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

dagster_k8s.K8sRunLauncher RunLauncher[source]

Config Schema:
job_image (Union[dagster.StringSource, None], optional)

Docker image to use for launched Jobs. If this field is empty, the image that was used to originally load the Dagster repository will be used. (Ex: “mycompany.com/dagster-k8s-image:latest”).

image_pull_policy (Union[dagster.StringSource, None], optional)

Image pull policy to set on launched Pods.

image_pull_secrets (Union[List[strict dict], None], optional)

Specifies that Kubernetes should get the credentials from the Secrets named in this list.

service_account_name (Union[dagster.StringSource, None], optional)

The name of the Kubernetes service account under which to run.

env_config_maps (Union[List[dagster.StringSource], None], optional)

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 (Union[List[dagster.StringSource], None], optional)

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

env_vars (Union[List[String], None], optional)

A list of environment variables to inject into the Job. Each can be of the form KEY=VALUE or just KEY (in which case the value will be pulled from the current process). 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

volume_mounts (List[permissive dict], optional)

A list of volume mounts to include in the job’s container. Default: []. See: https://v1-18.docs.kubernetes.io/docs/reference/generated/kubernetes-api/v1.18/#volumemount-v1-core

Default Value: []

volumes (List[permissive dict], optional)

A list of volumes to include in the Job’s Pod. Default: []. For the many possible volume source types that can be included, see: https://v1-18.docs.kubernetes.io/docs/reference/generated/kubernetes-api/v1.18/#volume-v1-core

Default Value: []

labels (permissive dict, optional)

Labels to apply to all created pods. See: https://kubernetes.io/docs/concepts/overview/working-with-objects/labels

resources (Union[strict dict, None], optional)

Compute resource requirements for the container. See: https://kubernetes.io/docs/concepts/configuration/manage-resources-containers/

instance_config_map (dagster.StringSource)

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 (dagster.StringSource, optional)

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.Secret must contain the key "postgresql-password" which will be exposed in the Job environment as the environment variable DAGSTER_PG_PASSWORD.

dagster_home (dagster.StringSource, optional)

The location of DAGSTER_HOME in the Job container; this is where the dagster.yaml file will be mounted from the instance ConfigMap specified here. Defaults to /opt/dagster/dagster_home.

Default Value: ‘/opt/dagster/dagster_home’

load_incluster_config (Bool, optional)

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 Value: True

kubeconfig_file (Union[String, None], optional)

The kubeconfig file from which to load config. Defaults to using the default kubeconfig.

Default Value: None

fail_pod_on_run_failure (Bool, optional)

Whether the launched Kubernetes Jobs and Pods should fail if the Dagster run fails

job_namespace (dagster.StringSource, optional)

Default Value: ‘default’

RunLauncher that starts a Kubernetes Job for each Dagster job run.

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

You can 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: your_service_account
    job_image: my_project/dagster_image:latest
    instance_config_map: dagster-instance
    postgres_password_secret: dagster-postgresql-secret
dagster_k8s.k8s_job_executor ExecutorDefinition[source]

Config Schema:
job_image (Union[dagster.StringSource, None], optional)

Docker image to use for launched Jobs. If this field is empty, the image that was used to originally load the Dagster repository will be used. (Ex: “mycompany.com/dagster-k8s-image:latest”).

image_pull_policy (Union[dagster.StringSource, None], optional)

Image pull policy to set on launched Pods.

image_pull_secrets (Union[List[strict dict], None], optional)

Specifies that Kubernetes should get the credentials from the Secrets named in this list.

service_account_name (Union[dagster.StringSource, None], optional)

The name of the Kubernetes service account under which to run.

env_config_maps (Union[List[dagster.StringSource], None], optional)

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 (Union[List[dagster.StringSource], None], optional)

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

env_vars (Union[List[String], None], optional)

A list of environment variables to inject into the Job. Each can be of the form KEY=VALUE or just KEY (in which case the value will be pulled from the current process). 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

volume_mounts (List[permissive dict], optional)

A list of volume mounts to include in the job’s container. Default: []. See: https://v1-18.docs.kubernetes.io/docs/reference/generated/kubernetes-api/v1.18/#volumemount-v1-core

Default Value: []

volumes (List[permissive dict], optional)

A list of volumes to include in the Job’s Pod. Default: []. For the many possible volume source types that can be included, see: https://v1-18.docs.kubernetes.io/docs/reference/generated/kubernetes-api/v1.18/#volume-v1-core

Default Value: []

labels (permissive dict, optional)

Labels to apply to all created pods. See: https://kubernetes.io/docs/concepts/overview/working-with-objects/labels

resources (Union[strict dict, None], optional)

Compute resource requirements for the container. See: https://kubernetes.io/docs/concepts/configuration/manage-resources-containers/

job_namespace (dagster.StringSource, optional)

retries (selector, optional)

Whether retries are enabled or not. By default, retries are enabled.

Default Value:
{
    "enabled": {}
}
Config Schema:
enabled (strict dict, optional)
Default Value:
{}
disabled (strict dict, optional)
Default Value:
{}
max_concurrent (dagster.IntSource, optional)

Limit on the number of pods that will run concurrently within the scope of a Dagster run. Note that this limit is per run, not global.

Executor which launches steps as Kubernetes Jobs.

To use the k8s_job_executor, set it as the executor_def when defining a job:

from dagster import job
from dagster_k8s import k8s_job_executor

@job(executor_def=k8s_job_executor)
def k8s_job():
    pass

Then you can configure the executor with run config as follows:

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

max_concurrent limits the number of pods that will execute concurrently for one run. By default there is no limit- it will maximally parallel as allowed by the DAG. Note that this is not a global limit.

Configuration set on the Kubernetes Jobs and Pods created by the K8sRunLauncher will also be set on Kubernetes Jobs and Pods created by the k8s_job_executor.

Ops

dagster_k8s.k8s_job_op = <dagster.core.definitions.op_definition.OpDefinition object>[source]

Config Schema:
image_pull_policy (Union[dagster.StringSource, None], optional)

Image pull policy to set on launched Pods.

image_pull_secrets (Union[List[strict dict], None], optional)

Specifies that Kubernetes should get the credentials from the Secrets named in this list.

service_account_name (Union[dagster.StringSource, None], optional)

The name of the Kubernetes service account under which to run.

env_config_maps (Union[List[dagster.StringSource], None], optional)

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 (Union[List[dagster.StringSource], None], optional)

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

env_vars (Union[List[String], None], optional)

A list of environment variables to inject into the Job. Each can be of the form KEY=VALUE or just KEY (in which case the value will be pulled from the current process). 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

volume_mounts (List[permissive dict], optional)

A list of volume mounts to include in the job’s container. Default: []. See: https://v1-18.docs.kubernetes.io/docs/reference/generated/kubernetes-api/v1.18/#volumemount-v1-core

Default Value: []

volumes (List[permissive dict], optional)

A list of volumes to include in the Job’s Pod. Default: []. For the many possible volume source types that can be included, see: https://v1-18.docs.kubernetes.io/docs/reference/generated/kubernetes-api/v1.18/#volume-v1-core

Default Value: []

labels (permissive dict, optional)

Labels to apply to all created pods. See: https://kubernetes.io/docs/concepts/overview/working-with-objects/labels

resources (Union[strict dict, None], optional)

Compute resource requirements for the container. See: https://kubernetes.io/docs/concepts/configuration/manage-resources-containers/

image (dagster.StringSource)

The image in which to launch the k8s job.

command (List[String], optional)

The command to run in the container within the launched k8s job.

args (List[String], optional)

The args for the command for the container.

namespace (dagster.StringSource, optional)

load_incluster_config (Bool, optional)

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 Value: True

kubeconfig_file (Union[String, None], optional)

The kubeconfig file from which to load config. Defaults to using the default kubeconfig.

Default Value: None

job_config (permissive dict, optional)

Raw k8s config for the v1Job. Keys can either snake_case or camelCase

timeout (Int, optional)

How long to wait for the job to succeed before raising an exception

An op that runs a Kubernetes job using the k8s API.

Contrast with the k8s_job_executor, which runs each Dagster op in a Dagster job in its own k8s job.

This op may be useful when:
  • You need to orchestrate a command that isn’t a Dagster op (or isn’t written in Python)

  • You want to run the rest of a Dagster job using a specific executor, and only a single op in k8s.

For example:

from dagster import job
from dagster_k8s import k8s_job_op

first_op = k8s_job_op.configured(
    {
        "image": "busybox",
        "command": ["/bin/sh", "-c"],
        "args": ["echo HELLO"],
    },
    name="first_op",
)
second_op = k8s_job_op.configured(
    {
        "image": "busybox",
        "command": ["/bin/sh", "-c"],
        "args": ["echo GOODBYE"],
    },
    name="second_op",
)

@job
def full_job():
    second_op(first_op())

The service account that is used to run this job should have the following RBAC permissions:

rules:
  - apiGroups: ["batch"]
      resources: ["jobs", "jobs/status"]
      verbs: ["*"]
  # The empty arg "" corresponds to the core API group
  - apiGroups: [""]
      resources: ["pods", "pods/log", "pods/status"]
      verbs: ["*"]'

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:

  1. Keep a warm kind cluster

  2. 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