ray-project /

Ray

Live map
View repository
Snapshotmaster · 97f7eec

How Python tasks and actors are scheduled across a cluster and exchange immutable objects.

Click a building or route
Choose a path

Submit work, acquire a lease, run it, and publish an immutable result.

System map7 components · 8 connections
130%
Ray architectureThe driver and CoreWorker enter at left; global metadata and per-node scheduling sit in the center; worker execution and the distributed object plane extend to the right.function, arguments, r…task specification and…resources and placemen…worker lease and taskimmutable return objectGlobal ControlServiceTasks andactors APICoreWorkerWorker processRayletRay AIlibrariesPlasmaobject store
Connections
BuildingsEntryServiceComputeStorageRuntimeToolingExternal

Terms, in plain English

ObjectRefA distributed reference to an immutable value that may not exist locally yet.

How to study this repo

  1. 1
    Follow one remote call

    Trace RemoteFunction.remote through CoreWorker submission, raylet leasing, and worker execution.

  2. 2
    Follow one large result

    Connect object creation and sealing in Plasma to a downstream pull and zero-copy read.