V1KFReplica

polyaxon.polyflow.run.kubeflow.replica.V1KFReplica(replicas=None, environment=None, connections=None, volumes=None, init=None, sidecars=None, container=None)

Kubeflow-Replica provides an interface to define a replica for TFJob/MPIJob/PytorchJob.

YAML usage

worker:
  replicas: 2
  environment:
  connections:
  volumes:
  init:
  sidecars:
  container:

Python usage

from polyaxon.polyflow import V1Environment, V1Init, V1KFReplica
from polyaxon.k8s import k8s_schemas
replica = V1KFReplica(
    replicas=2,
    environment=V1Environment(...),
    connections=["connection-name1"],
    volumes=[k8s_schemas.V1Volume(...)],
    init=[V1Init(...)],
    sidecars=[k8s_schemas.V1Container(...)],
    container=k8s_schemas.V1Container(...),
)

Fields

replicas

The number of processes of this replica type.

worker:
  replicas: 2

environment

Optional environment section, it provides a way to inject pod related information.

worker:
  environment:
    labels:
       key1: "label1"
       key2: "label2"
     annotations:
       key1: "value1"
       key2: "value2"
     nodeSelector:
       node_label: node_value
     ...
 ...

connections

A list of connection names to resolve for the job.

If you are referencing a connection it must be configured. All referenced connections will be checked:
  • If they are accessible in the context of the project of this run

  • If the user running the operation can have access to those connections

After checks, the connections will be resolved and inject any volumes, secrets, configMaps, environment variables for your main container to function correctly.

worker:
  connections: [connection1, connection2]

volumes

A list of Kubernetes Volumes to resolve and mount for your jobs.

This is an advanced use-case where configuring a connection is not an option.

When you add a volume you need to mount it manually to your container(s).

worker:
  volumes:
    - name: volume1
      persistentVolumeClaim:
        claimName: pvc1
  ...
  container:
    name: myapp-container
    image: busybox:1.28
    command: ['sh', '-c', 'echo custom init container']
    volumeMounts:
    - name: volume1
      mountPath: /mnt/vol/path

init

A list of init handlers and containers to resolve for the job.

If you are referencing a connection it must be configured. All referenced connections will be checked:
  • If they are accessible in the context of the project of this run

  • If the user running the operation can have access to those connections

worker:
  init:
    - artifacts:
        dirs: ["path/on/the/default/artifacts/store"]
    - connection: gcs-large-datasets
      artifacts:
        dirs: ["data"]
      container:
        resources:
          requests:
            memory: "256Mi"
            cpu: "500m"
    - container:
      name: myapp-container
      image: busybox:1.28
      command: ['sh', '-c', 'echo custom init container']

sidecars

A list of sidecar containers that will used as sidecars.

worker:
  sidecars:
    - name: sidecar2
      image: busybox:1.28
      command: ['sh', '-c', 'echo sidecar2']
    - name: sidecar1
      image: busybox:1.28
      command: ['sh', '-c', 'echo sidecar1']
      resources:
        requests:
          memory: "128Mi"
          cpu: "500m"

container

The main Kubernetes Container that will run your experiment training or data processing logic.

worker:
  container:
    name: tensorflow:2.1
    init:
      - connection: my-tf-code-repo
    command: ["python", "/plx-context/artifacts/my-tf-code-repo/model.py"]