Principle:Ray project Ray Actor Instance Creation
| Knowledge Sources | |
|---|---|
| Domains | Distributed_Computing, Actor_Model |
| Last Updated | 2026-02-13 17:00 GMT |
Overview
A mechanism for remotely instantiating a stateful actor on a cluster worker node, returning a serializable handle for subsequent method invocations.
Description
Actor Instance Creation is the process of scheduling a constructor call on a remote worker node, which creates a new process (or assigns a worker) dedicated to hosting the actor. The result is a handle — a lightweight, serializable reference to the remote actor that can be passed between tasks and stored in the object store. The handle enables remote method invocation on the actor.
Actor creation supports rich configuration:
- Resource requirements: CPU, GPU, and custom resource constraints
- Naming: Named actors are registered in the GCS for cross-process lookup
- Restart policies: Automatic restart on failure with configurable limits
- Lifetime: Detached actors survive the creating process; non-detached actors are tied to their creator
- Concurrency: Max concurrency and concurrency groups for async actors
Usage
Use actor creation when you need a long-lived, stateful service running on a cluster node. This is appropriate for model serving, stateful stream processing, in-memory caches, and coordination services.
Theoretical Basis
Actor creation is a special case of task submission where the task is a constructor call and the result is a handle rather than a value. The scheduler treats actor creation as a resource reservation — the worker hosting the actor retains the allocated resources for the actor's lifetime.