Hello everyone. I am experimenting a bit with the dask taskrunner, and I am running into some issues related to priority. I have the following test-flow:

from typing import List, Tuple

import dask
from prefect import flow, get_run_logger, task
from prefect.context import get_run_context
from prefect_dask import get_dask_client


def is_prime(number: int) -> Tuple[int, bool]:
    if number == 2 or number == 3:
        return number, True
    if number % 2 == 0 or number < 2:
        return number, False

    for i in range(3, int(number ** 0.5) + 1, 2):
        if number % i == 0:
            return number, False

    return number, True


@task
def get_primes_from_split(min_number, max_number) -> List[int]:

    if min_number % 2 == 0:
        min_number += 1

    with get_dask_client() as client:
        futures = [client.submit(is_prime, n) for n in range(min_number, max_number, 2)]

        maybe_primes = [future.result() for future in futures]

    return [value for value, flag in maybe_primes if flag]


@flow(name="example_prime_number_search")
def main(max_number: int = 1_000_000, split_size=10_000):
    log = get_run_logger()
    context = get_run_context()

    <http://log.info|log.info>("Task Runner: %s", context.task_runner.name)
    <http://log.info|log.info>("Searching for primes from up to %d", max_number)

    futures = [get_primes_from_split.submit(x, x + split_size) for x in range(0, max_number + 1, split_size)]
    primes = [value for future in futures for value in future.result()]

    if len(primes) > 10:
        <http://log.info|log.info>("Found %d primes: %s, ...", sorted(primes)[::-1][:10])
    else:
        <http://log.info|log.info>("Found %d primes: %s", sorted(primes)[::-1][:10]

When running this with the

DaskTaskrunner

, the task

get_primes_from_split

is scheduled first, then the dask-future

is_prime

. Since

get_primes_from_split

is scheduled first it gets higher priority, which causes the dask-execution to lock up, as it is waiting for the task to complete before executing anything else.

get_primes_from_split

naturally is waiting for

is_prime

to complete, which unfortunately will not execute at this point. Toying around with priorities, I managed to get

is_prime

to execute:

from typing import List, Tuple

import dask
from prefect import flow, get_run_logger, task
from prefect.context import get_run_context
from prefect_dask import get_dask_client


def is_prime(number: int) -> Tuple[int, bool]:
    if number == 2 or number == 3:
        return number, True
    if number % 2 == 0 or number < 2:
        return number, False

    for i in range(3, int(number ** 0.5) + 1, 2):
        if number % i == 0:
            return number, False

    return number, True


@task
def get_primes_from_split(min_number, max_number) -> List[int]:

    if min_number % 2 == 0:
        min_number += 1

    with get_dask_client() as client:
        futures = [client.submit(is_prime, n, priority=100) for n in range(min_number, max_number, 2)]

        maybe_primes = [future.result() for future in futures]

    return [value for value, flag in maybe_primes if flag]


@flow(name="example_prime_number_search")
def main(max_number: int = 1_000_000, split_size=10_000):
    log = get_run_logger()
    context = get_run_context()

    <http://log.info|log.info>("Task Runner: %s", context.task_runner.name)
    <http://log.info|log.info>("Searching for primes from up to %d", max_number)

    with dask.annotate(priority=0):
        futures = [get_primes_from_split.submit(x, x + split_size) for x in range(0, max_number + 1, split_size)]
        primes = [value for future in futures for value in future.result()]

    if len(primes) > 10:
        <http://log.info|log.info>("Found %d primes: %s, ...", sorted(primes)[::-1][:10])
    else:
        <http://log.info|log.info>("Found %d primes: %s", sorted(primes)[::-1][:10])

This causes dask to schedule a few instances of

get_primes_from_split

, which in turn schedules all its instances of

is_prime

.

is_prime

executes properly and starts returning its results, but it doesn't seem like

get_primes_from_split

picks up execution. I really don't understand what is going on here. Can anyone provide some insight into how do this kind of execution without reaching a deadlock like above?

Hi @Thomas Fredriksen, I don’t see you using DaskTaskRunner in the examples above. Am I missing something here?

Sorry, i should have mentioned. I am setting the taskrunner dynamically when deploying the flow. Some more information here: https://github.com/PrefectHQ/prefect/pull/7546

@Ryan Peden @Mason Menges Do you have any insights in this?

@Anna Geller Do you have any advice on this one?

My advice would be to build a minimal reproducible example and open an issue in the prefect-dask repo. Slack is really only helpful for quick ad hoc questions and this seems to be a bit more elaborate issue we would need to dive deeper