[linhongyu510]

Cancelling an ordinary asynchronous action currently records its step as completed, even though its state update never ran. Rebuilding the application with resume_at_next_action=True then skips that action and advances to its successor. For example, cancelling work while it waits for a response resumes at after with the old state.

Changes

• Preserve the original asyncio.CancelledError for _astep's post_run_step hooks before re-raising it. Persisters consequently record the step as failed and resume at the interrupted action.
• Cover astep and arun, with both function-based single-step actions and class-based multi-step actions. The tests cancel at an event-controlled await, rebuild from persisted state, retry, and verify the final result.

How I tested this

• On upstream 36e3ff3048c8cabd7aac6d5c373307934bf3c2df, all four new cases fail because the next action after recovery is after rather than work.
• With the fix, pytest tests/core/test_application.py tests/core/test_persistence.py tests/core/test_parallelism.py -q: 256 passed on Python 3.12.13.
• Black 23.11.0, isort 5.12.0, Flake8 6.1.0, ASF license-header checks, and git diff --check pass.
• Complete pre-commit run --files burr/core/application.py tests/core/test_application.py: passed, including the frontend lint-staged and ASF header hooks.

Notes

This concerns non-streaming step persistence. The streaming cleanup change in #1027 is already in the base. The top-level execute-hook reporting in #897 does not update _astep's step hooks or persistence status.

The existing cancellation still propagates to the caller. This does not shield persistence from repeated cancellation or change the handling of cancellation during a lifecycle hook.

This change was developed with an AI assistant. The reproduction and regression checks use no model or network service.

Checklist

• PR has an informative and human-readable title.
• Changes are limited to a single goal.
• Code passed the complete pre-commit check.
• The behavior change is tested.
• No new public functions or API are introduced.
• No placeholder code is added.
• The change corrects existing retry behavior without changing documented APIs.

[enekoreto]

The fix looks right to me. I checked out the branch (2a36017) on main (36e3ff3): the 4 new tests fail on main (resume goes to after) and pass here on Python 3.13, CI's test command passes, and pre-commit passes. I also tried it outside the tests with AsyncInMemoryPersister: a cancelled astep or arun is now saved as failed and resumes at work, and the local tracker now records the CancelledError on the step. A cancelled streaming action through astream_result was already saved as failed on main, and still is.

But the new tests fail on Python 3.10, which CI runs: all 4 stop at assert tracker.post_called[0][1]["exception"] is cancelled.value. On 3.10, awaiting a cancelled task raises a new CancelledError instead of the one raised inside the action, so the two aren't the same object (on 3.13 they are). isinstance(tracker.post_called[0][1]["exception"], asyncio.CancelledError) passes on both. The rest of the suite passes on 3.10.

Small thing, not blocking: Ctrl+C has the same problem, since KeyboardInterrupt isn't an Exception either. A KeyboardInterrupt inside an action, in step or astep, is still saved as completed and skipped on resume, on main and here. Catching it next to CancelledError in _astep, plus the same clause in _step, fixed it in my test, and the suite still passes. That could also be a follow-up.

I ran this on Python 3.13 on macOS, and the tests also on 3.10.