Hi, could you give an advice, is it possible to make a task based on class? Something like

@task
class ConfigurableFunction:
    def __init__(self, x:int):
        self._x = x
    def __call__(self, y:int) -> int:
        return self._x * y

And then usage

f10 = ConfigurableFunction(10)
f25 = ConfigurableFunction(25)

@workflow
def wf():
   a = 10
   b = f10(a)
   c = f25(b)
   return c

You will have to work a custom task resolver

Is it the same as writing Custom task plugin?

It’s slightly different. It tells flytekit how to find and load the task when it runs on the Flyte pod. See the default implementation here, and a few examples: • Airflow task resolver • CloudPickle resolver

I don’t understand how the task resolver appear on the Flyte pod. Do you know maybe where I can find information about that?

they are configured here: https://github.com/flyteorg/flytekit/blob/15dee9579a000696539b63745d0d036647987c0e/flytekit/core/python_auto_container.py#L142-L162

Does this work because

flykit.core.python_auto_container.default_task_resolver

is importable in the pod? How can I achieve the same with custom task resolver?

Flyte propeller only knows what container to run and what arguments to pass into it. The resolver and task are loaded in at the

entrypoint.py

: https://github.com/flyteorg/flytekit/blob/15dee9579a000696539b63745d0d036647987c0e/flytekit/bin/entrypoint.py#L392-L396 So as long as the task and its resolver are defined and loadable in your container via

pyflyte run

or

pyflyte register

it should work. Best way to iterate on this is using the flyte sandbox. I’d start by just subclassing

PythonFunctionTask

and supplying your own task resolver (just a copy of the default task resolver) into it to get a sense of how it works. Then start modifying the resolver and the

PythonFunctionTask

subclass to work with classes instead of functions.

Thank you! I'll try