Is there any way to define a

ContainerTask

with a toleration or affinity? It looks like I may have define a custom

Task

that overrides the pod spec construction method right now?

seems like a reasonable thing to forward a user-defined config into

super().__init__

, unless I’m missing something @high-accountant-32689 @freezing-airport-6809?

@melodic-magician-71351 have you seen the issue on adding PodTemplates to tasks? I believe this will cover your use-case right? Something like:

t = ContainerTask(
    name="foo",
    image="bar:latest",
    command=[
        "echo",
        "hello",
        "{{.inputs.b}}",
        "/var/outputs",
    ],
    pod_template=PodTemplate(
        spec=PodSpec(
            tolerations={...}
        )
    )
)

This have been pretty high priority for us, because many users are interested. We're planning on having a beta release out later this week.

@hallowed-mouse-14616 yes I think that's a good idea but not relevant here. I want to define a task without a python function (rather with a container image, and a command)

So we plan to support

ContainerTask

. We'll have a beta out later this week - but unfortunately this support will not be a part of it. It looks like you've dove into this a little bit, would you be interested in contributed support for specifying

pod_template

for

ContainerTask

?

Sounds good. I'll have to look into the design to see how easy it is to adapt to ContainerTask, but it doesn't seem like it would be too difficult.

Yeah, once we get the rest of the infrastructure to support this feature it should be trivial.

Yeah looks really straight forward, especially with the other MR already doing the heavy lifting on the SDK side. AFAICT

ContainerTask

becomes basically a no-op once you have a pod template? I'm assuming the sidecar container gets injected server-side for

raw-container

type tasks? because I don't see its definition anywhere in python.

I'm assuming the sidecar container gets injected server-side

absolutley, actually happens here

• Cool. Then the only question I have is how a primary container defined as part of a template should be merged with the one generated by the

ContainerTask

definition; ◦ Fail if both are defined at all. ◦ Deep merge failing on mismatch ◦ Deep merge warning/log on mismatch My understanding of how people use PodTemplates is incomplete, but I'm assuming there are compelling reasons to have primary container definitions included (e.g. volume mounts), so I would think the 3rd option is best.

@melodic-magician-71351 see also https://github.com/flyteorg/flyte/issues/3055

@quaint-diamond-37493 my understanding from @hallowed-mouse-14616 is that the pod plugin is going to be deprecated in favor of the

pod_template

argument, so this would solve that issue indirectly

yes, indeed

@melodic-magician-71351 I think you're right. A deep merge with logging may be the best approach. Currently we support a "default" PodTemplate as a k8s resource that gets applied to all Pods created by Flyte. This uses the PodTemplate as base configuration and then anything the user specifies overrides it. I think it makes sense to emulate this functionality - of course open to discussion here.

ContainerTask also seems be broken for local execution 😞

Yeah, unfortunately I'm not sure if it's necessarily "broken" as much as "not supported". Running a raw container means there needs to be a dependency on some CRI, so it can't really be run as a raw python func.

Taking an optional path to a CRI binary, and allowing string outputs or directory outputs to be output path or captured stdout from a subprocess seems reasonable to me? At the very least it should raise a NotImplemented or something

That would make sense to me! cc @high-accountant-32689 ^^ updating ContainerTask local execution.

Yeah, I actually wanted to start implementing local execution (assuming docker) via subprocess, but quickly realized that this would be of no help to me, as I need GPU support for all workflows anyway and I don't have it locally... So I immediately went with the remote cluster.

@melodic-magician-71351 / @quaint-diamond-37493 you can mock today - But I think some one added a dependency to docker. That is not acceptable to many people because of docker license issues - If you think you can add support and upstream highly welcome to do this. The core team currently does not have a plan to work on it. Caution use oci client

I wouldn't add a dependency, I would just ask users to pass a path to a client binary in their local environment, if they wanted local execution. Probably best set in config somewhere globally.

That sounds good