Great question. Assuming Prefect 3.x: to “finish the rest of the task” and then trigger a retry, just complete your logic and deliberately raise an exception at the end. Prefect will mark the attempt as failed and, if retries remain, schedule the next attempt using your configured delay/backoff. Quick patterns: - Force a retry after finishing work

from prefect import task, flow, get_run_logger
from prefect.tasks import exponential_backoff

@task(retries=3, retry_delay_seconds=exponential_backoff(2))  # 2s, 4s, 8s
def do_work_then_retry():
    logger = get_run_logger()

    # 1) Main work
    do_side_effects()
    finalize_things()

    # 2) Force a retry (this attempt will be marked Failed; next attempt will run)
    raise RuntimeError("Triggering retry after finishing required steps")

@flow
def run():
    do_work_then_retry()

- Conditional retry at the end

@task(retries=5, retry_delay_seconds=[5, 10, 20, 40, 60])
def maybe_retry():
    result = compute()
    write_artifacts(result)  # logs/artifacts are fine to emit before raising

    if need_another_attempt(result):
        raise Exception("Retry requested")
    return result

- Gate retries with a custom condition (decide which failures should retry)

def should_retry(task, task_run, state) -> bool:
    exc = state.result(raise_on_failure=False)
    # retry only on transient errors you care about
    return isinstance(exc, TimeoutError)

@task(retries=3, retry_condition_fn=should_retry, retry_delay_seconds=5)
def do_stuff():
    # ...
    raise TimeoutError("transient")

Notes: - There’s no special “retry” exception to raise in 3.x; any exception will trigger the retry logic if

retries

> 0 (and your

retry_condition_fn

returns True, if set). - If you need to avoid repeating side-effects on retries, make your work idempotent, checkpoint state externally, or split the task so the side-effecting part is separate from the retried part. If you can share what should trigger the retry and whether you need to persist any outputs before retrying, I can tailor the example (e.g., exact backoff, jitter, or task split).