What is Anyscale?
Anyscale is a unified AI platform from the creators of Ray. Anyscale adds optimizations, observability, data governance, and developer tooling that make it the best place to run Ray workloads. This page provides an overview of how Anyscale can accelerate and simplify getting AI and ML workloads into production at scale.
Anyscale uses a consumption-based model to manage Ray clusters configured with the RayTurbo optimized compute engine. Anyscale also offers additional paid professional services including customized training and dedicated support. See Anyscale pricing.
If you're new to Ray, see What is Ray?.
Deploy and manage Ray clusters
Anyscale helps you launch and manage Ray clusters of any size using resources in your cloud provider of choice. You deploy an Anyscale cloud to configure access to resources on AWS, Google Cloud, or Kubernetes. See Introduction to Anyscale clouds.
Compute configuration is highly customizable, and allows you to take advantage of your existing infrastructure and compute, including the following features:
- Deploy to on-prem or cloud-hosted Kubernetes clusters.
- Leverage virtual machine reservations in your cloud provider.
- Create clusters that combine CPUs and GPUs of varying sizes.
- Assign priority and preemption rules for Ray workloads, including the ability to schedule components of your workloads to certain worker nodes.
- Auto-scale Ray clusters by worker type, including adding new nodes to Kubernetes cluster or adding on demand virtual machines to your reserved instances.
When you launch a cluster, Anyscale builds and caches an optimized version of your container image. This cached image allows Anyscale to rapidly add new nodes to your cluster when needed. This allows for efficient upscaling, aggressive downscaling, and the ability to use spot instances with fallback.
You use container images and compute configs when configuring clusters for workspaces, jobs, and services. The Anyscale console provides interfaces for defining and managing custom container images and compute configs. See Define a Ray cluster.
Secure and share Ray workloads
Anyscale allows you to manage, observe, and share Ray workloads deployed across multiple cloud providers on VMs or Kubernetes from a single organization. Admins can restrict access to clouds or projects, ensuring that the right users can view, monitor, and run workloads in the correct cloud environments. Use SSO to manage logins and restrict access to users still active in your identity provider of choice. See Anyscale organization overview.
Anyscale uses a shared responsibility model, where you configure identity and access management policies to allow Anyscale to deploy and manage cloud infrastructure in your cloud provider of choice. Your data stays in your cloud storage and compute in the regions of your choice.
Anyscale has certified compliance with SOC 2 Type 2. Learn more about security from the Anyscale Trust Center.
For a complete overview of admin tasks and features on Anyscale, see Guide for admins.
Develop on Anyscale
The Anyscale console includes numerous features for accelerating developer workflows. Many of these features center on Anyscale workspaces for interactive development against Ray clusters.
Workspaces provide a hosted VS Code, Jupyter Lab, and web terminal for a similar developer experience to programming in virtual environments or cloud VMs. You can also connect IDEs such as VS Code and Cursor to your Anyscale workspace to run code directly in a containerized environment from your local machine.
The Anyscale console provides a graphical user interface for writing code, managing files, running applications, configuring dependencies, and monitoring logs and metrics for your workload. See Workspaces.
You can directly launch jobs or services from an Anyscale workspace, accelerating testing on the way to deploying code in production. See Launch jobs and services from an Anyscale workspace.
Because workspaces, jobs, and services use the same basic configurations and options, you can quickly package your code and dependencies and move from interactive development to production. See Develop Anyscale applications.
Anyscale provides a complete CLI and SDK for developers who prefer to develop code locally. See Anyscale API reference.
Run data processing and ML training jobs
Use Anyscale jobs to submit production Ray applications for workloads such as batch inference, model training, and data processing. See Get started with jobs.
You use the CLI or SDK to trigger jobs manually, or you can use the same tools to automate Anyscale jobs using CI/CD or scheduling tools. See Job API Reference.
Use job queues to enable multiple workloads to share compute resources, or use job schedules for workloads that need to run on a fixed cadence.