Skip to content

danielgafni/dagster-ray

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

39 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

dagster-ray

image image image CI pre-commit Checked with pyright Ruff


Ray integration for Dagster.

dagster-ray allows orchestrating distributed Ray compute from Dagster pipelines. It includes:

  • RayRunLauncher - a RunLauncher which submits Dagster runs as isolated Ray jobs (in cluster mode) to a Ray cluster.

  • ray_executor - an Executor which submits individual Dagster steps as isolated Ray jobs (in cluster mode) to a Ray cluster.

  • RayIOManager - an IOManager which stores intermediate values in Ray's object store. Ideal in conjunction with RayRunLauncher and ray_executor.

  • PipesRayjobClient, a Dagster Pipes client for launching and monitoring Ray jobs on remote clusters. Typically used with external Pythons scripts. Allows receiving rich logs, events and metadata from the job. Doesn't handle cluster management, can be used with any Ray cluster.

  • PipesKubeRayJobClient, a Dagster Pipes client for launching and monitoring KubeRay's RayJob CR on Kubernetes. Typically used with external Pythons scripts. Allows receiving rich logs, events and metadata from the job.

  • RayResource, a resource representing a Ray cluster. Interactions with Ray are performed in client mode (requires stable persistent connection), so it's most suitable for relatively short-lived jobs. It has implementations for KubeRay and local (mostly for testing) backends. dagster_ray.RayResource defines the common interface shared by all backends and can be used for backend-agnostic type annotations.

  • Miscellaneous utilities like @op, @job and @schedule for managing KubeRay clusters

dagster-ray is tested across multiple versions of Python, Ray, Dagster, and KubeRay Operator. It integrates with Dagster+ where possible.

Documentation can be found below.

Note

This project is in early development. APIs are unstable and can change at any time. Contributions are very welcome! See the Development section below.

Feature Matrix

There are different options available for running Dagster code on Ray. The following table summarizes the features of each option:

Feature RayRunLauncher ray_executor PipesRayJobClient PipesKubeRayJobClient KubeRayCluster
Creates Ray cluster ❌ ❌ ❌ ✅ ✅
Submits jobs in cluster mode ✅ ✅ ✅ ✅ ❌
For long-running jobs ✅ ✅ ✅ ✅ ❌
Enabled per-asset ❌ ❌ ✅ ✅ ✅
Configurable per-asset ❌ ✅ ✅ ✅ ✅
Doesn't need an external script ✅ ✅ ❌ ❌ ✅
Ray cluster doesn't need access to Dagster's metadata DB and logs storage ❌ ❌ ✅ ✅ ✅

Examples

See the examples directory.

Installation

pip install dagster-ray

To install with extra dependencies for a particular backend (like kuberay), run:

pip install 'dagster-ray[kuberay]'

Features

RunLauncher

Warning

The RayRunLauncher is a work in progress

pip install dagster-ray[run_launcher]

The RayRunLauncher can be configured via dagster.yaml:

run_launcher:
  module: dagster_ray
  class: RayRunLauncher
  config:
    ray:
      num_cpus: 1
      num_gpus: 0

Individual Runs can override Ray configuration:

from dagster import job


@job(
    tags={
        "dagster-ray/config": {
            "num_cpus": 16,
            "num_gpus": 1,
        }
    }
)
def my_job():
    return my_op()

Executor

Warning

The ray_executor is a work in progress

pip install dagster-ray[executor]

The ray_executor can be used to execute Dagster steps on an existing remote Ray cluster. The executor submits steps as Ray jobs. They are started directly in the Ray cluster. Example:

from dagster import job, op
from dagster_ray import ray_executor


@op(
    tags={
        "dagster-ray/config": {
            "num_cpus": 8,
            "num_gpus": 2,
            "runtime_env": {"pip": {"packages": ["torch"]}},
        }
    }
)
def my_op():
    import torch

    # your expensive computation here

    result = ...

    return result


@job(executor_def=ray_executor.configured({"ray": {"num_cpus": 1}}))
def my_job():
    return my_op()

Fields in the dagster-ray/config tag override corresponding fields in the Executor config.

IOManager

RayIOManager allows storing and retrieving intermediate values in Ray's object store. Most useful in conjunction with RayRunLauncher and ray_executor.

It works by storing Dagster step keys in a global Ray actor. This actor contains a mapping between step keys and Ray ObjectRefs. It can be used with any pickable Python objects.

from dagster import asset, Definitions
from dagster_ray import RayIOManager


@asset(io_manager_key="ray_io_manager")
def upstream() -> int:
    return 42


@asset
def downstream(upstream: int):
    return 0


definitions = Definitions(
    assets=[upstream, downstream], resources={"ray_io_manager": RayIOManager()}
)

It supports partitioned assets.

from dagster import (
    asset,
    Definitions,
    StaticPartitionsDefinition,
    AssetExecutionContext,
)
from dagster_ray import RayIOManager


partitions_def = StaticPartitionsDefinition(["a", "b", "c"])


@asset(io_manager_key="ray_io_manager", partitions_def=partitions_def)
def upstream(context: AssetExecutionContext):
    return context.partition_key


@asset(partitions_def=partitions_def)
def downstream(context: AssetExecutionContext, upstream: str) -> None:
    assert context.partition_key == upstream

It supports partition mappings. When loading multiple upstream partitions, they should be annotated with a Dict[str, ...], dict[str, ...], or Mapping[str, ...] type hint.

from dagster import (
    asset,
    Definitions,
    StaticPartitionsDefinition,
    AssetExecutionContext,
)
from dagster_ray import RayIOManager


partitions_def = StaticPartitionsDefinition(["A", "B", "C"])


@asset(io_manager_key="ray_io_manager", partitions_def=partitions_def)
def upstream(context: AssetExecutionContext):
    return context.partition_key.lower()


@asset
def downstream_unpartitioned(upstream: Dict[str, str]) -> None:
    assert upstream == {"A": "a", "B": "b", "C": "c"}

Pipes

PipesRayJobClient

A general-purpose Ray job client that can be used to submit Ray jobs and receive logs and Dagster events from them. It doesn't manage the cluser lifecycle and can be used with any Ray cluster.

Examples:

In Dagster code, import PipesRayJobClient and invoke it inside an @op or an @asset:

from dagster import AssetExecutionContext, Definitions, asset

from dagster_ray import PipesRayJobClient
from ray.job_submission import JobSubmissionClient


@asset
def my_asset(context: AssetExecutionContext, pipes_ray_job_client: PipesRayJobClient):
    pipes_ray_job_client.run(
        context=context,
        submit_job_params={
            "entrypoint": "python /app/my_script.py",
        },
        extra={"param": "value"},
    )


definitions = Definitions(
    resources={"pipes_ray_job_client": PipesRayJobClient(client=JobSubmissionClient())},
    assets=[my_asset],
)

In the Ray job, import dagster_pipes (must be provided as a dependency) and emit regular Dagster events such as logs or asset materializations:

from dagster_pipes import open_dagster_pipes


with open_dagster_pipes() as context:
    assert context.get_extra("param") == "value"
    context.log.info("Hello from Ray Pipes!")
    context.report_asset_materialization(
        metadata={"some_metric": {"raw_value": 57, "type": "int"}},
        data_version="alpha",
    )

A convenient way to provide dagster-pipes to the Ray job is runtime_env:

submit_job_params = {
    "entrypoint": "python /app/my_script.py",
    "runtime_env": {"pip": ["dagster-pipes"]},
}

Events emitted by the Ray job will be captured by PipesRayJobClient and will become available in the Dagster event log. Standard output and standard error streams will be forwarded to the standard output of the Dagster process.

Resources

LocalRay

A dummy resource which is useful for testing and development. It doesn't do anything, but provides the same interface as the other *Ray resources.

Examples:

Using the LocalRay resource

from dagster import asset, Definitions
from dagster_ray import LocalRay, RayResource
import ray


@asset
def my_asset(
    ray_cluster: RayResource,  # RayResource is only used as a type annotation
):  # this type annotation only defines the interface
    return ray.get(ray.put(42))


definitions = Definitions(resources={"ray_cluster": LocalRay()}, assets=[my_asset])

Conditionally using the LocalRay resource in development and KubeRayCluster in production:

from dagster import asset, Definitions
from dagster_ray import LocalRay, RayResource
from dagster_ray.kuberay import KubeRayCluster
import ray


@asset
def my_asset(
    ray_cluster: RayResource,  # RayResource is only used as a type annotation
):  # this type annotation only defines the interface
    return ray.get(ray.put(42))


IN_K8s = ...


definitions = Definitions(
    resources={"ray_cluster": KubeRayCluster() if IN_K8s else LocalRay()},
    assets=[my_asset],
)

KubeRay

pip install dagster-ray[kuberay]

This backend requires a Kubernetes cluster with KubeRay Operator installed.

Integrates with Dagster+ by injecting environment variables such as DAGSTER_CLOUD_DEPLOYMENT_NAME and tags such as dagster/user into default configuration values and Kubernetes labels.

To run ray code in client mode (from the Dagster Python process directly), use the KubeRayClient resource (see the KubeRayCluster section). To run ray code in job mode, use the PipesKubeRayJobClient with Dagster Pipes (see the Pipes section).

The public objects can be imported from dagster_ray.kuberay module.

Pipes

PipesKubeRayJobClient

dagster-ray provides the PipesKubeRayJobClient which can be used to execute remote Ray jobs on Kubernetes and receive Dagster events and logs from them. RayJob will manage the lifecycle of the underlying RayCluster, which will be cleaned up after the specified entrypoint exits. Doesn't require a persistent connection to the Ray cluster.

Examples:

In Dagster code, import PipesKubeRayJobClient and invoke it inside an @op or an @asset:

from dagster import AssetExecutionContext, Definitions, asset

from dagster_ray.kuberay import PipesKubeRayJobClient


@asset
def my_asset(
    context: AssetExecutionContext, pipes_kube_rayjob_client: PipesKubeRayJobClient
):
    pipes_kube_rayjob_client.run(
        context=context,
        ray_job={
            # RayJob manifest goes here
            # full reference: https://ray-project.github.io/kuberay/reference/api/#rayjob
            "metadata": {
                # .metadata.name is not required and will be generated if not provided
                "namespace": "ray"
            },
            "spec": {
                "entrypoint": "python /app/my_script.py",
                # *.container.image is not required and will be set to the current `dagster/image` tag if not provided
                "rayClusterSpec": {
                    "headGroupSpec": {...},
                    "workerGroupSpecs": [...],
                },
            },
        },
        extra={"param": "value"},
    )


definitions = Definitions(
    resources={"pipes_kube_rayjob_client": PipesKubeRayJobClient()}, assets=[my_asset]
)

In the Ray job, import dagster_pipes (must be provided as a dependency) and emit regular Dagster events such as logs or asset materializations:

from dagster_pipes import open_dagster_pipes


with open_dagster_pipes() as context:
    assert context.get_extra("param") == "value"
    context.log.info("Hello from Ray Pipes!")
    context.report_asset_materialization(
        metadata={"some_metric": {"raw_value": 57, "type": "int"}},
        data_version="alpha",
    )

A convenient way to provide dagster-pipes to the Ray job is with runtimeEnvYaml field:

import yaml

ray_job = {"spec": {"runtimeEnvYaml": yaml.safe_dump({"pip": ["dagster-pipes"]})}}

Events emitted by the Ray job will be captured by PipesKubeRayJobClient and will become available in the Dagster event log. Standard output and standard error streams will be forwarded to the standard output of the Dagster process.

Running locally

When running locally, the port_forward option has to be set to True in the PipesKubeRayJobClient resource in order to interact with the Ray job. For convenience, it can be set automatically with:

from dagster_ray.kuberay.configs import in_k8s

pipes_kube_rayjob_client = PipesKubeRayJobClient(..., port_forward=not in_k8s)

Resources

KubeRayCluster

KubeRayCluster can be used for running Ray computations on Kubernetes in client (interactive) mode. Requires stable persistent connection through the duration of the Dagster step.

When added as resource dependency to an @op/@asset, the KubeRayCluster:

  • Starts a dedicated RayCluster for it
  • Connects to the cluster in client mode with ray.init() (unless skip_init is set to True)
  • Tears down the cluster after the step is executed (unless skip_cleanup is set to True)

RayCluster comes with minimal default configuration, matching KubeRay defaults.

Examples:

Basic usage (will create a single-node, non-scaling RayCluster):

from dagster import asset, Definitions
from dagster_ray import RayResource
from dagster_ray.kuberay import KubeRayCluster
import ray


@asset
def my_asset(
    ray_cluster: RayResource,  # RayResource is a backend-agnostic type annotation
):
    return ray.get(ray.put(42))  # interact with the Ray cluster!


definitions = Definitions(
    resources={"ray_cluster": KubeRayCluster()}, assets=[my_asset]
)

Larger cluster with auto-scaling enabled:

from dagster_ray.kuberay import KubeRayCluster, RayClusterConfig

ray_cluster = KubeRayCluster(
    ray_cluster=RayClusterConfig(
        enable_in_tree_autoscaling=True,
        worker_group_specs=[
            {
                "groupName": "workers",
                "replicas": 2,
                "minReplicas": 1,
                "maxReplicas": 10,
                # ...
            }
        ],
    )
)

KubeRayClient

This resource can be used to interact with the Kubernetes API Server.

Examples:

Listing currently running RayClusters:

from dagster import op, Definitions
from dagster_ray.kuberay import KubeRayClient


@op
def list_ray_clusters(
    kube_ray_client: KubeRayClient,
):
    return kube_ray_client.client.list(namespace="kuberay")

Jobs

delete_kuberay_clusters

This job can be used to delete RayClusters from a given list of names.

cleanup_old_ray_clusters

This job can be used to delete old RayClusters which no longer correspond to any active Dagster Runs. They may be left behind if the automatic cluster cleanup was disabled or failed.

Schedules

Cleanup schedules can be trivially created using the cleanup_old_ray_clusters or delete_kuberay_clusters jobs.

cleanup_old_ray_clusters

dagster-ray provides an example daily cleanup schedule.

Development

uv sync --all-extras
uv run pre-commit install

Testing

uv run pytest

KubeRay

Required tools:

  • docker
  • kubectl
  • helm
  • minikube

Running pytest will automatically:

  • build an image with the local dagster-ray code
  • start a minikube Kubernetes cluster
  • load the built dagster-ray and loaded kuberay-operator images into the cluster
  • install KubeRay Operator into the cluster with helm
  • run the tests

Thus, no manual setup is required, just the presence of the tools listed above. This makes testing a breeze!

Note

Specifying a comma-separated list of KubeRay Operator versions in the PYTEST_KUBERAY_VERSIONS environment variable will spawn a new test for each version.

Note

it may take a while to download minikube and kuberay-operator images and build the local dagster-ray image during the first tests invocation