Airflow KubernetesJobOperator

An airflow operator that executes a task in a kubernetes cluster, given a yaml configuration or an image url.

If you like it * it, so other people would also use it.

Contributions are welcome. See here.

If you enjoyed using this repo, please consider posting in the use cases and testimonials issue.

Supports

1. Running tasks as Kubernetes Jobs.
2. Running tasks as Kubernetes Pods, with sidecar support (see metadata annotations).
3. Running tasks as multi resource (see example below, similar to kubectl apply).
4. Running tasks as custom resources (help).
5. Auto detection of kubernetes namespace and config.
6. Pod and participating resources logs -> airflow.
7. Full kubernetes error logs on failure.
8. Integrated operator airflow config, see below.
9. Integrated Jinja2 support for file templates with flag.
10. XCom
11. Log based events
12. Tested and working on google cloud composer.
13. Airflow > 2.0.0 (Airflow 1 is supported but deprecated)

Two operator classes are available

1. KubernetesJobOperator - Supply a kubernetes configuration (yaml file, yaml string or a list of python dictionaries) as the body of the task.
2. KubernetesLegacyJobOperator (only airflow 2.0 and up) - Defaults to a kubernetes job definition, and supports the same arguments as the KubernetesPodOperator. i.e. replace with the KubernetesPodOperator for legacy support.

Known issues

1. When running long tasks on AWS the job is activly disconnected from the server (Connection broken: ConnectionResetError, 104, 'Connection reset by peer'), see here and here. We are looking for a solution for it, but have no AWS resources for testing. Contributions welcome.

Install

To install using pip @ https://pypi.org/project/airflow-kubernetes-job-operator,

pip install airflow_kubernetes_job_operator

To install from master branch,

pip install git+https://github.com/LamaAni/KubernetesJobOperator.git@master

To install from a release (tag)

pip install git+https://github.com/LamaAni/KubernetesJobOperator.git@[tag]

TL;DR

When using from within kubernetes, please use the following RBAC rules for the pod service account.

Example airflow DAG

from airflow import DAG
from airflow_kubernetes_job_operator.kubernetes_job_operator import KubernetesJobOperator
from airflow_kubernetes_job_operator.kubernetes_legacy_job_operator import KubernetesLegacyJobOperator
from airflow.utils.dates import days_ago

default_args = {"owner": "tester", "start_date": days_ago(2), "retries": 0}
dag = DAG("job-tester", default_args=default_args, description="Test base job operator", schedule_interval=None)

job_task = KubernetesJobOperator(
    task_id="from-image",
    dag=dag,
    image="ubuntu",
    command=["bash", "-c", 'echo "all ok"'],
)

body = {"kind": "Pod"}  # The body or a yaml string (must be valid)
job_task_from_body = KubernetesJobOperator(dag=dag, task_id="from-body", body=body)

body_filepath = "./my_yaml_file.yaml" # Can be relative to this file, or abs path.
job_task_from_yaml = KubernetesJobOperator(dag=dag, task_id="from-yaml", body_filepath=body_filepath)

# Legacy compatibility to KubernetesPodOperator
legacy_job_task = KubernetesLegacyJobOperator(
    task_id="legacy-image-job",
    image="ubuntu",
    cmds=["bash", "-c", 'echo "all ok"'],
    dag=dag,
    is_delete_operator_pod=True,
)

Example (multi resource) task yaml

NOTE: that the success/failure of the task is tracked only on the first resource, no matter of its kind. Currently native support exists for Pods and Jobs only, though you can always add a custom resource.

# First resource: this resource will be tracked by the operator. Other resources will not be tracked.
apiVersion: batch/v1
kind: Job
metadata:
  name: test-job # not required. Will be a prefix to task name
  finalizers:
    - foregroundDeletion
spec:
  template:
    metadata:
      labels:
        app: test-task-pod
    spec:
      restartPolicy: Never
      containers:
        - name: job-executor
          image: ubuntu
          command:
            - bash
            - -c
            - |
              #/usr/bin/env bash
              echo "OK"
  backoffLimit: 0
---
apiVersion: v1
kind: Service
metadata:
  name: test-service # not required, will be a prefex to task name.
spec:
  selector:
    app: test-task-pod
  ports:
    - port: 8080
      targetPort: 8080

YAML field augmentation

The operator auto augments the following fields on the first resource containers,

1. envs are added to all pods
2. special envs are added:
   1. KUBERNETES_JOB_OPERATOR_RESOURCES - A list of all resource names.

All other changes are made on the main container (the first container) of the first resource only.

Configuration

Airflow config extra sections,

[kubernetes_job_operator]
# The task kube resources delete policy. Can be: Never, Always, IfFailed, IfSucceeded
delete_policy=IfSucceeded
# The default object type to execute with (legacy, or image). Can be: Pod, Job
default_execution_object=Job

# Logs
detect_kubernetes_log_level=True
show_kubernetes_timestamps=False
# Shows the runner id in the log (for all runner logs.)
show_runner_id=False

# Tasks (Defaults)
# Wait to first connect to kubernetes.
startup_timeout_seconds=120
# if true, will parse the body when building the dag. Otherwise only while executing.
validate_body_on_init=False

# Comma seperated list of where to look for the kube config file. Will be added to the top
# of the search list, in order.
kube_config_extra_locations=

To set these values through the environment follow airflow standards,

export AIRFLOW__[section]__[item] = value

Kubernetes job operator input variables

Metadata annotations

When executing Pod resources, via yaml, we can define the main container that the status is taken from. Thus we allow other sidecars to be loaded before the pod completes (this for now is not available for Job),

apiVersion: v1
kind: Pod
metadata:
  ...
  annotations:
    kubernetes_job_operator.main_container: main
spec:
  restartPolicy: Never
  containers:
    - name: main
      image: 'ubuntu:20.10'
      command:
        - sleep
        - '10'
    - name: side-car
      image: 'ubuntu:20.10'
      command:
        - sleep
        - infinity

When executed, only the status of the main container will be taken into account when determining the state of the task. Note, if the delete policy is not one of Always,IfFailed, the sidecar will continue executing on failure.

XCom

The implementation of XCom via the KubernetesJobOperator differes from the one by KuberenetesPodsOperator, and therefore no backward compatible implementation for XCom currently exists.

To use xcom with KubernetesJobOperator simply add a log line to your pod log output,

::kube_api:xcom={"a":2,"[key]":"[value]"}

Note the value of the xcom must be in json format (for the default parser).

Kubernetes RBAC rules

The following are the Kubernetes service account RBAC rules required for executing jobs via the operator. You can find how to configure a service account, with rules here.

apiVersion: rbac.authorization.k8s.io/v1
kind: Role
metadata:
  name: airflow-kube-access-user
rules:
  - verbs:
      - get
      - watch
      - list
    apiGroups:
      - ''
    resources:
      - pods/log
      - pods
  # Required for executing the job
  - verbs:
      - create
      - get
      - delete
      - watch
      - list
      - patch
      - update
    apiGroups:
      - batch
    resources:
      - jobs
  # Required for using configmaps in the job
  - verbs:
      - create
      - get
      - delete
      - watch
      - list
      - patch
      - update
    apiGroups:
      - ''
    resources:
      - configmaps
  # Required if reading secrets in the job.
  - verbs:
      - create
      - get
      - delete
      - watch
      - list
      - patch
      - update
    apiGroups:
      - ''
    resources:
      - secrets

Why would this be better than the KubernetesPodOperator?

The KubernetesJobOperator allows more advanced features such as multiple resource execution, custom resource executions, creation event and pod log tracking, proper resource deletion (after task is complete or on error, when task is cancelled) and more.

Further, in the KubernetesPodOperator the monitoring between the worker pod and airflow is done by an internal loop which consumes worker resources. In this operator, an async threaded approach was taken which reduces resource consumption on the worker. Further, since the logging and parsing of data is done outside of the main worker thread, the worker is "free" do handle other tasks without interruption.

Finally, using the KubernetesJobOperator you are free to use other resources like the kubernetes Job to execute your tasks. These are better designed for use in kubernetes and are "up with the times".

Contribution

Are welcome, please post issues or PR's if needed.

Licence

Copyright © Zav Shotan and other contributors. It is free software, released under the MIT licence, and may be redistributed under the terms specified in LICENSE.