Deferred follow-up from the #907 review (inline thread on http_connection_manager.py).
cancel_and_await_if_same_loop now (correctly) propagates unrelated task-cleanup exceptions. But its callers treat the whole teardown sequence as if it can't raise, so a propagated error aborts teardown midway:
ConnectionManager._close_on_owner_loop: an escape between self._closed = True and the session/connector close leaves the session open — and because _claim_close_request sees _closed, a retry close() is a no-op. The session leaks permanently and unrecoverably (probed during review).async_execution_manager.py (~:359): same shape — an abort skips the remaining task cancels, all state nulling, and active-execution cancellation.result_cache.py (~:237): same shape.
Not reachable today: every SDK background loop breaks on CancelledError, so cleanup can't currently raise — this is latent hardening, which is why it was split out of #907 rather than folded in.
Fix shape: wrap the cancel/await steps in try/finally so resource close + state nulling always run, and re-raise after. One test per site forcing a cleanup exception and asserting the session/connector still gets closed and state is nulled.
Deferred follow-up from the #907 review (inline thread on
http_connection_manager.py).cancel_and_await_if_same_loopnow (correctly) propagates unrelated task-cleanup exceptions. But its callers treat the whole teardown sequence as if it can't raise, so a propagated error aborts teardown midway:ConnectionManager._close_on_owner_loop: an escape betweenself._closed = Trueand the session/connector close leaves the session open — and because_claim_close_requestsees_closed, a retryclose()is a no-op. The session leaks permanently and unrecoverably (probed during review).async_execution_manager.py(~:359): same shape — an abort skips the remaining task cancels, all state nulling, and active-execution cancellation.result_cache.py(~:237): same shape.Not reachable today: every SDK background loop breaks on
CancelledError, so cleanup can't currently raise — this is latent hardening, which is why it was split out of #907 rather than folded in.Fix shape: wrap the cancel/await steps in
try/finallyso resource close + state nulling always run, and re-raise after. One test per site forcing a cleanup exception and asserting the session/connector still gets closed and state is nulled.