Repository navigation
Conversation
…ients worker.php already waits, inside the original 30 s budget, for the Worker's acknowledgement of its query result before it stops: the independent caller can see the committed result first. The Python and Rust clients stopped immediately, so a fast caller could end the Worker before its acknowledgement arrived (#708 attempt 1, MySQL 8.0, PHP target, timely). Both now wait the same bounded way. The submission and acknowledgement assertions are unchanged.
|
Security focused source review at exact 1cbe01f: accepted in principle. The Python and Rust clients now wait for the actual acknowledgement before initiating Worker shutdown, use the unchanged process-wide 30 s deadline, retain the terminal submission/acknowledgement count, authority, protocol, counter and no-history-mutation assertions, and do not add an observer retry or widen a product budget. The Rust change makes explicit a wait that previously passed only because shutdown happened to drain in-flight work. This correction is the right scope for the attempt-1 MySQL race. It does not erase that original and it does not resolve the independent SQLite poll refusal, now owned by #721 / Python #218. Final acceptance still needs the exact PR checks and its MySQL-family query-authority cohort to pass, followed by integration with current main and #708 before a fresh combined run. No merge or publication disposition is implied by this source acceptance. |
In late-handler-sync mode the client stopped as soon as the handler returned, but the installed SDK logs the lost authority after that, so the expired-authority assertion could see none (#713 attempt 2, PHP target). worker.php waits for its Worker's refusal; every non-timely Python mode now also waits, inside the 30 s original budget, for the expired-authority evidence before stopping. Assertions are unchanged.
|
Security source re-review at replacement head The wait remains inside the original process-wide 30-second No Security source blocker remains on this exact harness head. Fresh normal checks and the corrected query-authority cohort are still required. This correction classifies only the empty expiry-evidence race; it does not classify or rescue #713s MySQL 8.4 codec or two MySQL execution failures, and it does not resolve Server #721 / Python #218 poll-pressure behavior. |
|
|
Hosted result at
|
The query-authority Python and Rust clients now wait for the Worker's acknowledgement before stopping, as
worker.phpdoes (#325).What happened. In #708 attempt 1, the MySQL 8.0 job (run 40017239524, job 120644918636) failed with
target=php language=python mode=timely protocol=1.20. The run recorded one claim, one delivery, one submission and one returned result, but no acknowledgement. The query's authority was still valid and nothing was lost.worker.phpalready handles this race: the independent caller can see the committed result before the Worker receives the acknowledgement of its submission, so it waits for the acknowledgement inside the original budget before stopping.query-authority/python.pystopped the Worker immediately.query-authority/rust/src/main.rsdid too, and passed only because its shutdown drains in-flight requests.Change. In timely mode both clients now wait, inside the original 30 s budget, for the
acknowledgementevent before stopping the Worker. A Worker that exits early or a budget overrun still fails. Nothing else changes:submission == acknowledgement == 1for timely;Checked.
python.pyparses. The Rust client builds with its locked dependencies. The cohort runs on this PR's CI.Security's narrowing is on private Discussion 438. The separate SQLite poll-refusal observation from the same attempt has its own issue.