How to respond to state changes

Available state change hooks

Type	Flow	Task	Description
`on_completion`	✓	✓	Executes when a flow or task run enters a `Completed` state.
`on_failure`	✓	✓	Executes when a flow or task run enters a `Failed` state.
`on_cancellation`	✓	-	Executes when a flow run enters a `Cancelling` state.
`on_crashed`	✓	-	Executes when a flow run enters a `Crashed` state.
`on_running`	✓	✓	Executes when a flow or task run enters a `Running` state.

Note that the on_rollback hook for tasks is not a proper state change hook but instead is a transaction lifecycle hook. Rollback hooks accept one argument representing the transaction for the task.

Send a notification when a workflow run fails

To send a notification when a flow or task run fails, you can specify a on_failure hook.

from prefect import flow
from prefect.blocks.core import Block
from prefect.settings import get_current_settings


@flow(retries=1)
def failing_flow():
    raise ValueError("oops!")


@failing_flow.on_failure
def notify_slack(flow, flow_run, state):
    slack_webhook_block = Block.load(
        "slack-webhook/my-slack-webhook"
    )

    PREFECT_API_URL = get_current_settings().api.url

    slack_webhook_block.notify(
        (
            f"Your job {flow_run.name} entered {state.name} "
            f"with message:\n\n"
            f"See <https://{PREFECT_API_URL}/flow-runs/"
            f"flow-run/{flow_run.id}|the flow run in the UI>\n\n"
            f"Tags: {flow_run.tags}\n\n"
            f"Scheduled start: {flow_run.expected_start_time}"
        )
    )

if __name__ == "__main__":
    failing_flow()

Retries are configured in this example, so the on_failure hook will not run until all retries have completed and the flow run enters a Failed state.

State change hooks execute client sideState change hooks run in the same process as your workflow and execution cannot be guaranteed. For more robust execution of logic in response to state changes, use an Automation.

Execute code when a task starts running

The on_running hook executes when a task enters a Running state, before the task body executes. This is useful for logging, metrics, or setting up runtime state:

from prefect import flow, task
import time

def record_start_time(task, task_run, state):
    print(f"Task {task.name} started at {time.time()}")

@task(on_running=[record_start_time])
def my_task():
    return "hello"

# Or use the decorator pattern
@task
def another_task():
    return "world"

@another_task.on_running
def log_start(task, task_run, state):
    print(f"Starting {task.name}!")

@flow
def my_flow():
    my_task()
    another_task()

Note that on_running hooks execute synchronously before the task body runs. If your hook takes 10 seconds, the task waits 10 seconds before starting. When retries are configured, on_running hooks fire on each retry attempt, including the initial run. For example, a task configured with retries=2 will trigger its on_running hooks up to three times: once on the initial run and once for each retry attempt.

Log from state change hooks

State change hooks run outside the active flow or task run context. This means get_run_logger() will raise a MissingContextError if called inside a hook. To emit logs from a hook that appear in the Prefect UI, use flow_run_logger (or task_run_logger for task hooks) from prefect.logging.loggers. These functions create a logger tied to the run using the parameters your hook already receives:

from prefect import flow
from prefect.logging.loggers import flow_run_logger


def my_hook(flow, flow_run, state):
    logger = flow_run_logger(flow_run, flow)
    logger.info(f"Flow {flow_run.name} entered state {state.name}")


@flow(on_failure=[my_hook])
def failing_flow():
    raise ValueError("oops!")

For more information on logging options, see How to add logging.

Emit events from state change hooks

A common use of hooks is to emit a custom event when a run reaches a state—for example, a failure event that other automations react to. Because the hook runs a fraction of a second after the state change, your custom event and Prefect’s own state change event (like prefect.flow-run.Failed) occur within a second or two of each other, and Prefect does not guarantee the order they arrive at the automation system. If a trigger uses the state change event in after and your custom event in expect, that race can cause missed or spurious firings. To make the order deterministic, emit your event with follows referencing the state change event. The state argument your hook receives carries what you need: state.id is the ID of the state change event, and state.timestamp is when it occurred.

from prefect import flow
from prefect.events import emit_event
from prefect.events.schemas.events import Event


def emit_failure_event(flow, flow_run, state):
    emit_event(
        event="my-app.flow-run.failure-handled",
        resource={"prefect.resource.id": f"prefect.flow-run.{flow_run.id}"},
        follows=Event(
            event="prefect.flow-run.Failed",
            resource={"prefect.resource.id": f"prefect.flow-run.{flow_run.id}"},
            id=state.id,
            occurred=state.timestamp,
        ),
    )


@flow(on_failure=[emit_failure_event])
def failing_flow():
    raise ValueError("oops!")

The Event passed to follows is a stub. emit_event reads only its id and occurred fields—to fill in the follows field of the emitted event and to check that the two events occurred within five minutes of each other. The stub is not emitted: Prefect already emitted the real prefect.flow-run.Failed event server-side, and the stub only identifies it. With follows declared, the automation system always processes your custom event after the state change event it follows.

Pass `kwargs` to state change hooks

You can compose the with_options method to effectively pass arbitrary **kwargs to your hooks:

from functools import partial
from prefect import flow, task

data = {}

def my_hook(task, task_run, state, **kwargs):
    data.update(state=state, **kwargs)

@task
def bad_task():
    raise ValueError("meh")

@flow
def ok_with_failure_flow(x: str = "foo", y: int = 42):
    bad_task_with_a_hook = bad_task.with_options(
        on_failure=[partial(my_hook, **dict(x=x, y=y))]
    )
    # return a tuple of "bar" and the task run state
    # to avoid raising the task's exception
    return "bar", bad_task_with_a_hook(return_state=True)

_, task_run_state = ok_with_failure_flow()

assert data == {"x": "foo", "y": 42, "state": task_run_state}

Workflows

Deployments

Configuration

Automations

Workflow Infrastructure

Prefect Cloud

Prefect Self-hosted

AI

Migrate

How to respond to state changes

Available state change hooks

Send a notification when a workflow run fails

Execute code when a task starts running

Log from state change hooks

Emit events from state change hooks

Pass `kwargs` to state change hooks

​Available state change hooks

​Send a notification when a workflow run fails

​Execute code when a task starts running

​Log from state change hooks

​Emit events from state change hooks

​Pass kwargs to state change hooks

Available state change hooks

Send a notification when a workflow run fails

Execute code when a task starts running

Log from state change hooks

Emit events from state change hooks

Pass `kwargs` to state change hooks