#379 Fix deletion of elastic task resource requests Pull request opened by fg91 TL;DR Currently, the following elastic task's pods will not have any resource requests (1.8 release):

@task(
    task_config=Elastic(
        nnodes=2,
        nproc_per_node=2,
    ),
    requests=Resources(cpu="1", mem="1Gi", gpu="1"),
)
def train():
    ...

This PR fixes this. Type ☑︎ Bug Fix ☐ Feature ☐ Plugin Are all requirements met? ☑︎ Code completed ☑︎ Smoke tested ☑︎ Unit tests added ☐ Code documentation added ☑︎ Any pending items have an associated Issue Complete description The kubeflow plugin was recently refactored in flytekit (flyteorg/flytekit#1627) as well as in flyteplugins (#345). After the refactoring, users can set e.g. different resources for master and worker replicas of

PyTorch

tasks:

@task(
    task_config=PyTorch(
        worker=Worker(
            requests=Resources(cpu="2", mem="2Gi"),
            limits=Resources(cpu="4", mem="2Gi"),
        ),
        master=...
      )
      requests=Resources(cpu="1", mem="1Gi", gpu="1"),
)
....

If the user overrides one of the resource requests in the

task_config

, in the backend we override the resources configured on the task level here. Now, in a

@task(task_config=PyTorch()..)

task, when the user doesn't override the resources, we send

task_models.Resources(requests=[], limits=[])

(which is different from

None/nil

, see here) to the backend as part of the worker spec. We go into this

if

statement but not into this one within the first. As a consequence, in this case we don't overwrite the resource request set in

@task(requests=Resources())

. This makes sense. However, the

Elastic

task config in

flytekitplugins

currently doesn't allow overwriting the resource request since elastic pytorch training jobs don't have a differentiation between master and worker replicas. There are only workers which means we can simply use the normal

@task(requests=Resources())

to specify their resources. This, however, means that in this case we send

None/nil

to the backend as part of the worker specs resource request. Here we go into the

else

case and simply delete what the user set as

@task(requests=Resources())

. This seems wrong and this PR changes this. Tracking Issue NA fixes https://github.com/flyteorg/flyte/issues/ Follow-up issue NA flyteorg/flyteplugins Codecov: 62.96% (-0.01%) compared to 13ef7d6 Codecov: 0.00% of diff hit (target 62.96%) ✅ 5 other checks have passed 5/7 successful checks