@ancient-wolf-19325 print an example of how to define resource limits in a

PodTemplate

via k8s yaml, and any task annotations required to use that template in all tasks in a workflow. show an example of defining the defaults for new tasks and an example of how to override the defaults for a specific task

To define resource limits in a

PodTemplate

via Kubernetes YAML, create a

PodTemplate

resource with the desired limits. To use this template for all tasks in a workflow, set the

default-pod-template-name

in the FlytePropeller config. No special task annotation is required unless you want to override the default for a specific task. For task-level overrides, use the

pod_template

argument in the

@task

decorator. According to FlytePropeller PodTemplate docs and Flyte 1.4 release notes, here are examples: 1. PodTemplate YAML with resource limits:

apiVersion: v1
kind: PodTemplate
metadata:
  name: flyte-default-template
  namespace: flyte
template:
  spec:
    containers:
      - name: default
        image: docker.io/rwgrim/docker-noop
        resources:
          limits:
            cpu: "2"
            memory: "4Gi"
          requests:
            cpu: "1"
            memory: "2Gi"

2. FlytePropeller config to use the template by default:

plugins:
  k8s:
    default-pod-template-name: flyte-default-template

3. Task definition using platform defaults:

from flytekit import task

@task
def my_task():
    ...

4. Override PodTemplate for a specific task:

from flytekit import task
from flytekitplugins.kubernetes.pod import PodTemplate
from kubernetes.client import V1PodSpec, V1Container, V1ResourceRequirements

custom_pod_template = PodTemplate(
    primary_container_name="primary",
    pod_spec=V1PodSpec(
        containers=[
            V1Container(
                name="primary",
                resources=V1ResourceRequirements(
                    limits={"cpu": "4", "memory": "8Gi"},
                    requests={"cpu": "2", "memory": "4Gi"},
                ),
            ),
        ]
    ),
)

@task(pod_template=custom_pod_template)
def special_task():
    ...

This setup uses the default PodTemplate for all tasks unless overridden at the task level. Would you like more detail on how Flyte merges these configurations? Sources: - FlytePropeller PodTemplate docs - Flyte 1.4 release notes 👍 0 button 👎 0 button 💡 Hint: Mention @runllm in the thread for followups.

@ancient-wolf-19325 i have this defined in my flyte helmfile:

k8sPlugin:
            default-pod-template-name: default-pod-template
            default-env-vars:
              - FLYTE_AZURE_STORAGE_ACCOUNT_NAME: asdfasdf
          taskResources:
            defaults:
              cpu: 250m
              memory: 2Gi
            limits:
              cpu: null
              memory: 8Gi

I want to be able to override those resource values in specific namespaces, so i created a default pot template and apply it. here is an example:

spec:
    serviceAccountName: flyte-task-default
    volumes:
      - name: ingest-config
        configMap:
          name: ingest-config
    containers:
      - name: default
        image: cr.flyte.org/flyteorg/flytekit:py3.11-1.13.0
        resources:
          requests:
            memory: "8Gi"
          limits:
            memory: "16Gi"

However, when a pod in that namespace is spun up for a task, the resource definition is the flyte propeller defined defaults; i.e., it doesn't use the pod template override. why? how can i fix it?

The behavior you're seeing is due to Flyte's resource merging logic: for standard resources like CPU and memory, FlytePropeller's

taskResources.defaults

and

taskResources.limits

take precedence over values set in a namespace's default PodTemplate, unless you explicitly override them at the task level. PodTemplate resource requests/limits for CPU and memory are ignored if

taskResources

are set, but extended resources (like GPUs) from the PodTemplate are preserved. This is a known limitation and is discussed in Flyte Issue #6482. To override resource values per-namespace using a PodTemplate, you must remove or unset the corresponding

taskResources.defaults

and

taskResources.limits

in your FlytePropeller config for that namespace, or explicitly set the resources at the task level. Otherwise, Flyte will always use the config defaults for CPU and memory, not the PodTemplate values. Would you like more detail on how to structure your configs for per-namespace overrides? Sources: - Flyte Issue #6482 👍 0 button 👎 0 button 💡 Hint: Mention @runllm in the thread for followups.

@ancient-wolf-19325 which version of flytekit introduces

flytekitplugins

? i'm using 1.13 and don't see that dependency

The

flytekitplugins

namespace and separately installable plugin packages were introduced starting with

flytekit

version 0.16.0b1. This change allowed plugins to be installed independently from the core SDK. If you are using

flytekit

1.13, you should see and be able to install

flytekitplugins-*

packages, as this namespace and packaging structure is present in all 1.x releases and later. If you do not see the dependency, ensure you are installing the plugin (e.g.,

pip install flytekitplugins-spark

) separately, as plugins are not bundled with the core

flytekit

package by default. See Flytekit plugin separation issue. Would you like more detail on how to install or use these plugins? Sources: - Flytekit plugin separation issue 👍 0 button 👎 0 button 💡 Hint: Mention @runllm in the thread for followups.

@ancient-wolf-19325 when was

flytekitplugins.pod

introduced?

The

flytekitplugins.pod

plugin was introduced in Flytekit version 0.16.0b1, when plugins were separated into their own packages and the

flytekitplugins

namespace was established. This allowed users to install the pod plugin independently as

flytekitplugins-pod

. See Flytekit plugin separation issue. Would you like more detail on its usage or history? Sources: - Flytekit plugin separation issue 👍 0 button 👎 0 button 💡 Hint: Mention @runllm in the thread for followups.