States are rich objects that contain information about the status of a particular task run or flow run.

You can learn many things about a task or flow by examining its current state or the history of its states. For example, you might observe that a given run:

  • is Scheduled to run for the third time in an hour
  • is Late because the associated worker process is not responding
  • was Scheduled to run, but later Cancelled via button in the UI
  • was Cached from a previous run instead of Running again
  • has Completed after Running among its fellow task runs
  • ended up Failed because Claude left pydantic 1.x syntax in your code
  • ended up Crashed because you Ctrl-C’d the process

Only runs have states: Flows and tasks are templates that describe what a system does; only when we run the system does it also take on a state.

​
State types

Prefect states have names and types. A state’s name is often, but not always, synonymous with its type. For example, a task run that is running for the first time has a state with the name Running and the type RUNNING. However, if the task retries, that same task run will have the name Retrying and the type RUNNING.

State types drive orchestration logic, whereas state names provide visual bookkeeping.

The full list of states and state types includes:

NameTypeTerminal?Description
ScheduledSCHEDULEDNoThe run will begin at a particular time in the future.
LateSCHEDULEDNoThe run’s scheduled start time has passed, but it has not transitioned to PENDING (15 seconds by default).
AwaitingRetrySCHEDULEDNoThe run did not complete successfully because of a code issue and had remaining retry attempts.
PendingPENDINGNoThe run has been submitted to execute, but is waiting on necessary preconditions to be satisfied.
RunningRUNNINGNoThe run code is currently executing.
RetryingRUNNINGNoThe run code is currently executing after previously not completing successfully.
PausedPAUSEDNoThe run code has stopped executing until it receives manual approval to proceed.
CancellingCANCELLINGNoThe infrastructure on which the code was running is being cleaned up.
CancelledCANCELLEDYesThe run did not complete because a user determined that it should not.
CompletedCOMPLETEDYesThe run completed successfully.
CachedCOMPLETEDYesThe run result was loaded from a previously cached value.
RolledBackCOMPLETEDYesThe run completed successfully but the transaction rolled back and executed rollback hooks.
FailedFAILEDYesThe run did not complete because of a code issue and had no remaining retry attempts.
CrashedCRASHEDYesThe run did not complete because of an infrastructure issue.

​
State transitions

When a flow run changes states, you can often tell if it is behaving normally or not. Here are some common state transitions and what they mean:

FromToTriggered ByExplanation
✅NoneScheduledManual run, automation, or scheduleA flow run has been created and scheduled for future execution.
✅ScheduledPendingWorkerA worker is attempting to start the infrastructure for your flow run.
⛔ScheduledLateCloud or self-hosted serverTypically means that a worker did not pick up the run because 1) no workers are healthy, 2) workers are not polling the right work pool or work queue, or 3) workers are concurrency limited.
✅PendingRunningWorkerThe infrastructure was provisioned and the flow is running
⛔PendingCrashedWorker or runnerSomething went wrong. Some possibilities are 1) the worker could have failed to create the infrastructure, 2) the code is not present in storage or the worker failed to authenticate, or 3) the code has missing or broken imports, or syntax errors.
✅RunningCompletedFlow runThe flow completed successfully.
⛔RunningFailedFlow runThis usually means that your code raised an exception, check the flow run logs.
⛔RunningCrashedWorker or flow runProbably not a raised exception in your code, but could be an infrastructure issue such as 1) an out of memory error, 2) an evicted pod, or 3) a timeout.

​
Final state determination

The final state of a flow or task run depends on a number of factors; generally speaking there are three categories of terminal states:

  • COMPLETED: a run in any COMPLETED state did not encounter any errors or exceptions and returned successfully
  • FAILED: a run in any FAILED state encountered an error during execution, such as a raised exception
  • CRASHED: a run in any CRASHED state was interrupted by an OS signal such as a KeyboardInterrupt or SIGTERM

The flow of state transitions can be visualized here:

States are represented by their name, with boxes behind states clarifying their underlying type. Dotted lines lead to terminal states.

​
Task return values

A task will be placed into a Completed state if it returns any Python object, with one exception: if a task explicitly returns a Prefect Failed state, the task will be marked Failed.

from prefect import task, flow
from prefect.states import Completed, Failed


@task
def toggle_task(fail: bool):
    if fail:
        return Failed(message="I was instructed to fail.")
    else:
        return Completed(message="I was instructed to succeed.")


@flow
def example():
    # this run will be set to a `Failed` state
    state_one = toggle_task(fail=True)

    # this run will be set to a `Completed` state
    state_two = toggle_task(fail=False)

    # similarly, the flow run will fail because we return a `Failed` state
    return state_one, state_two

You can also access state objects directly within a flow through the return_state flag:

from prefect import flow, task


@task
def add_one(x):
    return x + 1


@flow
def my_flow():
    result = add_one(1)
    assert isinstance(result, int) and result == 2

    state = add_one(1, return_state=True)
    assert state.is_completed() is True
    assert state.result() == 2

Returning a State via return_state=True is useful when you want to conditionally respond to the terminal states of a task or flow. For example, if state.is_failed(): ....

​
Flow return values

The final state of a flow is determined by its return value. The following rules apply:

  • If an exception is raised directly in the flow function, the flow run is marked as FAILED.
  • If a flow returns a manually created state, it is used as the state of the final flow run. This allows for manual determination of final state.
  • If a flow returns an iterable of states, the presence of any FAILED state will cause the run to be marked as FAILED.

In any other situation in which the flow returns without error, it will be marked as COMPLETED.

If you manipulate states programmatically, you can create situations in which tasks within a flow can fail and not cause flow run failure. For example:

from prefect import flow, task 


@task 
def add_one(x):
    return x + 1


@flow 
def my_flow():
    # avoided raising an exception via `return_state=True`
    state = add_one("1", return_state=True)
    assert state.is_failed()

    # the flow function returns successfully!
    return

If state were returned from the flow function, the run would be marked as FAILED.

​
State change hooks

State change hooks execute code in response to client side changes in flow or task run states, enabling you to define actions for specific state transitions in a workflow.

State hooks have the following signature:

from prefect import Task, Flow
from prefect.states import State
from prefect.client.schemas.objects import TaskRun, FlowRun

def my_task_state_hook(task: Task, run: TaskRun, state: State) -> None:
    ...

def my_flow_state_hook(flow: Flow, run: FlowRun, state: State) -> None:
    ...

State change hooks are versatile, allowing you to specify multiple state change hooks for the same state transition, or to use the same state change hook for different transitions.

​
Further reading