[RFC] Terminology Clarification: Worker Node vs Worker Process

Summary

This RFC proposes a change in Ray's terminology to distinguish between "worker nodes" (hardware resources) and "worker processes" (individual processes executing tasks/actors), in an effort to reduce recurring user confusion and improve the overall clarity of Ray's API and documentation.

Proposal

The proposed changes are as follows:

Rename all instances of worker referring to worker nodes to worker_node in the Ray API, VM launcher, KubeRay YAMLs, and other relevant locations. This includes fields such as max_workers and min_workers in the VM launcher YAML, which would be renamed to max_worker_nodes and min_worker_nodes.
Maintain backward compatibility by supporting both worker and worker_node in our interfaces.
Update the documentation to always use "worker node" when referring to hardware resources, and clarify the distinction between "worker node" and "worker".

Alternatives Considered

One alternative would be to introduce a different term, such as "executor", to replace "worker" in the context of worker processes. However, there seems to be a consensus that "worker" is a term commonly used and understood in the machine learning community to refer to individual execution units (processes), as seen in libraries like PyTorch and TensorFlow. Therefore, maintaining the term worker for worker processes and introducing worker_node for hardware resources is the preferred approach.

Status

Comments, feedback, and alternative suggestions are welcomed for this proposal. After 2 weeks, we'll make this change in a backward-compatible manner as part of the Ray 2.6 release.

cc @pcmoritz @richardliaw @GokuMohandas @rkooo567 @gvspraveen @scottsun94

ray-project / ray