dabla opened a new issue, #77240: URL: https://github.com/apache/airflow/issues/77240
### Apache Airflow version main (development) ### What happened? `CommsDecoder._exchange` writes a request and then reads the next frame as its response. `_get_response` does not check the frame id against the request id. When an exception unwinds `_exchange` between the write and the read, the response stays on the socket and the next request reads it as its own. On the main thread, a signal handler that raises can cause this: - the `AirflowTaskTimeout` raised by the `execution_timeout` alarm, which any task can hit while it makes an SDK call; - a `DeadlockImminentError` raised into a request by a handler that calls `send()` itself (for example `on_kill()` making an SDK call). The task may then get a wrong value back, for example the previous `GetVariable`'s answer for a `GetConnection`, or fail to decode it. ### What you think should happen instead? A request never reads a response that belongs to another request. Possible fixes: check the response id and drain frames that belong to an abandoned request, read a partly read frame to the end before the next request, or mark the channel unusable after an interrupted exchange so the next call fails loudly. ### How to reproduce Make the response of a `send()` arrive after an `execution_timeout` alarm fires while the task waits for it, then make another SDK call in the same process: it returns the first request's response. ### Anything else? Found while working on #76545 (and earlier on #62922). ### Are you willing to submit PR? - [x] Yes I am willing to submit a PR! -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
