Kubernetes (dagster-k8s)¶

See also the Kubernetes deployment guide.

This library contains utilities for running Dagster with Kubernetes. This includes a Python API allowing Dagit to launch runs as Kubernetes Jobs, as well as a Helm chart you can use as the basis for a Dagster deployment on a Kubernetes cluster.

APIs¶

class dagster_k8s.K8sRunLauncher(service_account_name, instance_config_map, postgres_password_secret=None, dagster_home=None, job_image=None, image_pull_policy='Always', image_pull_secrets=None, load_incluster_config=True, kubeconfig_file=None, inst_data=None, job_namespace='default', env_config_maps=None, env_secrets=None, k8s_client_batch_api=None)[source]¶

RunLauncher that starts a Kubernetes Job for each pipeline run.

Encapsulates each pipeline run in a separate, isolated invocation of dagster-graphql.

You may configure a Dagster instance to use this RunLauncher by adding a section to your dagster.yaml like the following:

run_launcher:
    module: dagster_k8s.launcher
    class: K8sRunLauncher
    config:
        service_account_name: pipeline_run_service_account
        job_image: my_project/dagster_image:latest
        instance_config_map: dagster-instance
        postgres_password_secret: dagster-postgresql-secret

As always when using a ConfigurableClass, the values under the config key of this YAML block will be passed to the constructor. The full list of acceptable values is given below by the constructor args.

Parameters

service_account_name (str) – The name of the Kubernetes service account under which to run the Job.
job_image (Optional[str]) – The name of the image to use for the Job’s Dagster container. This image will be run with the command dagster api execute_run. When using user code deployments, the image should not be specified.
instance_config_map (str) – The name of an existing Volume to mount into the pod in order to provide a ConfigMap for the Dagster instance. This Volume should contain a dagster.yaml with appropriate values for run storage, event log storage, etc.
postgres_password_secret (Optional[str]) – The name of the Kubernetes Secret where the postgres password can be retrieved. Will be mounted and supplied as an environment variable to the Job Pod.
dagster_home (str) – The location of DAGSTER_HOME in the Job container; this is where the dagster.yaml file will be mounted from the instance ConfigMap specified above.
load_incluster_config (Optional[bool]) – Set this value if you are running the launcher within a k8s cluster. If True, we assume the launcher is running within the target cluster and load config using kubernetes.config.load_incluster_config. Otherwise, we will use the k8s config specified in kubeconfig_file (using kubernetes.config.load_kube_config) or fall back to the default kubeconfig. Default: True.
kubeconfig_file (Optional[str]) – The kubeconfig file from which to load config. Defaults to None (using the default kubeconfig).
image_pull_secrets (Optional[List[Dict[str, str]]]) – Optionally, a list of dicts, each of which corresponds to a Kubernetes LocalObjectReference (e.g., {'name': 'myRegistryName'}). This allows you to specify the `imagePullSecrets on a pod basis. Typically, these will be provided through the service account, when needed, and you will not need to pass this argument. See: https://kubernetes.io/docs/concepts/containers/images/#specifying-imagepullsecrets-on-a-pod and https://kubernetes.io/docs/reference/generated/kubernetes-api/v1.17/#podspec-v1-core.
image_pull_policy (Optional[str]) – Allows the image pull policy to be overridden, e.g. to facilitate local testing with kind. Default: "Always". See: https://kubernetes.io/docs/concepts/containers/images/#updating-images.
job_namespace (Optional[str]) – The namespace into which to launch new jobs. Note that any other Kubernetes resources the Job requires (such as the service account) must be present in this namespace. Default: "default"
env_config_maps (Optional[List[str]]) – A list of custom ConfigMapEnvSource names from which to draw environment variables (using envFrom) for the Job. Default: []. See: https://kubernetes.io/docs/tasks/inject-data-application/define-environment-variable-container/#define-an-environment-variable-for-a-container
env_secrets (Optional[List[str]]) – A list of custom Secret names from which to draw environment variables (using envFrom) for the Job. Default: []. See: https://kubernetes.io/docs/tasks/inject-data-application/distribute-credentials-secure/#configure-all-key-value-pairs-in-a-secret-as-container-environment-variables

dagster_k8s.k8s_job_executor ExecutorDefinition[source]¶

Executor which launches steps as Kubernetes Jobs. This executor is experimental.

To add the Kubernetes Job executor in addition to the default_executors, you should add it to the executor_defs defined on a ModeDefinition as follows:

from dagster import ModeDefinition, default_executors, pipeline
from dagster_k8s import k8s_job_executor


@pipeline(mode_defs=[ModeDefinition(executor_defs=default_executors + [k8s_job_executor])])
def k8s_pipeline():
    pass

Then you can configure the executor with run config (either via a PresetDefinition or the Dagit playground) as follows:

execution:
  k8s:
    config:
      job_namespace: 'some-namespace'
      image_pull_policy: ...
      image_pull_secrets: ...
      service_account_name: ...
      env_config_maps: ...
      env_secrets: ...
      job_image: ... # leave out if using userDeployments

class dagster_k8s.K8sScheduler(dagster_home, service_account_name, instance_config_map, postgres_password_secret, job_image, load_incluster_config=True, scheduler_namespace='default', image_pull_policy='Always', image_pull_secrets=None, kubeconfig_file=None, inst_data=None, env_config_maps=None, env_secrets=None)[source]¶

Scheduler implementation on top of Kubernetes CronJob.

Enable this scheduler by adding it to your dagster.yaml, or by configuring the scheduler section of the Helm chart https://github.com/dagster-io/dagster/tree/master/helm/dagster

Python API¶

The K8sRunLauncher allows Dagit instances to be configured to launch new runs by starting per-run Kubernetes Jobs. To configure the K8sRunLauncher, your dagster.yaml should include a section like:

run_launcher:
  module: dagster_k8s.launcher
  class: K8sRunLauncher
  config:
    image_pull_secrets:
    service_account_name: dagster
    job_image: "my-company.com/image:latest"
    dagster_home: "/opt/dagster/dagster_home"
    postgres_password_secret: "dagster-postgresql-secret"
    image_pull_policy: "IfNotPresent"
    job_namespace: "dagster"
    instance_config_map: "dagster-instance"
    env_config_maps:
      - "dagster-k8s-job-runner-env"
    env_secrets:
      - "dagster-k8s-some-secret"

Helm chart¶

For local dev (e.g., on kind or minikube):

helm install \
    --set dagit.image.repository="dagster.io/buildkite-test-image" \
    --set dagit.image.tag="py37-latest" \
    --set job_runner.image.repository="dagster.io/buildkite-test-image" \
    --set job_runner.image.tag="py37-latest" \
    --set imagePullPolicy="IfNotPresent" \
    dagster \
    helm/dagster/

Upon installation, the Helm chart will provide instructions for port forwarding Dagit and Flower (if configured).

Running tests¶

To run the unit tests:

pytest -m "not integration"

To run the integration tests, you must have Docker, kind, and helm installed.

On macOS:

brew install kind
brew install helm

Docker must be running.

You may experience slow first test runs thanks to image pulls (run pytest -svv --fulltrace for visibility). Building images and loading them to the kind cluster is slow, and there is no visibility into the progress of the load.

NOTE: This process is quite slow, as it requires bootstrapping a local kind cluster with Docker images and the dagster-k8s Helm chart. For faster development, you can either:

Keep a warm kind cluster
Use a remote K8s cluster, e.g. via AWS EKS or GCP GKE

Instructions are below.

Faster local development (with kind)¶

You may find that the kind cluster creation, image loading, and kind cluster creation loop is too slow for effective local dev.

You may bypass cluster creation and image loading in the following way. First add the --no-cleanup flag to your pytest invocation:

pytest --no-cleanup -s -vvv -m "not integration"

The tests will run as before, but the kind cluster will be left running after the tests are completed.

For subsequent test runs, you can run:

pytest --kind-cluster="cluster-d9971c84d44d47f382a2928c8c161faa" --existing-helm-namespace="dagster-test-95590a" -s -vvv -m "not integration"

This will bypass cluster creation, image loading, and Helm chart installation, for much faster tests.

The kind cluster name and Helm namespace for this command can be found in the logs, or retrieved via the respective CLIs, using kind get clusters and kubectl get namespaces. Note that for kubectl and helm to work correctly with a kind cluster, you should override your kubeconfig file location with:

kind get kubeconfig --name kind-test > /tmp/kubeconfig
export KUBECONFIG=/tmp/kubeconfig

Manual kind cluster setup¶

The test fixtures provided by dagster-k8s automate the process described below, but sometimes it’s useful to manually configure a kind cluster and load images onto it.

First, ensure you have a Docker image appropriate for your Python version. Run, from the root of the repo:

./python_modules/dagster-test/dagster_test/test_project/build.sh 3.7.6 \
    dagster.io.priv/buildkite-test-image:py37-latest

In the above invocation, the Python majmin version should be appropriate for your desired tests.

Then run the following commands to create the cluster and load the image. Note that there is no feedback from the loading process.

kind create cluster --name kind-test
kind load docker-image --name kind-test dagster.io/dagster-docker-buildkite:py37-latest

If you are deploying the Helm chart with an in-cluster Postgres (rather than an external database), and/or with dagster-celery workers (and a RabbitMQ), you’ll also want to have images present for rabbitmq and postgresql:

docker pull docker.io/bitnami/rabbitmq
docker pull docker.io/bitnami/postgresql

kind load docker-image --name kind-test docker.io/bitnami/rabbitmq:latest
kind load docker-image --name kind-test docker.io/bitnami/postgresql:latest

Then you can run pytest as follows:

pytest --kind-cluster=kind-test

Faster local development (with an existing K8s cluster)¶

If you already have a development K8s cluster available, you can run tests on that cluster vs. running locally in kind.

For this to work, first build and deploy the test image to a registry available to your cluster. For example, with a private ECR repository:

./python_modules/dagster-test/dagster_test/test_project/build.sh 3.7.6
docker tag dagster-docker-buildkite:latest $AWS_ACCOUNT_ID.dkr.ecr.us-west-2.amazonaws.com/dagster-k8s-tests:2020-04-21T21-04-06

aws ecr get-login --no-include-email --region us-west-1 | sh
docker push $AWS_ACCOUNT_ID.dkr.ecr.us-west-1.amazonaws.com/dagster-k8s-tests:2020-04-21T21-04-06

Then, you can run tests on EKS with:

export DAGSTER_DOCKER_IMAGE_TAG="2020-04-21T21-04-06"
export DAGSTER_DOCKER_REPOSITORY="$AWS_ACCOUNT_ID.dkr.ecr.us-west-2.amazonaws.com"
export DAGSTER_DOCKER_IMAGE="dagster-k8s-tests"

# First run with --no-cleanup to leave Helm chart in place
pytest --cluster-provider="kubeconfig" --no-cleanup -s -vvv

# Subsequent runs against existing Helm chart
pytest --cluster-provider="kubeconfig" --existing-helm-namespace="dagster-test-<some id>" -s -vvv

Validating Helm charts¶

To test / validate Helm charts, you can run:

helm install dagster --dry-run --debug helm/dagster
helm lint

Enabling GCR access from Minikube¶

To enable GCR access from Minikube:

kubectl create secret docker-registry element-dev-key \
    --docker-server=https://gcr.io \
    --docker-username=oauth2accesstoken \
    --docker-password="$(gcloud auth print-access-token)" \
    --docker-email=my@email.com

A note about PVCs¶

Both the Postgres and the RabbitMQ Helm charts will store credentials using Persistent Volume Claims, which will outlive test invocations and calls to helm uninstall. These must be deleted if you want to change credentials. To view your pvcs, run:

kubectl get pvc

Testing Redis¶

The Redis Helm chart installs w/ a randomly-generated password by default; turn this off:

helm install dagredis stable/redis --set usePassword=false

Then, to connect to your database from outside the cluster execute the following commands:

kubectl port-forward --namespace default svc/dagredis-master 6379:6379
redis-cli -h 127.0.0.1 -p 6379