[WORKFLOW SDK FEATURE REQUEST] Retry WaitForInstanceCompletion/Start on a server-sent CANCELLED

[CasperGN]

Proposal

wait_for_workflow_completion and wait_for_workflow_start (sync and aio) should re-issue WaitForInstanceCompletion / WaitForInstanceStart when the server ends the call with CANCELLED and the caller's own timeout hasn't expired. Today that status reaches the caller as an error, even though the workflow is still running. This asks maintainers to decide on the behaviour first, because it reverses assertions added in #1112. It's a proposal, not a PR.

Why the server sends CANCELLED

When daprd shuts down or restarts mid-wait (rollout, pod restart, hot-reload, a config change that restarts the runtime), the actors router cancels every in-flight call it tracks. The wait then returns CANCELLED "context canceled", even though the caller is still waiting. The mechanism and a repro are in dapr/dapr#10566.

Reproduced against native daprd 1.18.0:

• a 60 s workflow with wait_for_workflow_completion(id) and no timeout returns normally;
• the same run with daprd sent SIGTERM about 10 s in fails with StatusCode.CANCELLED "context canceled".

Why retrying is safe

• A blocking unary call only gets CANCELLED from the server. The sync client can't cancel it mid-flight, and in the aio client a caller's cancellation raises asyncio.CancelledError, not AioRpcError. A client-side deadline shows up as DEADLINE_EXCEEDED, which already maps to TimeoutError.
• The wait only reads state, so re-issuing it is idempotent. It returns straight away if the workflow already finished.

What would change

• Add CANCELLED to the retried codes, when the caller's timeout hasn't expired, in _durabletask/client.py L335 and _durabletask/aio/client.py L230. _is_deadline_cancellation (L349) already turns a CANCELLED after the deadline into TimeoutError.
• Restart the 30 s transient-retry window after each successful re-issue, rather than anchoring it on the first transient error. Otherwise a sidecar that restarts more than once during a long wait would still fail it.
• These tests from Add stateful-history delta delivery to the workflow worker #1112 currently assert that a CANCELLED reaches the caller, and would flip:
  • test_cancelled_within_deadline_still_propagates
  • test_cancelled_unbounded_wait_still_propagates
  • test_cancelled_within_deadline_still_propagates (aio)

Relationship to the runtime fix

The proper fix is in daprd: return UNAVAILABLE in this case (dapr/dapr#10566). The SDK already retries UNAVAILABLE. This change would protect users on current and older runtimes until that ships, and it would still help afterwards with proxies that reset the stream.

Question for maintainers

Is it acceptable to treat a server-sent CANCELLED on these two waits as retryable, and change the #1112 tests? If yes, it's a small change in both clients plus tests.