Define a Ray cluster
This page provides an overview of how you define a Ray cluster on Anyscale. When you launch a workspace, job, or service, Anyscale deploys a Ray cluster using cloud resources according to your cluster definition. Because workspaces, jobs, and services rely on the same basic constructs for defining clusters, you can move between stages in your development lifecycle by updating a handful of configurations.
The following diagram highlights the core concepts in this model:
How to define a cluster on Anyscale
A cluster definition consists of the following two components:
- A container image specifies the compute environment deployed to each node. It includes package requirements, dependencies, and environmental variables.
- A compute config specifies the number and type of Ray nodes to deploy, as well as instructions about scaling and various cloud-specific configuration options.
The Anyscale console provides user interfaces for the following:
- Creating custom container images.
- Defining custom compute configs.
- Configuring and managing workspaces.
- Monitoring and terminating jobs and services.
The CLI and SDK provide commands for defining, managing, and deploying all the configurations related to workspaces, jobs, and services.
You use the CLI or SDK to define clusters when submitting jobs and deploying services.
Workspaces, jobs, and services each have additional options and parameters to further customize clusters for an application.
Container image
The container image defines the standardized package of dependencies and code that Anyscale deploys as containerized Ray nodes on in a cluster.
Anyscale provides a selection of base images for all Ray versions 2.8.0 and above. Anyscale recommends starting development of new workloads using base images with newer versions of Ray and Python. To review available images and installed libraries, see Anyscale base images.
If you're using Anyscale workspaces for interactive development, you can iterate on the environment in your container image using workspace dependency features.
If you're submitting a job or deploying a service, Anyscale recommends building and using a custom image. Custom images use standard Dockerfile syntax to allow you to extend Anyscale-provided base images to include your own environmental variables, package dependencies, and other configurations. See Build your own image.
Compute configuration
The compute config defines the number and type of nodes that Anyscale adds to the Ray cluster. You use the compute config to specify settings such as the following:
- The size of the head node.
- The number of CPUs and GPUs in the cluster.
- The size and type of worker nodes.
- Whether to use spot instances.
- The minimum and maximum number of nodes for scaling.
Compute configs are specific to an Anyscale cloud. The Anyscale cloud determines where the Ray cluster deploys (for example, AWS, Google Cloud, or Kubernetes), the default permissions assumed by the cluster in that ecosystem, and defaults for other settings such as networking and storage. You can optionally use the compute config to specify advanced configurations including custom networking or IAM roles.
The console only shows compute configs for the active Anyscale cloud. You can use the console to define a new compute config for the active cloud. If you're using the CLI or SDK, you can either specify an Anyscale cloud in your compute config or name a previously defined compute config in a specified cloud.
The Anyscale console provides human-friendly names for node configuration when working with clouds hosted on AWS or Google Cloud. These map back to virtual machine instances.
For example, the 8CPU-32GB instance type maps to a m5.2xlarge machine on AWS or a n2-standard-8 machine on Google Cloud.
Anyscale clouds deployed on Kubernetes use pods instead of virtual machines as worker nodes.
You can use the same compute config for multiple jobs, services, and workspaces. To learn more about compute configs, see Compute configurations.
Cluster storage
Workspaces, jobs, and services use the same basic model for cluster storage. Anyscale cloud admins configure the default storage permissions and configurations used when deploying clusters.
Anyscale configures and deploys the following storage assets each time a cluster starts:
- Ephemeral block storage local to each machine in your cluster.
- Ephemeral storage for sharing files across nodes in the cluster.
- Default object storage configured by Anyscale for persisting files between workloads.
- Shared storage available to all users in a cloud.
- Shared storage available to all clusters launched by a user.
To learn more about these different types of storage, see Storage overview.
Grant access to cloud provider resources
Anyscale recommends configuring access to resources in your cloud provider to add persistent object storage and secret management to your cluster, as shown in the following table.
You must have sufficient privileges in your cloud provider account to configure the required permissions.
| AWS | Google Cloud | |
|---|---|---|
| Object storage | S3 | GCS |
| Secret management | AWS Secrets Manager | Secret Manager |