates cheaply but does not prove equivalence — a verification pass is still required. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016PnTz6Zxqa4jHocK8kCbyx
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
What
Ran the pairs scrape end-to-end (build-plan stage 2), which exposed and fixed
several scraper bugs, and produced the first real cross-scheduler pair candidates.
Scraper fixes
--array/-l/-Wleads with a dash and is now parsed as a NOT operator,so every slurm and pbs query returned 422 — the corpus scraper was silently
broken for both HPC schedulers. Quoting the phrases (
"#SBATCH --array") fixes it.Retry-Afteron 403/429 secondary (abuse) rate limits;RemoteDisconnected) and on 408/5xx from thesearch backend;
can't misrepresent the current results.
First pairs run
3 candidate pair-repos, all volcano↔kueue (manifest committed; specs stay
gitignored per corpus policy):
{{._NAME_}}, N replicas, ubuntu, 8 GPUs, identical kwok annotations) rendered once per scheduler. Usable as conversion ground truth now.Finding
Repo-level co-location surfaces candidates cheaply, but does not prove
equivalence. A false positive from an earlier run (
converged-computing/ fractale-experiments, slurm+pbs) turned out to be unrelated jobs a repo merelycollects. The gold seam is benchmarking/harness repos like knavigator that
deliberately express one workload across schedulers. HPC↔HPC and HPC↔K8s real-world
co-located pairs remain scarce, matching the earlier corpus cross-reference.
Testing
--listunchanged; pythonast.parseclean.🤖 Generated with Claude Code