Repository navigation
Flaky: field_expiry_ms dump/restore/migrate test reads a field deadline back up to 2 ms late under load #494
Description
tests/field_expiry_ms.rs a_field_deadline_is_kept_to_the_millisecond_by_dump_restore_and_migrate failed once in a full parallel cargo test --features lua,tls run on an Azure D16ps_v5. That was during the #490 trials, but the code involved isn't changed by #490. After MIGRATE, 1 of 900 field deadlines read back 2 ms late (h74, field a: …979 instead of …977).
Likely cause: FieldClock (src/commands/helpers.rs) reads the monotonic and wall clocks as a pair. Since #488 it rounds so that deadlines are never early, so a thread preempted between the reads makes a deadline late. Retries cap the gap at 50 µs, but the conversion happens twice (save and load), and rounding to the millisecond can turn a small gap into 1-2 ms.
Options: let the test allow a small late tolerance (never early), or carry the absolute Unix ms through DUMP/RESTORE and MIGRATE so they don't convert twice. Found during #490's trials (3 full runs; 1 failure in 1 run).
Activity