[Bug] non determinism when using `asyncio.gather` with local activities

test_workflow_concurrent_local_activity_replay の最小再現から始め、Replayer を通して実行し、asyncio.gather 内の execute_local_activity 呼び出しに焦点を当てます。初回実行時とリプレイ時のローカルアクティビティの完了順序を比較します。workflow が完了し、その履歴を NondeterminismError なしでリプレイできれば完了です。

What are you really trying to do?

At a high level, our code is using some local activities at the beginning and end of some workflows parts. We got some reports of non determinism errors when those workflow parts are called using asyncio.gather.

Describe the bug

According to my understanding (might be wrong of course):
During first execution, local activity resolve_activity jobs arrive in completion order (wall-clock timing). During replay, they arrive in sequence number order. This ordering difference causes coroutines to resume in a different order, assigning different sequence numbers to subsequent commands, which no longer match the recorded markers.

Minimal Reproduction

Extracted from: https://github.com/temporalio/sdk-python/pull/1573

@activity.defn
async def local_activity_slow(index: int) -> None:
    if index % 2 == 0:
        await asyncio.sleep(0.05)


@activity.defn
async def local_activity_fast(index: int) -> None:
    return None


@activity.defn
async def local_activity_gate() -> None:
    await asyncio.sleep(0.05)


@workflow.defn
class ConcurrentLocalActivityReplayWorkflow:
    """Workflow that runs two concurrent coroutines with local activities.

    This reproduces a replay nondeterminism bug: during first execution,
    local activities take real time, creating a deterministic interleaving.
    During replay, all local activities return instantly from markers, which
    can reorder coroutine scheduling and produce a different command sequence.
    """

    @workflow.run
    async def run(self) -> list[int]:
        async def lifecycle_a(index: int) -> int:
            await workflow.execute_local_activity(
                local_activity_slow,
                args=[index * 2],
                start_to_close_timeout=timedelta(seconds=5),
            )
            await workflow.execute_local_activity(
                local_activity_fast,
                args=[index * 2],
                start_to_close_timeout=timedelta(seconds=5),
            )
            return index * 2

        async def lifecycle_b(index: int) -> int:
            await workflow.execute_local_activity(
                local_activity_gate,
                start_to_close_timeout=timedelta(seconds=5),
            )
            await workflow.execute_local_activity(
                local_activity_slow,
                args=[index * 2 + 1],
                start_to_close_timeout=timedelta(seconds=5),
            )
            await workflow.execute_local_activity(
                local_activity_fast,
                args=[index * 2 + 1],
                start_to_close_timeout=timedelta(seconds=5),
            )
            return index * 2 + 1

        results: list[int] = []
        for index in range(20):
            results.extend(await asyncio.gather(lifecycle_a(index), lifecycle_b(index)))
        return results


async def test_workflow_concurrent_local_activity_replay(client: Client):
    """Test that concurrent local activities replay deterministically.

    Runs a workflow with two concurrent coroutines that each issue multiple
    local activities, then replays the history. Without the fix, replay
    fails with NondeterminismError because the local activity command order
    diverges from the recorded marker order.
    """
    async with new_worker(
        client,
        ConcurrentLocalActivityReplayWorkflow,
        activities=[local_activity_slow, local_activity_fast, local_activity_gate],
    ) as worker:
        handle = await client.start_workflow(
            ConcurrentLocalActivityReplayWorkflow.run,
            id=f"workflow-{uuid.uuid4()}",
            task_queue=worker.task_queue,
        )
        expected = [v for i in range(20) for v in (i * 2, i * 2 + 1)]
        assert await handle.result() == expected

        history = await handle.fetch_history()

    await Replayer(
        workflows=[ConcurrentLocalActivityReplayWorkflow],
    ).replay_workflow(history)

Environment/Versions

• OS and processor: M1 Mac and Linux at least (I believe this is more general than that)
• Temporal Version: server=1.30.4, sdk=1.27.2, but also reproducing on main of sdk-python
• Building Temporal server from source