Ray

Ray is the distributed compute engine that OpenAI uses to train its models. It scales Python code from your laptop to thousands of GPUs with minimal code changes — distributed training, hyperparameter tuning, model serving, and data processing in one framework.

If you're doing ML at scale — training large models, running distributed inference, or orchestrating complex AI pipelines — Ray is the serious choice. Apache Spark handles distributed data processing but isn't optimized for GPU workloads or low-latency inference. Dask is Python-native and lighter but lacks Ray's actor model and ML libraries. Horovod focuses specifically on distributed training. Commercially, SageMaker and Vertex AI offer managed ML platforms.

Ray's libraries (Ray Train, Ray Serve, Ray Data) cover the full ML lifecycle. The actor model makes stateful distributed computing natural.

The catch: Ray has significant operational complexity. Cluster management, memory tuning, and debugging distributed failures require expertise. The learning curve from "hello world" to "production deployment" is steep. For indie hackers, if your model fits on a single GPU, you don't need Ray — and most models do. This is infrastructure for teams with real scale problems.