Skip to main content

Introduction to Anyscale clouds

This page provides an overview of Anyscale clouds and how to configure cloud resources for different environments.

What is an Anyscale cloud?

An Anyscale cloud abstracts the resources and infrastructure necessary for managing Anyscale clusters. An Anyscale cloud is a logical entity that connects the Anyscale control plane with cloud resources in your AWS account, Google Cloud project, or Kubernetes cluster. When you configure cloud resources for an Anyscale cloud, you create a trust relationship between the Anyscale control plane and resources in your cloud provider environment.

note

Cloud resources include both the infrastructure you configure during initial setup (VPCs, IAM roles, storage) and additional resources Anyscale deploys on your behalf (MemoryDB for fault tolerance, EC2 instances for Ray clusters). All of these resources reside in your cloud provider account as part of the data plane.

Anyscale clouds sit between your Anyscale organization and projects in the platform hierarchy. Your organization can contain one or many Anyscale clouds. Organization admins and cloud owners can manage which users can access an Anyscale cloud. See Anyscale organization overview and Cloud roles.

An Anyscale cloud performs the following functions:

  • Maintains information on where and how to start clusters.
  • Deploys Ray clusters in your AWS, Google Cloud, or Kubernetes environments.
  • Defines the collection of resources necessary to manage Anyscale clusters.
  • Provides isolated environments of resources for users within your organization.

When you sign up, Anyscale provisions a serverless Anyscale cloud (also called an Anyscale-hosted cloud) for you to start running workloads. Anyscale recommends serverless clouds mostly for product trials and platform evaluation, as most users require access to additional customizations or features only present in self-hosted clouds.

Types of cloud resource configurations

You configure an Anyscale cloud to deploy Ray clusters using either virtual machines or Kubernetes. The following table shows the options for each configuration:

Compute stackConfiguration options
Virtual machines AWS
Google Cloud
Serverless Anyscale clouds (Anyscale-hosted clouds)
Kubernetes Amazon Elastic Kubernetes Service (EKS)
Google Kubernetes Engine (GKE)
Azure Kubernetes Service (AKS)
Any Kubernetes cluster with configured network access, including on-prem

Anyscale supports the following options for configuring cloud resources:

OptionDescription
Anyscale-managed resource configurationUse anyscale cloud setup to let Anyscale configure resources and permissions in your AWS or Google Cloud account. Only available for clouds that use virtual machines.
Customer-defined resource configurationUse anyscale cloud register to create a trust relationship between Anyscale and resources you've configured in your cloud environment or Kubernetes cluster. Available for AWS and Google Cloud configurations, and required for deploying Anyscale on Kubernetes.

How should you configure your Anyscale cloud?

You should choose the configuration type that meets the needs of your organization. The following table describes the different configuration types in order of increasing complexity:

Configuration typeDetails
Serverless Anyscale (Anyscale-hosted) cloudDefault for new organizations. Provides a fully managed cloud experience. Good for initial product exploration and for customers who prefer not to manage any cloud configurations.
Anyscale-managed resource configurationA simpler way to configure AWS or Google Cloud resources for Anyscale. Use this method to let Anyscale automatically configure the necessary resources in your cloud provider account.
Customer-defined resource configuration on AWS or Google CloudRecommended for most use cases, unless you prefer to work on Kubernetes. This provides the greatest customization and includes support for custom networking.
Kubernetes deploymentUse Kubernetes when you have established Kubernetes infrastructure that you want to leverage for Ray workloads on Anyscale. When deploying Anyscale on Kubernetes, you install the Anyscale operator in your cluster. When possible, use first-party offerings from cloud providers such as GKE, AKS, and EKS for the best Anyscale product experience.
important

Deploying Anyscale on Kubernetes requires custom configuration for your environment. Contact Anyscale support for assistance deploying the Anyscale operator on Kubernetes.

Serverless clouds

When you create a new Anyscale organization, Anyscale provisions a serverless cloud (sometimes called an Anyscale-hosted cloud) in your organization.

In this model, Anyscale manages all of the underlying infrastructure and storage used in your workloads.

Anyscale recommends serverless clouds primarily for users that are familiarizing themselves with the platform or trying out Ray for the first time.

Generally speaking, you can interact with serverless clouds the same way you would with self-hosted clouds, just without the need to worry about configuring underlying cloud infrastructure.

Some users might prefer to use serverless clouds to eliminate cloud configuration complexities. Review the Anyscale pricing page to estimate costs for your workloads.

Contact Anyscale support if you're having trouble with a serverless Anyscale cloud or need access to the following features on a serverless cloud:

  • GPU compute instances.
  • A serverless cloud running on Google Cloud.
  • Custom cloud storage for data retention.
important

Anyscale provides temporary object storage as scratch storage on Anyscale-hosted clouds. $ANYSCALE_ARTIFACT_STORAGE retains data for up to 90 days from the date of creation. Anyscale removes all user data within a maximum of 120 days following account deactivation. Anyscale-hosted clouds may retain logs for up to 2 years. Users can request deletion of their Anyscale-hosted cloud account by contacting Anyscale support.