feat(intel): run-all measures every provider through the production job path; fix: zero-length mpack objects are readable (#8 Phase D)
intel_experiment run-all <repo_id> runs the 15 code providers sequentially in dependency order through worker._run_intel_job (prerequisite check, supersession check, atomic persist, outcome recording), with the same gates, never-claimable job row and hard SIGALRM stop as 'run', and records per-provider seconds / peak RSS / outcome. The runner passes the worker's LOCAL object-store read settings (refused unless the endpoint is local) so snapshot providers read real mpack objects.
Real data found a bug the fixtures could not: read_object_bytes returns None for a zero-length object stored in an mpack (the byte-range read has nothing to fetch), so the fail-closed unreadable-object check failed the type provider for every repository with an empty __init__.py. An object whose recorded size is 0 is now an empty file.
Measured (1 GiB, no swap, 1 CPU, 60 s hard limit, --init): 100 reachable commits -> all 15 providers ok in 49.8 s, 169 MiB peak (index providers <= 0.17 s each, snapshot providers 7-10 s each). 300 reachable commits -> hit the 60 s limit; the run was stopped there (limits not raised).
Semantic Changes
8 symbols
0 comments
muse hub commit comment sha256:78744d379faf9e46bfc8a45b2998d385b92f232df4fd33a1d06bfe62efa0b1d2 --body "your comment"
No comments yet. Be the first to start the discussion.