[andystaples]

Summary

Add two large-payload regression cases using the existing in-memory backend, fake payload store, and real Python gRPC worker:

• Remove a payload referenced by an orchestration or activity delivery and observe the affected work item's abandonment.
• Keep two unrelated large-payload activities in progress during the failure.
• Verify fresh work completes before releasing the held activities, then verify both original orchestrations complete with their expected outputs.

The tests use synchronization events rather than timing-dependent sleeps. They run in the existing core test suite without Docker, Azure Storage, a Functions host, or additional CI jobs.

Scope

Existing abandonment behavior is unchanged. The missing payload is restored after abandonment to bound immediate orchestration redelivery; these tests do not assert a new terminal-failure or retry policy.

This is native Python SDK worker isolation coverage, not Functions-host or DTS-provider isolation coverage. It does not validate host-managed payload externalization or the upstream fix for that path.

Closes #312 as superseded by this narrower SDK regression coverage; the proposed DTS-host test matrix is intentionally not implemented.

[Copilot]

🟡 Changes recommended

The regression test is timing-dependent because the healthy activity’s 10-second timeout spans two sequential waits that each independently allow 10 seconds, so valid execution can still fail.

1 open issue

• Avoid overlapping activity timeout with bounded controller waits · New

What changed in this PR

Adds native Python SDK worker regression coverage for missing-payload download isolation, without changing abandonment behavior. This is narrower than the Functions-host/DTS coverage proposed in #312.

Changes:

• Tests missing payloads in both orchestration and activity deliveries.
• Verifies fresh work completes while two unrelated activities remain blocked, then checks their expected outputs.
• Restores missing payloads after abandonment to bound immediate redelivery.

File	Description
tests/​durabletask/​test_large_payload.py	Adds parameterized isolation tests using the in-memory backend, fake payload store, and real gRPC worker.

_{Review effort: Balanced}

[Bernd Verst (berndverst)]

The added cases exercise the intended native-worker missing-download paths and provide useful isolation coverage without changing runtime behavior, performance, or public compatibility. I independently confirmed the existing timeout concern: a six-second missing-download/abandonment stage followed by a six-second fresh activity causes both parameterized cases to fail their original healthy orchestrations even though each controller stage is within its ten-second timeout. Please remove the healthy activity's independent ten-second release deadline and rely on the bounded controller waits plus finally-based release before merging. The broader Functions/DTS acceptance criteria from #312 remain outside this coverage.

[Bernd Verst (berndverst)]

The held-activity timeout race is fixed. I reran the original reproduction with a six-second missing-download delay followed by a six-second fresh-activity delay; both parameterized cases now pass while preserving the isolation and output assertions. Controller-failure cleanup also releases the held activities correctly. No remaining actionable findings in the fix; the coverage remains scoped to native Python SDK worker isolation.