Skip to content

prefect.engine

Client-side execution and orchestration of flows and tasks.

Engine process overview

Flows

  • The flow is called by the user or an existing flow run is executed in a new process.

    See Flow.__call__ and prefect.engine.__main__ (python -m prefect.engine)

  • A synchronous function acts as an entrypoint to the engine. The engine executes on a dedicated "global loop" thread. For asynchronous flow calls, we return a coroutine from the entrypoint so the user can enter the engine without blocking their event loop.

    See enter_flow_run_engine_from_flow_call, enter_flow_run_engine_from_subprocess

  • The thread that calls the entrypoint waits until orchestration of the flow run completes. This thread is referred to as the "user" thread and is usually the "main" thread. The thread is not blocked while waiting — it allows the engine to send work back to it. This allows us to send calls back to the user thread from the global loop thread.

    See wait_for_call_in_loop_thread and call_soon_in_waiting_thread

  • The asynchronous engine branches depending on if the flow run exists already and if there is a parent flow run in the current context.

    See create_then_begin_flow_run, create_and_begin_subflow_run, and retrieve_flow_then_begin_flow_run

  • The asynchronous engine prepares for execution of the flow run. This includes starting the task runner, preparing context, etc.

    See begin_flow_run

  • The flow run is orchestrated through states, calling the user's function as necessary. Generally the user's function is sent for execution on the user thread. If the flow function cannot be safely executed on the user thread, e.g. it is a synchronous child in an asynchronous parent it will be scheduled on a worker thread instead.

    See orchestrate_flow_run, call_soon_in_waiting_thread, call_soon_in_new_thread

Tasks

  • The task is called or submitted by the user. We require that this is always within a flow.

    See Task.__call__ and Task.submit

  • A synchronous function acts as an entrypoint to the engine. Unlike flow calls, this will not block until completion if submit was used.

    See enter_task_run_engine

  • A future is created for the task call. Creation of the task run and submission to the task runner is scheduled as a background task so submission of many tasks can occur concurrently.

    See create_task_run_future and create_task_run_then_submit

  • The engine branches depending on if a future, state, or result is requested. If a future is requested, it is returned immediately to the user thread. Otherwise, the engine will wait for the task run to complete and return the final state or result.

    See get_task_call_return_value

  • An engine function is submitted to the task runner. The task runner will schedule this function for execution on a worker. When executed, it will prepare for orchestration and wait for completion of the run.

    See create_task_run_then_submit and begin_task_run

  • The task run is orchestrated through states, calling the user's function as necessary. The user's function is always executed in a worker thread for isolation.

    See orchestrate_task_run, call_soon_in_new_thread

    _Ideally, for local and sequential task runners we would send the task run to the user thread as we do for flows. See #9855.

begin_flow_run async

Begins execution of a flow run; blocks until completion of the flow run

  • Starts a task runner
  • Determines the result storage block to use
  • Orchestrates the flow run (runs the user-function and generates tasks)
  • Waits for tasks to complete / shutsdown the task runner
  • Sets a terminal state for the flow run

Note that the flow_run contains a parameters attribute which is the serialized parameters sent to the backend while the parameters argument here should be the deserialized and validated dictionary of python objects.

Returns:

Type Description
State

The final state of the run

Source code in prefect/engine.py
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
async def begin_flow_run(
    flow: Flow,
    flow_run: FlowRun,
    parameters: Dict[str, Any],
    client: PrefectClient,
    user_thread: threading.Thread,
) -> State:
    """
    Begins execution of a flow run; blocks until completion of the flow run

    - Starts a task runner
    - Determines the result storage block to use
    - Orchestrates the flow run (runs the user-function and generates tasks)
    - Waits for tasks to complete / shutsdown the task runner
    - Sets a terminal state for the flow run

    Note that the `flow_run` contains a `parameters` attribute which is the serialized
    parameters sent to the backend while the `parameters` argument here should be the
    deserialized and validated dictionary of python objects.

    Returns:
        The final state of the run
    """
    logger = flow_run_logger(flow_run, flow)

    log_prints = should_log_prints(flow)
    flow_run_context = FlowRunContext.construct(log_prints=log_prints)

    async with AsyncExitStack() as stack:
        await stack.enter_async_context(
            report_flow_run_crashes(flow_run=flow_run, client=client, flow=flow)
        )

        # Create a task group for background tasks
        flow_run_context.background_tasks = await stack.enter_async_context(
            anyio.create_task_group()
        )

        # If the flow is async, we need to provide a portal so sync tasks can run
        flow_run_context.sync_portal = (
            stack.enter_context(start_blocking_portal()) if flow.isasync else None
        )

        task_runner = flow.task_runner.duplicate()
        if task_runner is NotImplemented:
            # Backwards compatibility; will not support concurrent flow runs
            task_runner = flow.task_runner
            logger.warning(
                f"Task runner {type(task_runner).__name__!r} does not implement the"
                " `duplicate` method and will fail if used for concurrent execution of"
                " the same flow."
            )

        logger.debug(
            f"Starting {type(flow.task_runner).__name__!r}; submitted tasks "
            f"will be run {CONCURRENCY_MESSAGES[flow.task_runner.concurrency_type]}..."
        )

        flow_run_context.task_runner = await stack.enter_async_context(
            task_runner.start()
        )

        flow_run_context.result_factory = await ResultFactory.from_flow(
            flow, client=client
        )

        if log_prints:
            stack.enter_context(patch_print())

        terminal_or_paused_state = await orchestrate_flow_run(
            flow,
            flow_run=flow_run,
            parameters=parameters,
            wait_for=None,
            client=client,
            partial_flow_run_context=flow_run_context,
            # Orchestration needs to be interruptible if it has a timeout
            interruptible=flow.timeout_seconds is not None,
            user_thread=user_thread,
        )

    if terminal_or_paused_state.is_paused():
        timeout = terminal_or_paused_state.state_details.pause_timeout
        msg = "Currently paused and suspending execution."
        if timeout:
            msg += f" Resume before {timeout.to_rfc3339_string()} to finish execution."
        logger.log(level=logging.INFO, msg=msg)
        await APILogHandler.aflush()

        return terminal_or_paused_state
    else:
        terminal_state = terminal_or_paused_state

    # If debugging, use the more complete `repr` than the usual `str` description
    display_state = repr(terminal_state) if PREFECT_DEBUG_MODE else str(terminal_state)

    logger.log(
        level=logging.INFO if terminal_state.is_completed() else logging.ERROR,
        msg=f"Finished in state {display_state}",
    )

    # When a "root" flow run finishes, flush logs so we do not have to rely on handling
    # during interpreter shutdown
    await APILogHandler.aflush()

    return terminal_state

begin_task_map async

Async entrypoint for task mapping

Source code in prefect/engine.py
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
async def begin_task_map(
    task: Task,
    flow_run_context: Optional[FlowRunContext],
    parameters: Dict[str, Any],
    wait_for: Optional[Iterable[PrefectFuture]],
    return_type: EngineReturnType,
    task_runner: Optional[BaseTaskRunner],
    autonomous: bool = False,
) -> List[Union[PrefectFuture, Awaitable[PrefectFuture], TaskRun]]:
    """Async entrypoint for task mapping"""
    # We need to resolve some futures to map over their data, collect the upstream
    # links beforehand to retain relationship tracking.
    task_inputs = {
        k: await collect_task_run_inputs(v, max_depth=0) for k, v in parameters.items()
    }

    # Resolve the top-level parameters in order to get mappable data of a known length.
    # Nested parameters will be resolved in each mapped child where their relationships
    # will also be tracked.
    parameters = await resolve_inputs(parameters, max_depth=1)

    # Ensure that any parameters in kwargs are expanded before this check
    parameters = explode_variadic_parameter(task.fn, parameters)

    iterable_parameters = {}
    static_parameters = {}
    annotated_parameters = {}
    for key, val in parameters.items():
        if isinstance(val, (allow_failure, quote)):
            # Unwrap annotated parameters to determine if they are iterable
            annotated_parameters[key] = val
            val = val.unwrap()

        if isinstance(val, unmapped):
            static_parameters[key] = val.value
        elif isiterable(val):
            iterable_parameters[key] = list(val)
        else:
            static_parameters[key] = val

    if not len(iterable_parameters):
        raise MappingMissingIterable(
            "No iterable parameters were received. Parameters for map must "
            f"include at least one iterable. Parameters: {parameters}"
        )

    iterable_parameter_lengths = {
        key: len(val) for key, val in iterable_parameters.items()
    }
    lengths = set(iterable_parameter_lengths.values())
    if len(lengths) > 1:
        raise MappingLengthMismatch(
            "Received iterable parameters with different lengths. Parameters for map"
            f" must all be the same length. Got lengths: {iterable_parameter_lengths}"
        )

    map_length = list(lengths)[0]

    task_runs = []
    for i in range(map_length):
        call_parameters = {key: value[i] for key, value in iterable_parameters.items()}
        call_parameters.update({key: value for key, value in static_parameters.items()})

        # Add default values for parameters; these are skipped earlier since they should
        # not be mapped over
        for key, value in get_parameter_defaults(task.fn).items():
            call_parameters.setdefault(key, value)

        # Re-apply annotations to each key again
        for key, annotation in annotated_parameters.items():
            call_parameters[key] = annotation.rewrap(call_parameters[key])

        # Collapse any previously exploded kwargs
        call_parameters = collapse_variadic_parameters(task.fn, call_parameters)

        if autonomous:
            task_runs.append(
                await create_autonomous_task_run(
                    task=task,
                    parameters=call_parameters,
                )
            )
        else:
            task_runs.append(
                partial(
                    get_task_call_return_value,
                    task=task,
                    flow_run_context=flow_run_context,
                    parameters=call_parameters,
                    wait_for=wait_for,
                    return_type=return_type,
                    task_runner=task_runner,
                    extra_task_inputs=task_inputs,
                )
            )

    if autonomous:
        return task_runs

    # Maintain the order of the task runs when using the sequential task runner
    runner = task_runner if task_runner else flow_run_context.task_runner
    if runner.concurrency_type == TaskConcurrencyType.SEQUENTIAL:
        return [await task_run() for task_run in task_runs]

    return await gather(*task_runs)

begin_task_run async

Entrypoint for task run execution.

This function is intended for submission to the task runner.

This method may be called from a worker so we ensure the settings context has been entered. For example, with a runner that is executing tasks in the same event loop, we will likely not enter the context again because the current context already matches:

main thread: --> Flow called with settings A --> begin_task_run executes same event loop --> Profile A matches and is not entered again

However, with execution on a remote environment, we are going to need to ensure the settings for the task run are respected by entering the context:

main thread: --> Flow called with settings A --> begin_task_run is scheduled on a remote worker, settings A is serialized remote worker: --> Remote worker imports Prefect (may not occur) --> Global settings is loaded with default settings --> begin_task_run executes on a different event loop than the flow --> Current settings is not set or does not match, settings A is entered

Source code in prefect/engine.py
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
async def begin_task_run(
    task: Task,
    task_run: TaskRun,
    parameters: Dict[str, Any],
    wait_for: Optional[Iterable[PrefectFuture]],
    result_factory: ResultFactory,
    log_prints: bool,
    settings: prefect.context.SettingsContext,
):
    """
    Entrypoint for task run execution.

    This function is intended for submission to the task runner.

    This method may be called from a worker so we ensure the settings context has been
    entered. For example, with a runner that is executing tasks in the same event loop,
    we will likely not enter the context again because the current context already
    matches:

    main thread:
    --> Flow called with settings A
    --> `begin_task_run` executes same event loop
    --> Profile A matches and is not entered again

    However, with execution on a remote environment, we are going to need to ensure the
    settings for the task run are respected by entering the context:

    main thread:
    --> Flow called with settings A
    --> `begin_task_run` is scheduled on a remote worker, settings A is serialized
    remote worker:
    --> Remote worker imports Prefect (may not occur)
    --> Global settings is loaded with default settings
    --> `begin_task_run` executes on a different event loop than the flow
    --> Current settings is not set or does not match, settings A is entered
    """
    maybe_flow_run_context = prefect.context.FlowRunContext.get()

    async with AsyncExitStack() as stack:
        # The settings context may be null on a remote worker so we use the safe `.get`
        # method and compare it to the settings required for this task run
        if prefect.context.SettingsContext.get() != settings:
            stack.enter_context(settings)
            setup_logging()

        if maybe_flow_run_context:
            # Accessible if on a worker that is running in the same thread as the flow
            client = maybe_flow_run_context.client
            # Only run the task in an interruptible thread if it in the same thread as
            # the flow _and_ the flow run has a timeout attached. If the task is on a
            # worker, the flow run timeout will not be raised in the worker process.
            interruptible = maybe_flow_run_context.timeout_scope is not None
        else:
            # Otherwise, retrieve a new clien`t
            client = await stack.enter_async_context(get_client())
            interruptible = False
            await stack.enter_async_context(anyio.create_task_group())

        await stack.enter_async_context(report_task_run_crashes(task_run, client))

        # TODO: Use the background tasks group to manage logging for this task

        if log_prints:
            stack.enter_context(patch_print())

        await check_api_reachable(
            client, f"Cannot orchestrate task run '{task_run.id}'"
        )
        try:
            state = await orchestrate_task_run(
                task=task,
                task_run=task_run,
                parameters=parameters,
                wait_for=wait_for,
                result_factory=result_factory,
                log_prints=log_prints,
                interruptible=interruptible,
                client=client,
            )

            if not maybe_flow_run_context:
                # When a a task run finishes on a remote worker flush logs to prevent
                # loss if the process exits
                await APILogHandler.aflush()

        except Abort as abort:
            # Task run probably already completed, fetch its state
            task_run = await client.read_task_run(task_run.id)

            if task_run.state.is_final():
                task_run_logger(task_run).info(
                    f"Task run '{task_run.id}' already finished."
                )
            else:
                # TODO: This is a concerning case; we should determine when this occurs
                #       1. This can occur when the flow run is not in a running state
                task_run_logger(task_run).warning(
                    f"Task run '{task_run.id}' received abort during orchestration: "
                    f"{abort} Task run is in {task_run.state.type.value} state."
                )
            state = task_run.state

        except Pause:
            # A pause signal here should mean the flow run suspended, so we
            # should do the same. We'll look up the flow run's pause state to
            # try and reuse it, so we capture any data like timeouts.
            flow_run = await client.read_flow_run(task_run.flow_run_id)
            if flow_run.state and flow_run.state.is_paused():
                state = flow_run.state
            else:
                state = Suspended()

            task_run_logger(task_run).info(
                "Task run encountered a pause signal during orchestration."
            )

        return state

create_and_begin_subflow_run async

Async entrypoint for flows calls within a flow run

Subflows differ from parent flows in that they - Resolve futures in passed parameters into values - Create a dummy task for representation in the parent flow - Retrieve default result storage from the parent flow rather than the server

Returns:

Type Description
Any

The final state of the run

Source code in prefect/engine.py
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
@inject_client
async def create_and_begin_subflow_run(
    flow: Flow,
    parameters: Dict[str, Any],
    wait_for: Optional[Iterable[PrefectFuture]],
    return_type: EngineReturnType,
    client: PrefectClient,
    user_thread: threading.Thread,
) -> Any:
    """
    Async entrypoint for flows calls within a flow run

    Subflows differ from parent flows in that they
    - Resolve futures in passed parameters into values
    - Create a dummy task for representation in the parent flow
    - Retrieve default result storage from the parent flow rather than the server

    Returns:
        The final state of the run
    """
    parent_flow_run_context = FlowRunContext.get()
    parent_logger = get_run_logger(parent_flow_run_context)
    log_prints = should_log_prints(flow)
    terminal_state = None

    parent_logger.debug(f"Resolving inputs to {flow.name!r}")
    task_inputs = {k: await collect_task_run_inputs(v) for k, v in parameters.items()}

    if wait_for:
        task_inputs["wait_for"] = await collect_task_run_inputs(wait_for)

    rerunning = (
        parent_flow_run_context.flow_run.run_count > 1
        if getattr(parent_flow_run_context, "flow_run", None)
        else False
    )

    # Generate a task in the parent flow run to represent the result of the subflow run
    dummy_task = Task(name=flow.name, fn=flow.fn, version=flow.version)
    parent_task_run = await client.create_task_run(
        task=dummy_task,
        flow_run_id=(
            parent_flow_run_context.flow_run.id
            if getattr(parent_flow_run_context, "flow_run", None)
            else None
        ),
        dynamic_key=_dynamic_key_for_task_run(parent_flow_run_context, dummy_task),
        task_inputs=task_inputs,
        state=Pending(),
    )

    # Resolve any task futures in the input
    parameters = await resolve_inputs(parameters)

    if parent_task_run.state.is_final() and not (
        rerunning and not parent_task_run.state.is_completed()
    ):
        # Retrieve the most recent flow run from the database
        flow_runs = await client.read_flow_runs(
            flow_run_filter=FlowRunFilter(
                parent_task_run_id={"any_": [parent_task_run.id]}
            ),
            sort=FlowRunSort.EXPECTED_START_TIME_ASC,
        )
        flow_run = flow_runs[-1]

        # Set up variables required downstream
        terminal_state = flow_run.state
        logger = flow_run_logger(flow_run, flow)

    else:
        flow_run = await client.create_flow_run(
            flow,
            parameters=flow.serialize_parameters(parameters),
            parent_task_run_id=parent_task_run.id,
            state=parent_task_run.state if not rerunning else Pending(),
            tags=TagsContext.get().current_tags,
        )

        parent_logger.info(
            f"Created subflow run {flow_run.name!r} for flow {flow.name!r}"
        )

        logger = flow_run_logger(flow_run, flow)
        ui_url = PREFECT_UI_URL.value()
        if ui_url:
            logger.info(
                f"View at {ui_url}/flow-runs/flow-run/{flow_run.id}",
                extra={"send_to_api": False},
            )

        result_factory = await ResultFactory.from_flow(
            flow, client=parent_flow_run_context.client
        )

        if flow.should_validate_parameters:
            try:
                parameters = flow.validate_parameters(parameters)
            except Exception:
                message = "Validation of flow parameters failed with error:"
                logger.exception(message)
                terminal_state = await propose_state(
                    client,
                    state=await exception_to_failed_state(
                        message=message, result_factory=result_factory
                    ),
                    flow_run_id=flow_run.id,
                )

        if terminal_state is None or not terminal_state.is_final():
            async with AsyncExitStack() as stack:
                await stack.enter_async_context(
                    report_flow_run_crashes(flow_run=flow_run, client=client, flow=flow)
                )

                task_runner = flow.task_runner.duplicate()
                if task_runner is NotImplemented:
                    # Backwards compatibility; will not support concurrent flow runs
                    task_runner = flow.task_runner
                    logger.warning(
                        f"Task runner {type(task_runner).__name__!r} does not implement"
                        " the `duplicate` method and will fail if used for concurrent"
                        " execution of the same flow."
                    )

                await stack.enter_async_context(task_runner.start())

                if log_prints:
                    stack.enter_context(patch_print())

                terminal_state = await orchestrate_flow_run(
                    flow,
                    flow_run=flow_run,
                    parameters=parameters,
                    wait_for=wait_for,
                    # If the parent flow run has a timeout, then this one needs to be
                    # interruptible as well
                    interruptible=parent_flow_run_context.timeout_scope is not None,
                    client=client,
                    partial_flow_run_context=FlowRunContext.construct(
                        sync_portal=parent_flow_run_context.sync_portal,
                        task_runner=task_runner,
                        background_tasks=parent_flow_run_context.background_tasks,
                        result_factory=result_factory,
                        log_prints=log_prints,
                    ),
                    user_thread=user_thread,
                )

    # Display the full state (including the result) if debugging
    display_state = repr(terminal_state) if PREFECT_DEBUG_MODE else str(terminal_state)
    logger.log(
        level=logging.INFO if terminal_state.is_completed() else logging.ERROR,
        msg=f"Finished in state {display_state}",
    )

    # Track the subflow state so the parent flow can use it to determine its final state
    parent_flow_run_context.flow_run_states.append(terminal_state)

    if return_type == "state":
        return terminal_state
    elif return_type == "result":
        return await terminal_state.result(fetch=True)
    else:
        raise ValueError(f"Invalid return type for flow engine {return_type!r}.")

create_autonomous_task_run async

Create a task run in the API for an autonomous task submission and store the provided parameters using the existing result storage mechanism.

Source code in prefect/engine.py
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
async def create_autonomous_task_run(task: Task, parameters: Dict[str, Any]) -> TaskRun:
    """Create a task run in the API for an autonomous task submission and store
    the provided parameters using the existing result storage mechanism.
    """
    async with get_client() as client:
        state = Scheduled()
        if parameters:
            parameters_id = uuid4()
            state.state_details.task_parameters_id = parameters_id

            # TODO: Improve use of result storage for parameter storage / reference
            task.persist_result = True

            factory = await ResultFactory.from_autonomous_task(task, client=client)
            await factory.store_parameters(parameters_id, parameters)

        task_run = await client.create_task_run(
            task=task,
            flow_run_id=None,
            dynamic_key=f"{task.task_key}-{str(uuid4())[:NUM_CHARS_DYNAMIC_KEY]}",
            state=state,
        )

        engine_logger.debug(f"Submitted run of task {task.name!r} for execution")

    return task_run

create_then_begin_flow_run async

Async entrypoint for flow calls

Creates the flow run in the backend, then enters the main flow run engine.

Source code in prefect/engine.py
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
@inject_client
async def create_then_begin_flow_run(
    flow: Flow,
    parameters: Dict[str, Any],
    wait_for: Optional[Iterable[PrefectFuture]],
    return_type: EngineReturnType,
    client: PrefectClient,
    user_thread: threading.Thread,
) -> Any:
    """
    Async entrypoint for flow calls

    Creates the flow run in the backend, then enters the main flow run engine.
    """
    # TODO: Returns a `State` depending on `return_type` and we can add an overload to
    #       the function signature to clarify this eventually.

    await check_api_reachable(client, "Cannot create flow run")

    flow_run = None
    state = Pending()
    if flow.should_validate_parameters:
        try:
            parameters = flow.validate_parameters(parameters)
        except Exception:
            state = await exception_to_failed_state(
                message="Validation of flow parameters failed with error:"
            )
            flow_run = await _make_flow_run(flow, parameters, state, client)
            await _run_flow_hooks(flow, flow_run, state)

    flow_run = flow_run or await _make_flow_run(flow, parameters, state, client)

    engine_logger.info(f"Created flow run {flow_run.name!r} for flow {flow.name!r}")

    logger = flow_run_logger(flow_run, flow)

    ui_url = PREFECT_UI_URL.value()
    if ui_url:
        logger.info(
            f"View at {ui_url}/flow-runs/flow-run/{flow_run.id}",
            extra={"send_to_api": False},
        )

    if state.is_failed():
        logger.error(state.message)
        engine_logger.info(
            f"Flow run {flow_run.name!r} received invalid parameters and is marked as"
            " failed."
        )
    else:
        state = await begin_flow_run(
            flow=flow,
            flow_run=flow_run,
            parameters=parameters,
            client=client,
            user_thread=user_thread,
        )

    if return_type == "state":
        return state
    elif return_type == "result":
        return await state.result(fetch=True)
    else:
        raise ValueError(f"Invalid return type for flow engine {return_type!r}.")

enter_flow_run_engine_from_flow_call

Sync entrypoint for flow calls.

This function does the heavy lifting of ensuring we can get into an async context for flow run execution with minimal overhead.

Source code in prefect/engine.py
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
def enter_flow_run_engine_from_flow_call(
    flow: Flow,
    parameters: Dict[str, Any],
    wait_for: Optional[Iterable[PrefectFuture]],
    return_type: EngineReturnType,
) -> Union[State, Awaitable[State]]:
    """
    Sync entrypoint for flow calls.

    This function does the heavy lifting of ensuring we can get into an async context
    for flow run execution with minimal overhead.
    """
    setup_logging()

    registry = PrefectObjectRegistry.get()
    if registry and registry.block_code_execution:
        engine_logger.warning(
            f"Script loading is in progress, flow {flow.name!r} will not be executed."
            " Consider updating the script to only call the flow if executed"
            f' directly:\n\n\tif __name__ == "__main__":\n\t\t{flow.fn.__name__}()'
        )
        return None

    parent_flow_run_context = FlowRunContext.get()
    is_subflow_run = parent_flow_run_context is not None

    if wait_for is not None and not is_subflow_run:
        raise ValueError("Only flows run as subflows can wait for dependencies.")

    begin_run = create_call(
        create_and_begin_subflow_run if is_subflow_run else create_then_begin_flow_run,
        flow=flow,
        parameters=parameters,
        wait_for=wait_for,
        return_type=return_type,
        client=parent_flow_run_context.client if is_subflow_run else None,
        user_thread=threading.current_thread(),
    )

    # On completion of root flows, wait for the global thread to ensure that
    # any work there is complete
    done_callbacks = (
        [create_call(wait_for_global_loop_exit)] if not is_subflow_run else None
    )

    # WARNING: You must define any context managers here to pass to our concurrency
    # api instead of entering them in here in the engine entrypoint. Otherwise, async
    # flows will not use the context as this function _exits_ to return an awaitable to
    # the user. Generally, you should enter contexts _within_ the async `begin_run`
    # instead but if you need to enter a context from the main thread you'll need to do
    # it here.
    contexts = [capture_sigterm(), collapse_excgroups()]

    if flow.isasync and (
        not is_subflow_run or (is_subflow_run and parent_flow_run_context.flow.isasync)
    ):
        # return a coro for the user to await if the flow is async
        # unless it is an async subflow called in a sync flow
        retval = from_async.wait_for_call_in_loop_thread(
            begin_run,
            done_callbacks=done_callbacks,
            contexts=contexts,
        )

    else:
        retval = from_sync.wait_for_call_in_loop_thread(
            begin_run,
            done_callbacks=done_callbacks,
            contexts=contexts,
        )

    return retval

enter_flow_run_engine_from_subprocess

Sync entrypoint for flow runs that have been submitted for execution by an agent

Differs from enter_flow_run_engine_from_flow_call in that we have a flow run id but not a flow object. The flow must be retrieved before execution can begin. Additionally, this assumes that the caller is always in a context without an event loop as this should be called from a fresh process.

Source code in prefect/engine.py
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
def enter_flow_run_engine_from_subprocess(flow_run_id: UUID) -> State:
    """
    Sync entrypoint for flow runs that have been submitted for execution by an agent

    Differs from `enter_flow_run_engine_from_flow_call` in that we have a flow run id
    but not a flow object. The flow must be retrieved before execution can begin.
    Additionally, this assumes that the caller is always in a context without an event
    loop as this should be called from a fresh process.
    """

    # Ensure collections are imported and have the opportunity to register types before
    # loading the user code from the deployment
    prefect.plugins.load_prefect_collections()

    setup_logging()

    state = from_sync.wait_for_call_in_loop_thread(
        create_call(
            retrieve_flow_then_begin_flow_run,
            flow_run_id,
            user_thread=threading.current_thread(),
        ),
        contexts=[capture_sigterm(), collapse_excgroups()],
    )

    APILogHandler.flush()
    return state

enter_task_run_engine

Sync entrypoint for task calls

Source code in prefect/engine.py
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
def enter_task_run_engine(
    task: Task,
    parameters: Dict[str, Any],
    wait_for: Optional[Iterable[PrefectFuture]],
    return_type: EngineReturnType,
    task_runner: Optional[BaseTaskRunner],
    mapped: bool,
    entering_from_task_run: Optional[bool] = False,
) -> Union[PrefectFuture, Awaitable[PrefectFuture], TaskRun]:
    """Sync entrypoint for task calls"""

    flow_run_context = FlowRunContext.get()

    if not flow_run_context:
        if return_type == "future" or mapped:
            raise RuntimeError(
                " If you meant to submit a background task, you need to set"
                " `prefect config set PREFECT_EXPERIMENTAL_ENABLE_TASK_SCHEDULING=true`"
                " and use `your_task.submit()` instead of `your_task()`."
            )
        from prefect.task_engine import submit_autonomous_task_run_to_engine

        return submit_autonomous_task_run_to_engine(
            task=task,
            task_run=None,
            parameters=parameters,
            task_runner=task_runner,
            wait_for=wait_for,
            return_type=return_type,
            client=get_client(),
        )

    if flow_run_context.timeout_scope and flow_run_context.timeout_scope.cancel_called:
        raise TimeoutError("Flow run timed out")

    call_arguments = {
        "task": task,
        "flow_run_context": flow_run_context,
        "parameters": parameters,
        "wait_for": wait_for,
        "return_type": return_type,
        "task_runner": task_runner,
    }

    if not mapped:
        call_arguments["entering_from_task_run"] = entering_from_task_run

    begin_run = create_call(
        begin_task_map if mapped else get_task_call_return_value, **call_arguments
    )

    if task.isasync and (
        flow_run_context.flow is None or flow_run_context.flow.isasync
    ):
        # return a coro for the user to await if an async task in an async flow
        return from_async.wait_for_call_in_loop_thread(begin_run)
    else:
        return from_sync.wait_for_call_in_loop_thread(begin_run)

orchestrate_flow_run async

Executes a flow run.

Note on flow timeouts

Since async flows are run directly in the main event loop, timeout behavior will match that described by anyio. If the flow is awaiting something, it will immediately return; otherwise, the next time it awaits it will exit. Sync flows are being task runner in a worker thread, which cannot be interrupted. The worker thread will exit at the next task call. The worker thread also has access to the status of the cancellation scope at FlowRunContext.timeout_scope.cancel_called which allows it to raise a TimeoutError to respect the timeout.

Returns:

Type Description
State

The final state of the run

Source code in prefect/engine.py
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
async def orchestrate_flow_run(
    flow: Flow,
    flow_run: FlowRun,
    parameters: Dict[str, Any],
    wait_for: Optional[Iterable[PrefectFuture]],
    interruptible: bool,
    client: PrefectClient,
    partial_flow_run_context: FlowRunContext,
    user_thread: threading.Thread,
) -> State:
    """
    Executes a flow run.

    Note on flow timeouts:
        Since async flows are run directly in the main event loop, timeout behavior will
        match that described by anyio. If the flow is awaiting something, it will
        immediately return; otherwise, the next time it awaits it will exit. Sync flows
        are being task runner in a worker thread, which cannot be interrupted. The worker
        thread will exit at the next task call. The worker thread also has access to the
        status of the cancellation scope at `FlowRunContext.timeout_scope.cancel_called`
        which allows it to raise a `TimeoutError` to respect the timeout.

    Returns:
        The final state of the run
    """

    logger = flow_run_logger(flow_run, flow)

    flow_run_context = None
    parent_flow_run_context = FlowRunContext.get()

    try:
        # Resolve futures in any non-data dependencies to ensure they are ready
        if wait_for is not None:
            await resolve_inputs({"wait_for": wait_for}, return_data=False)
    except UpstreamTaskError as upstream_exc:
        return await propose_state(
            client,
            Pending(name="NotReady", message=str(upstream_exc)),
            flow_run_id=flow_run.id,
            # if orchestrating a run already in a pending state, force orchestration to
            # update the state name
            force=flow_run.state.is_pending(),
        )

    state = await propose_state(client, Running(), flow_run_id=flow_run.id)

    # flag to ensure we only update the flow run name once
    run_name_set = False

    await _run_flow_hooks(flow=flow, flow_run=flow_run, state=state)

    while state.is_running():
        waited_for_task_runs = False

        # Update the flow run to the latest data
        flow_run = await client.read_flow_run(flow_run.id)
        try:
            with FlowRunContext(
                **{
                    **partial_flow_run_context.dict(),
                    **{
                        "flow_run": flow_run,
                        "flow": flow,
                        "client": client,
                        "parameters": parameters,
                    },
                }
            ) as flow_run_context:
                # update flow run name
                if not run_name_set and flow.flow_run_name:
                    flow_run_name = _resolve_custom_flow_run_name(
                        flow=flow, parameters=parameters
                    )

                    await client.update_flow_run(
                        flow_run_id=flow_run.id, name=flow_run_name
                    )
                    logger.extra["flow_run_name"] = flow_run_name
                    logger.debug(
                        f"Renamed flow run {flow_run.name!r} to {flow_run_name!r}"
                    )
                    flow_run.name = flow_run_name
                    run_name_set = True

                args, kwargs = parameters_to_args_kwargs(flow.fn, parameters)
                logger.debug(
                    f"Executing flow {flow.name!r} for flow run {flow_run.name!r}..."
                )

                if PREFECT_DEBUG_MODE:
                    logger.debug(f"Executing {call_repr(flow.fn, *args, **kwargs)}")
                else:
                    logger.debug(
                        "Beginning execution...", extra={"state_message": True}
                    )

                flow_call = create_call(flow.fn, *args, **kwargs)

                # This check for a parent call is needed for cases where the engine
                # was entered directly during testing
                parent_call = get_current_call()

                if parent_call and (
                    not parent_flow_run_context
                    or (
                        getattr(parent_flow_run_context, "flow", None)
                        and parent_flow_run_context.flow.isasync == flow.isasync
                    )
                ):
                    from_async.call_soon_in_waiting_thread(
                        flow_call,
                        thread=user_thread,
                        timeout=flow.timeout_seconds,
                    )
                else:
                    from_async.call_soon_in_new_thread(
                        flow_call, timeout=flow.timeout_seconds
                    )

                result = await flow_call.aresult()

                waited_for_task_runs = await wait_for_task_runs_and_report_crashes(
                    flow_run_context.task_run_futures, client=client
                )
        except PausedRun as exc:
            # could get raised either via utility or by returning Paused from a task run
            # if a task run pauses, we set its state as the flow's state
            # to preserve reschedule and timeout behavior
            paused_flow_run = await client.read_flow_run(flow_run.id)
            if paused_flow_run.state.is_running():
                state = await propose_state(
                    client,
                    state=exc.state,
                    flow_run_id=flow_run.id,
                )

                return state
            paused_flow_run_state = paused_flow_run.state
            return paused_flow_run_state
        except CancelledError as exc:
            if not flow_call.timedout():
                # If the flow call was not cancelled by us; this is a crash
                raise
            # Construct a new exception as `TimeoutError`
            original = exc
            exc = TimeoutError()
            exc.__cause__ = original
            logger.exception("Encountered exception during execution:")
            terminal_state = await exception_to_failed_state(
                exc,
                message=f"Flow run exceeded timeout of {flow.timeout_seconds} seconds",
                result_factory=flow_run_context.result_factory,
                name="TimedOut",
            )
        except Exception:
            # Generic exception in user code
            logger.exception("Encountered exception during execution:")
            terminal_state = await exception_to_failed_state(
                message="Flow run encountered an exception.",
                result_factory=flow_run_context.result_factory,
            )
        else:
            if result is None:
                # All tasks and subflows are reference tasks if there is no return value
                # If there are no tasks, use `None` instead of an empty iterable
                result = (
                    flow_run_context.task_run_futures
                    + flow_run_context.task_run_states
                    + flow_run_context.flow_run_states
                ) or None

            terminal_state = await return_value_to_state(
                await resolve_futures_to_states(result),
                result_factory=flow_run_context.result_factory,
            )

        if not waited_for_task_runs:
            # An exception occurred that prevented us from waiting for task runs to
            # complete. Ensure that we wait for them before proposing a final state
            # for the flow run.
            await wait_for_task_runs_and_report_crashes(
                flow_run_context.task_run_futures, client=client
            )

        # Before setting the flow run state, store state.data using
        # block storage and send the resulting data document to the Prefect API instead.
        # This prevents the pickled return value of flow runs
        # from being sent to the Prefect API and stored in the Prefect database.
        # state.data is left as is, otherwise we would have to load
        # the data from block storage again after storing.
        state = await propose_state(
            client,
            state=terminal_state,
            flow_run_id=flow_run.id,
        )

        await _run_flow_hooks(flow=flow, flow_run=flow_run, state=state)

        if state.type != terminal_state.type and PREFECT_DEBUG_MODE:
            logger.debug(
                (
                    f"Received new state {state} when proposing final state"
                    f" {terminal_state}"
                ),
                extra={"send_to_api": False},
            )

        if not state.is_final() and not state.is_paused():
            logger.info(
                (
                    f"Received non-final state {state.name!r} when proposing final"
                    f" state {terminal_state.name!r} and will attempt to run again..."
                ),
            )
            # Attempt to enter a running state again
            state = await propose_state(client, Running(), flow_run_id=flow_run.id)

    return state

orchestrate_task_run async

Execute a task run

This function should be submitted to a task runner. We must construct the context here instead of receiving it already populated since we may be in a new environment.

Proposes a RUNNING state, then - if accepted, the task user function will be run - if rejected, the received state will be returned

When the user function is run, the result will be used to determine a final state - if an exception is encountered, it is trapped and stored in a FAILED state - otherwise, return_value_to_state is used to determine the state

If the final state is COMPLETED, we generate a cache key as specified by the task

The final state is then proposed - if accepted, this is the final state and will be returned - if rejected and a new final state is provided, it will be returned - if rejected and a non-final state is provided, we will attempt to enter a RUNNING state again

Returns:

Type Description
State

The final state of the run

Source code in prefect/engine.py
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
async def orchestrate_task_run(
    task: Task,
    task_run: TaskRun,
    parameters: Dict[str, Any],
    wait_for: Optional[Iterable[PrefectFuture]],
    result_factory: ResultFactory,
    log_prints: bool,
    interruptible: bool,
    client: PrefectClient,
) -> State:
    """
    Execute a task run

    This function should be submitted to a task runner. We must construct the context
    here instead of receiving it already populated since we may be in a new environment.

    Proposes a RUNNING state, then
    - if accepted, the task user function will be run
    - if rejected, the received state will be returned

    When the user function is run, the result will be used to determine a final state
    - if an exception is encountered, it is trapped and stored in a FAILED state
    - otherwise, `return_value_to_state` is used to determine the state

    If the final state is COMPLETED, we generate a cache key as specified by the task

    The final state is then proposed
    - if accepted, this is the final state and will be returned
    - if rejected and a new final state is provided, it will be returned
    - if rejected and a non-final state is provided, we will attempt to enter a RUNNING
        state again

    Returns:
        The final state of the run
    """
    flow_run_context = prefect.context.FlowRunContext.get()
    if flow_run_context:
        flow_run = flow_run_context.flow_run
    else:
        flow_run = await client.read_flow_run(task_run.flow_run_id)
    logger = task_run_logger(task_run, task=task, flow_run=flow_run)

    partial_task_run_context = TaskRunContext.construct(
        task_run=task_run,
        task=task,
        client=client,
        result_factory=result_factory,
        log_prints=log_prints,
    )
    task_introspection_start_time = time.perf_counter()
    try:
        # Resolve futures in parameters into data
        resolved_parameters = await resolve_inputs(parameters)
        # Resolve futures in any non-data dependencies to ensure they are ready
        await resolve_inputs({"wait_for": wait_for}, return_data=False)
    except UpstreamTaskError as upstream_exc:
        return await propose_state(
            client,
            Pending(name="NotReady", message=str(upstream_exc)),
            task_run_id=task_run.id,
            # if orchestrating a run already in a pending state, force orchestration to
            # update the state name
            force=task_run.state.is_pending(),
        )
    task_introspection_end_time = time.perf_counter()

    introspection_time = round(
        task_introspection_end_time - task_introspection_start_time, 3
    )
    threshold = PREFECT_TASK_INTROSPECTION_WARN_THRESHOLD.value()
    if threshold and introspection_time > threshold:
        logger.warning(
            f"Task parameter introspection took {introspection_time} seconds "
            f", exceeding `PREFECT_TASK_INTROSPECTION_WARN_THRESHOLD` of {threshold}. "
            "Try wrapping large task parameters with "
            "`prefect.utilities.annotations.quote` for increased performance, "
            "e.g. `my_task(quote(param))`. To disable this message set "
            "`PREFECT_TASK_INTROSPECTION_WARN_THRESHOLD=0`."
        )

    # Generate the cache key to attach to proposed states
    # The cache key uses a TaskRunContext that does not include a `timeout_context``

    task_run_context = TaskRunContext(
        **partial_task_run_context.dict(), parameters=resolved_parameters
    )

    cache_key = (
        task.cache_key_fn(
            task_run_context,
            resolved_parameters,
        )
        if task.cache_key_fn
        else None
    )

    # Ignore the cached results for a cache key, default = false
    # Setting on task level overrules the Prefect setting (env var)
    refresh_cache = (
        task.refresh_cache
        if task.refresh_cache is not None
        else PREFECT_TASKS_REFRESH_CACHE.value()
    )

    # Emit an event to capture that the task run was in the `PENDING` state.
    last_event = emit_task_run_state_change_event(
        task_run=task_run, initial_state=None, validated_state=task_run.state
    )
    last_state = (
        Pending()
        if flow_run_context and flow_run_context.autonomous_task_run
        else task_run.state
    )

    # Completed states with persisted results should have result data. If it's missing,
    # this could be a manual state transition, so we should use the Unknown result type
    # to represent that we know we don't know the result.
    if (
        last_state
        and last_state.is_completed()
        and result_factory.persist_result
        and not last_state.data
    ):
        state = await propose_state(
            client,
            state=Completed(data=await UnknownResult.create()),
            task_run_id=task_run.id,
            force=True,
        )

    # Transition from `PENDING` -> `RUNNING`
    try:
        state = await propose_state(
            client,
            Running(
                state_details=StateDetails(
                    cache_key=cache_key, refresh_cache=refresh_cache
                )
            ),
            task_run_id=task_run.id,
        )
    except Pause as exc:
        # We shouldn't get a pause signal without a state, but if this happens,
        # just use a Paused state to assume an in-process pause.
        state = exc.state if exc.state else Paused()

        # If a flow submits tasks and then pauses, we may reach this point due
        # to concurrency timing because the tasks will try to transition after
        # the flow run has paused. Orchestration will send back a Paused state
        # for the task runs.
        if state.state_details.pause_reschedule:
            # If we're being asked to pause and reschedule, we should exit the
            # task and expect to be resumed later.
            raise

    if state.is_paused():
        BACKOFF_MAX = 10  # Seconds
        backoff_count = 0

        async def tick():
            nonlocal backoff_count
            if backoff_count < BACKOFF_MAX:
                backoff_count += 1
            interval = 1 + backoff_count + random.random() * backoff_count
            await anyio.sleep(interval)

        # Enter a loop to wait for the task run to be resumed, i.e.
        # become Pending, and then propose a Running state again.
        while True:
            await tick()

            # Propose a Running state again. We do this instead of reading the
            # task run because if the flow run times out, this lets
            # orchestration fail the task run.
            try:
                state = await propose_state(
                    client,
                    Running(
                        state_details=StateDetails(
                            cache_key=cache_key, refresh_cache=refresh_cache
                        )
                    ),
                    task_run_id=task_run.id,
                )
            except Pause as exc:
                if not exc.state:
                    continue

                if exc.state.state_details.pause_reschedule:
                    # If the pause state includes pause_reschedule, we should exit the
                    # task and expect to be resumed later. We've already checked for this
                    # above, but we check again here in case the state changed; e.g. the
                    # flow run suspended.
                    raise
                else:
                    # Propose a Running state again.
                    continue
            else:
                break

    # Emit an event to capture the result of proposing a `RUNNING` state.
    last_event = emit_task_run_state_change_event(
        task_run=task_run,
        initial_state=last_state,
        validated_state=state,
        follows=last_event,
    )
    last_state = state

    # flag to ensure we only update the task run name once
    run_name_set = False

    run_on_completion_hooks_on_cached = (
        PREFECT_RUN_ON_COMPLETION_HOOKS_ON_CACHED
        and state.is_completed()
        and state.name == "Cached"
    )

    if run_on_completion_hooks_on_cached:
        await _run_task_hooks(
            task=task,
            task_run=task_run,
            state=state,
        )

    # Only run the task if we enter a `RUNNING` state
    while state.is_running():
        # Retrieve the latest metadata for the task run context
        task_run = await client.read_task_run(task_run.id)

        with task_run_context.copy(
            update={"task_run": task_run, "start_time": pendulum.now("UTC")}
        ):
            try:
                args, kwargs = parameters_to_args_kwargs(task.fn, resolved_parameters)
                # update task run name
                if not run_name_set and task.task_run_name:
                    task_run_name = _resolve_custom_task_run_name(
                        task=task, parameters=resolved_parameters
                    )
                    await client.set_task_run_name(
                        task_run_id=task_run.id, name=task_run_name
                    )
                    logger.extra["task_run_name"] = task_run_name
                    logger.debug(
                        f"Renamed task run {task_run.name!r} to {task_run_name!r}"
                    )
                    task_run.name = task_run_name
                    run_name_set = True

                if PREFECT_DEBUG_MODE.value():
                    logger.debug(f"Executing {call_repr(task.fn, *args, **kwargs)}")
                else:
                    logger.debug(
                        "Beginning execution...", extra={"state_message": True}
                    )

                call = from_async.call_soon_in_new_thread(
                    create_call(task.fn, *args, **kwargs), timeout=task.timeout_seconds
                )
                result = await call.aresult()

            except (CancelledError, asyncio.CancelledError) as exc:
                if not call.timedout():
                    # If the task call was not cancelled by us; this is a crash
                    raise
                # Construct a new exception as `TimeoutError`
                original = exc
                exc = TimeoutError()
                exc.__cause__ = original
                logger.exception("Encountered exception during execution:")
                terminal_state = await exception_to_failed_state(
                    exc,
                    message=(
                        f"Task run exceeded timeout of {task.timeout_seconds} seconds"
                    ),
                    result_factory=task_run_context.result_factory,
                    name="TimedOut",
                )
            except Exception as exc:
                logger.exception("Encountered exception during execution:")
                terminal_state = await exception_to_failed_state(
                    exc,
                    message="Task run encountered an exception",
                    result_factory=task_run_context.result_factory,
                )
            else:
                terminal_state = await return_value_to_state(
                    result,
                    result_factory=task_run_context.result_factory,
                )

                # for COMPLETED tasks, add the cache key and expiration
                if terminal_state.is_completed():
                    terminal_state.state_details.cache_expiration = (
                        (pendulum.now("utc") + task.cache_expiration)
                        if task.cache_expiration
                        else None
                    )
                    terminal_state.state_details.cache_key = cache_key

            if terminal_state.is_failed():
                # Defer to user to decide whether failure is retriable
                terminal_state.state_details.retriable = (
                    await _check_task_failure_retriable(task, task_run, terminal_state)
                )
            state = await propose_state(client, terminal_state, task_run_id=task_run.id)
            last_event = emit_task_run_state_change_event(
                task_run=task_run,
                initial_state=last_state,
                validated_state=state,
                follows=last_event,
            )
            last_state = state

            await _run_task_hooks(
                task=task,
                task_run=task_run,
                state=state,
            )

            if state.type != terminal_state.type and PREFECT_DEBUG_MODE:
                logger.debug(
                    (
                        f"Received new state {state} when proposing final state"
                        f" {terminal_state}"
                    ),
                    extra={"send_to_api": False},
                )

            if not state.is_final() and not state.is_paused():
                logger.info(
                    (
                        f"Received non-final state {state.name!r} when proposing final"
                        f" state {terminal_state.name!r} and will attempt to run"
                        " again..."
                    ),
                )
                # Attempt to enter a running state again
                state = await propose_state(client, Running(), task_run_id=task_run.id)
                last_event = emit_task_run_state_change_event(
                    task_run=task_run,
                    initial_state=last_state,
                    validated_state=state,
                    follows=last_event,
                )
                last_state = state

    # If debugging, use the more complete `repr` than the usual `str` description
    display_state = repr(state) if PREFECT_DEBUG_MODE else str(state)

    logger.log(
        level=logging.INFO if state.is_completed() else logging.ERROR,
        msg=f"Finished in state {display_state}",
    )
    return state

pause_flow_run async

Pauses the current flow run by blocking execution until resumed.

When called within a flow run, execution will block and no downstream tasks will run until the flow is resumed. Task runs that have already started will continue running. A timeout parameter can be passed that will fail the flow run if it has not been resumed within the specified time.

Parameters:

Name Type Description Default
flow_run_id UUID

a flow run id. If supplied, this function will attempt to pause the specified flow run outside of the flow run process. When paused, the flow run will continue execution until the NEXT task is orchestrated, at which point the flow will exit. Any tasks that have already started will run until completion. When resumed, the flow run will be rescheduled to finish execution. In order pause a flow run in this way, the flow needs to have an associated deployment and results need to be configured with the persist_results option.

None
timeout int

the number of seconds to wait for the flow to be resumed before failing. Defaults to 1 hour (3600 seconds). If the pause timeout exceeds any configured flow-level timeout, the flow might fail even after resuming.

3600
poll_interval int

The number of seconds between checking whether the flow has been resumed. Defaults to 10 seconds.

10
reschedule bool

Flag that will reschedule the flow run if resumed. Instead of blocking execution, the flow will gracefully exit (with no result returned) instead. To use this flag, a flow needs to have an associated deployment and results need to be configured with the persist_results option.

False
key str

An optional key to prevent calling pauses more than once. This defaults to the number of pauses observed by the flow so far, and prevents pauses that use the "reschedule" option from running the same pause twice. A custom key can be supplied for custom pausing behavior.

None
wait_for_input Optional[Type[T]]

a subclass of RunInput or any type supported by Pydantic. If provided when the flow pauses, the flow will wait for the input to be provided before resuming. If the flow is resumed without providing the input, the flow will fail. If the flow is resumed with the input, the flow will resume and the input will be loaded and returned from this function.

None
@task
def task_one():
    for i in range(3):
        sleep(1)

@flow
def my_flow():
    terminal_state = task_one.submit(return_state=True)
    if terminal_state.type == StateType.COMPLETED:
        print("Task one succeeded! Pausing flow run..")
        pause_flow_run(timeout=2)
    else:
        print("Task one failed. Skipping pause flow run..")
Source code in prefect/engine.py
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
@sync_compatible
@deprecated_parameter(
    "flow_run_id", start_date="Dec 2023", help="Use `suspend_flow_run` instead."
)
@deprecated_parameter(
    "reschedule",
    start_date="Dec 2023",
    when=lambda p: p is True,
    help="Use `suspend_flow_run` instead.",
)
@experimental_parameter(
    "wait_for_input", group="flow_run_input", when=lambda y: y is not None
)
async def pause_flow_run(
    wait_for_input: Optional[Type[T]] = None,
    flow_run_id: UUID = None,
    timeout: int = 3600,
    poll_interval: int = 10,
    reschedule: bool = False,
    key: str = None,
) -> Optional[T]:
    """
    Pauses the current flow run by blocking execution until resumed.

    When called within a flow run, execution will block and no downstream tasks will
    run until the flow is resumed. Task runs that have already started will continue
    running. A timeout parameter can be passed that will fail the flow run if it has not
    been resumed within the specified time.

    Args:
        flow_run_id: a flow run id. If supplied, this function will attempt to pause
            the specified flow run outside of the flow run process. When paused, the
            flow run will continue execution until the NEXT task is orchestrated, at
            which point the flow will exit. Any tasks that have already started will
            run until completion. When resumed, the flow run will be rescheduled to
            finish execution. In order pause a flow run in this way, the flow needs to
            have an associated deployment and results need to be configured with the
            `persist_results` option.
        timeout: the number of seconds to wait for the flow to be resumed before
            failing. Defaults to 1 hour (3600 seconds). If the pause timeout exceeds
            any configured flow-level timeout, the flow might fail even after resuming.
        poll_interval: The number of seconds between checking whether the flow has been
            resumed. Defaults to 10 seconds.
        reschedule: Flag that will reschedule the flow run if resumed. Instead of
            blocking execution, the flow will gracefully exit (with no result returned)
            instead. To use this flag, a flow needs to have an associated deployment and
            results need to be configured with the `persist_results` option.
        key: An optional key to prevent calling pauses more than once. This defaults to
            the number of pauses observed by the flow so far, and prevents pauses that
            use the "reschedule" option from running the same pause twice. A custom key
            can be supplied for custom pausing behavior.
        wait_for_input: a subclass of `RunInput` or any type supported by
            Pydantic. If provided when the flow pauses, the flow will wait for the
            input to be provided before resuming. If the flow is resumed without
            providing the input, the flow will fail. If the flow is resumed with the
            input, the flow will resume and the input will be loaded and returned
            from this function.

    Example:
    ```python
    @task
    def task_one():
        for i in range(3):
            sleep(1)

    @flow
    def my_flow():
        terminal_state = task_one.submit(return_state=True)
        if terminal_state.type == StateType.COMPLETED:
            print("Task one succeeded! Pausing flow run..")
            pause_flow_run(timeout=2)
        else:
            print("Task one failed. Skipping pause flow run..")
    ```

    """
    if flow_run_id:
        if wait_for_input is not None:
            raise RuntimeError("Cannot wait for input when pausing out of process.")

        return await _out_of_process_pause(
            flow_run_id=flow_run_id,
            timeout=timeout,
            reschedule=reschedule,
            key=key,
        )
    else:
        return await _in_process_pause(
            timeout=timeout,
            poll_interval=poll_interval,
            reschedule=reschedule,
            key=key,
            wait_for_input=wait_for_input,
        )

report_flow_run_crashes async

Detect flow run crashes during this context and update the run to a proper final state.

This context must reraise the exception to properly exit the run.

Source code in prefect/engine.py
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
@asynccontextmanager
async def report_flow_run_crashes(flow_run: FlowRun, client: PrefectClient, flow: Flow):
    """
    Detect flow run crashes during this context and update the run to a proper final
    state.

    This context _must_ reraise the exception to properly exit the run.
    """
    try:
        with collapse_excgroups():
            yield
    except (Abort, Pause):
        # Do not capture internal signals as crashes
        raise
    except BaseException as exc:
        state = await exception_to_crashed_state(exc)
        logger = flow_run_logger(flow_run)
        with anyio.CancelScope(shield=True):
            logger.error(f"Crash detected! {state.message}")
            logger.debug("Crash details:", exc_info=exc)
            flow_run_state = await propose_state(client, state, flow_run_id=flow_run.id)
            engine_logger.debug(
                f"Reported crashed flow run {flow_run.name!r} successfully!"
            )

            # Only `on_crashed` and `on_cancellation` flow run state change hooks can be called here.
            # We call the hooks after the state change proposal to `CRASHED` is validated
            # or rejected (if it is in a `CANCELLING` state).
            await _run_flow_hooks(
                flow=flow,
                flow_run=flow_run,
                state=flow_run_state,
            )

        # Reraise the exception
        raise

report_task_run_crashes async

Detect task run crashes during this context and update the run to a proper final state.

This context must reraise the exception to properly exit the run.

Source code in prefect/engine.py
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
@asynccontextmanager
async def report_task_run_crashes(task_run: TaskRun, client: PrefectClient):
    """
    Detect task run crashes during this context and update the run to a proper final
    state.

    This context _must_ reraise the exception to properly exit the run.
    """
    try:
        with collapse_excgroups():
            yield
    except (Abort, Pause):
        # Do not capture internal signals as crashes
        raise
    except BaseException as exc:
        state = await exception_to_crashed_state(exc)
        logger = task_run_logger(task_run)
        with anyio.CancelScope(shield=True):
            logger.error(f"Crash detected! {state.message}")
            logger.debug("Crash details:", exc_info=exc)
            await client.set_task_run_state(
                state=state,
                task_run_id=task_run.id,
                force=True,
            )
            engine_logger.debug(
                f"Reported crashed task run {task_run.name!r} successfully!"
            )

        # Reraise the exception
        raise

resume_flow_run async

Resumes a paused flow.

Parameters:

Name Type Description Default
flow_run_id

the flow_run_id to resume

required
run_input Optional[Dict]

a dictionary of inputs to provide to the flow run.

None
Source code in prefect/engine.py
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
@deprecated_callable(
    start_date="Jun 2024",
    help="Will be moved in Prefect 3 to prefect.flow_runs:resume_flow_run",
)
@sync_compatible
async def resume_flow_run(flow_run_id, run_input: Optional[Dict] = None):
    """
    Resumes a paused flow.

    Args:
        flow_run_id: the flow_run_id to resume
        run_input: a dictionary of inputs to provide to the flow run.
    """
    client = get_client()
    async with client:
        flow_run = await client.read_flow_run(flow_run_id)

        if not flow_run.state.is_paused():
            raise NotPausedError("Cannot resume a run that isn't paused!")

        response = await client.resume_flow_run(flow_run_id, run_input=run_input)

    if response.status == SetStateStatus.REJECT:
        if response.state.type == StateType.FAILED:
            raise FlowPauseTimeout("Flow run can no longer be resumed.")
        else:
            raise RuntimeError(f"Cannot resume this run: {response.details.reason}")

retrieve_flow_then_begin_flow_run async

Async entrypoint for flow runs that have been submitted for execution by an agent

  • Retrieves the deployment information
  • Loads the flow object using deployment information
  • Updates the flow run version
Source code in prefect/engine.py
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
@inject_client
async def retrieve_flow_then_begin_flow_run(
    flow_run_id: UUID,
    client: PrefectClient,
    user_thread: threading.Thread,
) -> State:
    """
    Async entrypoint for flow runs that have been submitted for execution by an agent

    - Retrieves the deployment information
    - Loads the flow object using deployment information
    - Updates the flow run version
    """
    flow_run = await client.read_flow_run(flow_run_id)

    entrypoint = os.environ.get("PREFECT__FLOW_ENTRYPOINT")

    try:
        flow = (
            # We do not want to use a placeholder flow at runtime
            load_flow_from_entrypoint(entrypoint, use_placeholder_flow=False)
            if entrypoint
            else await load_flow_from_flow_run(
                flow_run, client=client, use_placeholder_flow=False
            )
        )
    except Exception:
        message = (
            "Flow could not be retrieved from"
            f" {'entrypoint' if entrypoint else 'deployment'}."
        )
        flow_run_logger(flow_run).exception(message)
        state = await exception_to_failed_state(message=message)
        await client.set_flow_run_state(
            state=state, flow_run_id=flow_run_id, force=True
        )
        return state

    # Update the flow run policy defaults to match settings on the flow
    # Note: Mutating the flow run object prevents us from performing another read
    #       operation if these properties are used by the client downstream
    if flow_run.empirical_policy.retry_delay is None:
        flow_run.empirical_policy.retry_delay = flow.retry_delay_seconds

    if flow_run.empirical_policy.retries is None:
        flow_run.empirical_policy.retries = flow.retries

    await client.update_flow_run(
        flow_run_id=flow_run_id,
        flow_version=flow.version,
        empirical_policy=flow_run.empirical_policy,
    )

    if flow.should_validate_parameters:
        failed_state = None
        try:
            parameters = flow.validate_parameters(flow_run.parameters)
        except Exception:
            message = "Validation of flow parameters failed with error: "
            flow_run_logger(flow_run).exception(message)
            failed_state = await exception_to_failed_state(message=message)

        if failed_state is not None:
            await propose_state(
                client,
                state=failed_state,
                flow_run_id=flow_run_id,
            )
            return failed_state
    else:
        parameters = flow_run.parameters

    # Ensure default values are populated
    parameters = {**get_parameter_defaults(flow.fn), **parameters}

    return await begin_flow_run(
        flow=flow,
        flow_run=flow_run,
        parameters=parameters,
        client=client,
        user_thread=user_thread,
    )

suspend_flow_run async

Suspends a flow run by stopping code execution until resumed.

When suspended, the flow run will continue execution until the NEXT task is orchestrated, at which point the flow will exit. Any tasks that have already started will run until completion. When resumed, the flow run will be rescheduled to finish execution. In order suspend a flow run in this way, the flow needs to have an associated deployment and results need to be configured with the persist_results option.

Parameters:

Name Type Description Default
flow_run_id Optional[UUID]

a flow run id. If supplied, this function will attempt to suspend the specified flow run. If not supplied will attempt to suspend the current flow run.

None
timeout Optional[int]

the number of seconds to wait for the flow to be resumed before failing. Defaults to 1 hour (3600 seconds). If the pause timeout exceeds any configured flow-level timeout, the flow might fail even after resuming.

3600
key Optional[str]

An optional key to prevent calling suspend more than once. This defaults to a random string and prevents suspends from running the same suspend twice. A custom key can be supplied for custom suspending behavior.

None
wait_for_input Optional[Type[T]]

a subclass of RunInput or any type supported by Pydantic. If provided when the flow suspends, the flow will remain suspended until receiving the input before resuming. If the flow is resumed without providing the input, the flow will fail. If the flow is resumed with the input, the flow will resume and the input will be loaded and returned from this function.

None
Source code in prefect/engine.py
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
@sync_compatible
@inject_client
@experimental_parameter(
    "wait_for_input", group="flow_run_input", when=lambda y: y is not None
)
async def suspend_flow_run(
    wait_for_input: Optional[Type[T]] = None,
    flow_run_id: Optional[UUID] = None,
    timeout: Optional[int] = 3600,
    key: Optional[str] = None,
    client: PrefectClient = None,
) -> Optional[T]:
    """
    Suspends a flow run by stopping code execution until resumed.

    When suspended, the flow run will continue execution until the NEXT task is
    orchestrated, at which point the flow will exit. Any tasks that have
    already started will run until completion. When resumed, the flow run will
    be rescheduled to finish execution. In order suspend a flow run in this
    way, the flow needs to have an associated deployment and results need to be
    configured with the `persist_results` option.

    Args:
        flow_run_id: a flow run id. If supplied, this function will attempt to
            suspend the specified flow run. If not supplied will attempt to
            suspend the current flow run.
        timeout: the number of seconds to wait for the flow to be resumed before
            failing. Defaults to 1 hour (3600 seconds). If the pause timeout
            exceeds any configured flow-level timeout, the flow might fail even
            after resuming.
        key: An optional key to prevent calling suspend more than once. This
            defaults to a random string and prevents suspends from running the
            same suspend twice. A custom key can be supplied for custom
            suspending behavior.
        wait_for_input: a subclass of `RunInput` or any type supported by
            Pydantic. If provided when the flow suspends, the flow will remain
            suspended until receiving the input before resuming. If the flow is
            resumed without providing the input, the flow will fail. If the flow is
            resumed with the input, the flow will resume and the input will be
            loaded and returned from this function.
    """
    context = FlowRunContext.get()

    if flow_run_id is None:
        if TaskRunContext.get():
            raise RuntimeError("Cannot suspend task runs.")

        if context is None or context.flow_run is None:
            raise RuntimeError(
                "Flow runs can only be suspended from within a flow run."
            )

        logger = get_run_logger(context=context)
        logger.info(
            "Suspending flow run, execution will be rescheduled when this flow run is"
            " resumed."
        )
        flow_run_id = context.flow_run.id
        suspending_current_flow_run = True
        pause_counter = _observed_flow_pauses(context)
        pause_key = key or str(pause_counter)
    else:
        # Since we're suspending another flow run we need to generate a pause
        # key that won't conflict with whatever suspends/pauses that flow may
        # have. Since this method won't be called during that flow run it's
        # okay that this is non-deterministic.
        suspending_current_flow_run = False
        pause_key = key or str(uuid4())

    proposed_state = Suspended(timeout_seconds=timeout, pause_key=pause_key)

    if wait_for_input:
        wait_for_input = run_input_subclass_from_type(wait_for_input)
        run_input_keyset = keyset_from_paused_state(proposed_state)
        proposed_state.state_details.run_input_keyset = run_input_keyset

    try:
        state = await propose_state(
            client=client,
            state=proposed_state,
            flow_run_id=flow_run_id,
        )
    except Abort as exc:
        # Aborted requests mean the suspension is not allowed
        raise RuntimeError(f"Flow run cannot be suspended: {exc}")

    if state.is_running():
        # The orchestrator rejected the suspended state which means that this
        # suspend has happened before and the flow run has been resumed.
        if wait_for_input:
            # The flow run wanted input, so we need to load it and return it
            # to the user.
            return await wait_for_input.load(run_input_keyset)
        return

    if not state.is_paused():
        # If we receive anything but a PAUSED state, we are unable to continue
        raise RuntimeError(
            f"Flow run cannot be suspended. Received unexpected state from API: {state}"
        )

    if wait_for_input:
        await wait_for_input.save(run_input_keyset)

    if suspending_current_flow_run:
        # Exit this process so the run can be resubmitted later
        raise Pause()