Ray

**Ray Distributed AI Framework** is **a distributed execution engine providing low-latency task scheduling, distributed actors, and object store for efficient machine learning and AI workloads, enabling fine-grained parallelism with minimal overhead** — optimized for dynamic, heterogeneous AI computations. Ray unifies batch, streaming, and serving. **Tasks and Parallelism** @ray.remote decorator designates functions as distributed tasks. task.remote() submits asynchronously, returning ObjectRef (future). ray.get() blocks retrieving result. Fine-grained task submission enables dynamic parallelism without DAG pre-specification. **Actors and Stateful Computation** @ray.remote classes define actors—processes maintaining state. Actors handle multiple method calls sequentially, enabling stateful service. Useful for parameter servers, replay buffers, rollout workers. **Distributed Object Store** Ray's object store enables efficient data sharing: local store on each node, distributed with replication. Objects auto-spilled to external storage (S3, HDFS) if memory insufficient. Zero-copy sharing: tasks on same node access object in local store without serialization. **Scheduling and Locality** scheduler assigns tasks to nodes considering data locality and resource requirements. CPU/GPU resource specification ensures proper placement. Minimizes data movement. **Fault Tolerance** lineage-based recovery: Ray tracks task dependencies, re-executes failed tasks recomputing lost data. Effective for deterministic tasks. **Ray Tune** hyperparameter optimization: automatic distributed hyperparameter search with early stopping, population-based training. **Ray RLlib** reinforcement learning library: distributed training algorithms (A3C, PPO, QMIX). Actors organize rollout workers, training workers, parameter servers. **Ray Serve** serving predictions from trained models. **Ray Data** distributed data processing with lazy evaluation, similar to Spark but Ray-optimized. **Named Actor Handles** actors can be named and retrieved globally, enabling loosely-coupled microservice architectures. **Dynamic Task Graphs** unlike static DAG frameworks (Spark, Dask), Ray supports dynamic task creation—task outcomes determine future tasks. Essential for tree search, early stopping, RL. **Heterogeneous Resources** specify CPU, GPU, memory, custom resources. Scheduler respects constraints. **Applications** include hyperparameter optimization, reinforcement learning training, distributed ML inference, batch RL, parameter sweeps. **Ray's fine-grained scheduling, distributed object store, and dynamic task graphs make it ideal for heterogeneous, resource-intensive AI workloads** compared to traditional batch frameworks.