steveahnahn opened a new pull request, #69676:
URL: https://github.com/apache/airflow/pull/69676
Supersedes #69631 (closed during a verification rework; GitHub cannot reopen
a PR after its branch is rebased).
### Problem
`RedshiftDataOperator.on_kill` cancels the running statement, but once a
deferrable run enters the triggerer that method can no longer run: the worker
process exits at `defer()`. When a user then marks the deferred task failed,
clears it, or marks it success, `RedshiftDataTrigger` had no `on_kill` hook
(the `BaseTrigger` default is a no op), so the statement kept executing against
the cluster or workgroup. The operator's documented kill semantics silently
stop applying the moment the task defers. On Redshift Serverless the orphaned
statement keeps billing RPU hours, and on a provisioned cluster it holds a WLM
slot, until it finishes on its own.
Sibling deferrable triggers already cancel their remote work on user kill:
EMR in this provider, and BigQuery, Dataproc and Dataflow in the Google
provider.
### Change
`RedshiftDataTrigger` gains an `on_kill` that cancels the running statement
through the hook's async connection (`cancel_statement`). A `cancel_on_kill`
flag (default `True`) is added to both `RedshiftDataOperator` and
`RedshiftDataTrigger` so users can opt out, and the operator threads it into
the trigger at the defer site. Cancellation is best effort: failures while
cancelling are logged and swallowed so trigger teardown is not blocked.
### Live verification
Verified in a running Airflow instance (`airflow standalone`: real
scheduler, triggerer and API server) against a live `redshift-data` API
endpoint served by moto, with no mocks anywhere in the execution path:
1. A deferrable `RedshiftDataOperator` submitted a statement
(`ExecuteStatement` over HTTP, status `STARTED`) and the task reached
`deferred`, with the trigger polling the statement from the triggerer's event
loop through the real aiobotocore client.
2. Marking the deferred task failed from the UI drove the user action cancel
path. Triggerer log:
```
Trigger cancelled by user action, invoking on_kill
Cancelling Redshift statement 93e26017-b365-4c9f-9c58-9d387af5e6ea.
Redshift statement 93e26017-b365-4c9f-9c58-9d387af5e6ea cancelled.
```
No exception escaped `on_kill`.
3. `DescribeStatement` afterwards returned `Status: ABORTED`, and without
the change the same kill leaves the statement `STARTED` (the base `on_kill` is
a no op, so nothing contacts the API).
This exercises the full deferred kill chain end to end: trigger
serialization round trip, `on_kill` invocation by the triggerer, async
connection creation inside the event loop, and the `CancelStatement` HTTP call.
What it does not cover is AWS side cancellation semantics; the trigger issues
the same `cancel_statement` call the operator's existing non deferred `on_kill`
already ships and the unit tests cover.
### Tests
Trigger tests cover cancel on kill, the disabled and missing statement id no
ops, best effort error swallowing, and serialization of the new flag. Operator
tests cover the `cancel_on_kill=False` guard and that the flag is threaded into
the trigger on defer. Every new assertion fails without the source change.
---
##### Was generative AI tooling used to co-author this PR?
- [X] Yes, Claude Code (Fable 5)
Generated-by: Claude Code (Fable 5) following [the
guidelines](https://github.com/apache/airflow/blob/main/contributing-docs/05_pull_requests.rst#gen-ai-assisted-contributions)
--
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]
