This article was accepted into the corpus but its outbound wikilinks were never NER-processed — typical at the deepest BFS hop or when the run's entity cap was reached. No expansion funnel to show.
| Ray (distributed execution) | |
|---|---|
| Name | Ray |
| Developer | Ray Project |
| Released | 2016 |
| Programming language | Python (programming language), C++ |
| Operating system | Linux, macOS, Windows |
| License | Apache License |
Ray (distributed execution)
Ray is an open-source distributed execution framework designed for building scalable, high-performance applications for machine learning, reinforcement learning, and data processing. It provides primitives for task-parallel and actor-based computation, integrates with popular projects, and targets use cases spanning research at Berkeley AI Research to production deployments at companies like Uber Technologies, OpenAI, and Huawei. Ray emphasizes low-latency scheduling, fault tolerance, and a unified programming model that interoperates with ecosystems around TensorFlow, PyTorch, Dask (software), and Kubernetes.
Ray originated from research at University of California, Berkeley and was incubated by the Ray Project and contributors across industry and academia, drawing on concepts from MapReduce, Actor model, Apache Spark, and MPI. It exposes primitives that let developers express fine-grained parallelism for workloads encountered in projects such as RLlib, Tune (software), Modin (software), and Serve (Ray). Ray's design aims to bridge the gap between research prototypes and production systems used by organizations including Microsoft, Amazon (company), Intel, and NVIDIA.
Ray's architecture uses a distributed control plane and a scalable worker pool with a lightweight scheduler and object store derived from techniques used in Plasma (object store). Core components include a global control store reminiscent of ZooKeeper, a node-local object store leveraging shared memory and ZeroMQ-style fast messaging, and local schedulers that coordinate with a global scheduler akin to approaches in Mesos and Kubernetes. The runtime supports multi-language frontends, primarily Python (programming language) and C++, and integrates with container orchestration platforms such as Docker and Kubernetes for cluster management and resource isolation.
Ray's programming model revolves around tasks and actors, concepts inspired by the Actor model and distributed task systems like Celery (software). A task is a stateless function executed remotely, while an actor is a stateful worker with methods invoked remotely; both return futures similar to Promise (programming). Ray manages object references in a distributed object store with lineage-based reconstruction for fault tolerance, drawing parallels to Resilient Distributed Dataset semantics from Apache Spark. Resource-aware scheduling allows allocation of GPU and CPU resources per task, and Ray's placement groups implement strategies comparable to bin packing and affinity scheduling used in cluster managers.
Ray exposes high-level APIs for parallel and distributed programming in Python (programming language) and C++. Libraries such as RLlib, Tune (software), Datasets (Ray), and Serve (Ray) provide domain-specific abstractions for reinforcement learning, hyperparameter tuning, data ingestion, and model serving, integrating with frameworks like TensorFlow, PyTorch, and scikit-learn. The API supports asynchronous futures, remote actor handles, and placement groups, enabling patterns found in systems like Dask (software), Ray Tune, and Horovod. Users can combine task and actor abstractions with third-party tools such as Prometheus, Grafana, and Jaeger for observability.
Ray is used for large-scale reinforcement learning projects from groups like DeepMind-adjacent research, hyperparameter optimization in enterprises including LinkedIn and Netflix, and serving machine learning models for companies such as Instacart and Spotify. It supports distributed training workloads comparable to those run on Horovod and Kubeflow pipelines, and its streaming and batch capabilities have been applied in contexts similar to Flink and Storm (software). Research labs and startups leverage Ray for multi-agent simulation, automated machine learning workflows, real-time inference, and data preprocessing at scale.
Ray targets low-latency scheduling and high task throughput, employing techniques akin to those in Nimble and Sparrow (scheduling), including decentralized scheduling and lightweight RPC to reduce overhead. Benchmarks presented by Ray contributors compare favorably with Apache Spark for fine-grained tasks and with Dask (software) on mixed workloads; performance depends on workload characteristics, hardware from vendors like NVIDIA and Intel, and deployment platforms such as Kubernetes or bare-metal clusters. Ray's object store and plasma-inspired shared-memory mechanisms reduce serialization costs for large arrays used in NumPy and Pandas (software) pipelines.
An active ecosystem surrounds Ray, including managed services, third-party libraries, and integrations with TensorFlow, PyTorch, scikit-learn, Hugging Face, and orchestration platforms like Kubernetes and Docker Swarm. The community includes contributors from Anyscale, Amazon (company), Microsoft, Intel, and academic institutions such as MIT and Stanford University. Commercial products and managed offerings by organizations like Anyscale (company) and cloud vendors have accelerated enterprise adoption for production deployments in fintech, adtech, and autonomous systems.
Ray implements lineage-based fault recovery and actor checkpointing patterns similar to strategies in Spark Streaming and Checkpointing (computing). Security features rely on underlying infrastructure controls from Kubernetes, TLS, and cloud IAM services provided by Amazon Web Services, Google Cloud Platform, and Microsoft Azure for authentication and authorization. For multi-tenant or regulated environments, operators integrate Ray with secrets managers like HashiCorp Vault and observability stacks including Prometheus and ELK (software stack) to meet compliance and incident response requirements.
Category:Distributed computing Category:Open-source software