Reading poses in parallel multithreaded may cause errors #177

Activity

1. AmitMY commented on Jul 23, 2026

   

   AmitMY

   on Jul 23, 2026

   CollaboratorAuthor

   More actions

   Fixed in #239: the root cause was PoseHeaderCache being updated non-atomically — concurrent Pose.read calls on files with different headers could crash ("buffer is too small for requested array") or silently return a pose parsed with another file's header. All cache access now goes through a lock, and check_cache returns the header and end offset from the same critical section. The test this issue asked for exists as test_concurrent_read_of_different_files_is_safe (fails in under a second on the old code; an external repro went from 122 exceptions + 153 wrong headers to 0 across 4,800 concurrent reads).
2. AmitMY

   closed this as completedon Jul 23, 2026