Ray
A distributed computing framework designed to address the complexity of deploying machine learning workloads at scale on Kubernetes clusters, commonly used in MLOps and ML platform engineering.
Core Facts
Timeline (last 90 days)
落地时需注意：镜像版本与 Ray 版本需匹配；节点间网络带宽是常被低估的瓶颈；多节点场景下可观测性应在部署阶段就纳入考量。
RLlib is provided by the Ray framework, focuses on distributed large-scale training, and is suited for scenarios requiring parallel simulation.
Ray Serve is built on top of the distributed computing framework Ray and is suited for complex service scenarios that require flexible orchestration of multiple models and business logic
Ray provides a Global Control Store (GCS) and a distributed object store (Plasma), making cross-node data sharing and task orchestration overhead extremely low
Ray supports native handling of PyTorch tensors without requiring serialization conversion
Ray Serve is a production-oriented model serving capability that transforms training results into inference services
Unlike frameworks such as Spark and Dask that are oriented toward batch data processing, Ray's design goal is general-purpose distributed computing, and it excels at handling AI workloads that require dynamic scheduling, state maintenance, and low-latency communication
Ray Tune is the hyperparameter tuning framework provided by Ray
Ray was incubated by UC Berkeley's RISELab and open-sourced in 2017
Ray's core abstraction is to wrap Python functions and classes as Tasks and Actors respectively, coordinating their execution through a distributed scheduler
2 more timeline events
All Facts (13)
Ray, as a distributed computing framework, spans multiple layers, including Ray Tune (hyperparameter search), Ray Train (distributed training), and Ray Serve (online inference).
65%Unverified落地时需注意：镜像版本与 Ray 版本需匹配；节点间网络带宽是常被低估的瓶颈；多节点场景下可观测性应在部署阶段就纳入考量。
50%UnverifiedRLlib is provided by the Ray framework, focuses on distributed large-scale training, and is suited for scenarios requiring parallel simulation.
50%UnverifiedRay Serve is built on top of the distributed computing framework Ray and is suited for complex service scenarios that require flexible orchestration of multiple models and business logic
50%UnverifiedRay's core abstraction is to wrap Python functions and classes as Tasks and Actors respectively, coordinating their execution through a distributed scheduler
50%UnverifiedRay provides a Global Control Store (GCS) and a distributed object store (Plasma), making cross-node data sharing and task orchestration overhead extremely low
50%UnverifiedRay supports native handling of PyTorch tensors without requiring serialization conversion
50%UnverifiedRay Tune is the hyperparameter tuning framework provided by Ray
50%UnverifiedRay Serve is a production-oriented model serving capability that transforms training results into inference services
50%UnverifiedUnlike frameworks such as Spark and Dask that are oriented toward batch data processing, Ray's design goal is general-purpose distributed computing, and it excels at handling AI workloads that require dynamic scheduling, state maintenance, and low-latency communication
50%UnverifiedRay was incubated by UC Berkeley's RISELab and open-sourced in 2017
50%UnverifiedvLLM uses ZeroMQ (zmq) as the RPC framework in single-machine scenarios and leverages the Ray framework to manage multi-node workers in distributed scenarios.
50%UnverifiedOpen-source projects led by Berkeley include the Caffe deep learning framework, the Ray distributed computing framework, and the large language model projects Vicuna and LMSYS Chatbot Arena.
50%