Repository navigation
Release v3.3.0: version bumps, CHANGELOG release section, README sync - #1819
Conversation
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… zero-warning Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…key UPDATE The #1814 staleness test aged its snapshot by UPDATEing collection_time to now-2h. collection_time is the hypertable's partition key and TimescaleDB never re-routes a row on UPDATE, so between 00:00 and 02:00 UTC the new value crosses the midnight chunk boundary and violates the chunk's slice CHECK constraint (23514) - a nightly two-hour time bomb that passed its own CI and two nightlies because they all ran outside the window. First detonation was this release PR's 00:42 UTC run. A DELETE + re-INSERT routes to the correct chunk at any hour; the stale and relaxed-profile assertions are unchanged. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
The Darling PG failure at 00:42 UTC was a real find, not a flake: the #1814 staleness test aged its snapshot by UPDATEing collection_time (the partition key) to now-2h — TimescaleDB never re-routes rows on UPDATE, so between 00:00 and 02:00 UTC that crosses the midnight chunk boundary and violates the chunk slice CHECK constraint (23514). It passed its own CI (19:16) and two nightlies (19:16, 20:39) because all ran outside the two-hour nightly window; this PR's run was the first inside it. Fixed here by DELETE + re-INSERT at the aged timestamp (chunk routing then works at any hour), assertions unchanged, one-site sweep confirmed no siblings. The re-run on this head executes inside the same window — green there is the in-window proof of the mechanism. |
…ction Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Release prep for v3.3.0, per the release checklist:
No product code changes beyond the CS0105 fix.
🤖 Generated with Claude Code