sql-machine-learning/elasticdl

Leverages Kubernetes API calls to dynamically manage worker and parameter server pods, enabling fault-tolerant training where job failures don't interrupt execution—workers can be preempted or killed without requiring checkpoint recovery. Supports TensorFlow (Estimator and Keras) and PyTorch with a minimal CLI interface that automatically distributes training across clusters while integrating with Kubernetes priority-based preemption for elastic resource scheduling.

746 stars. No commits in the last 6 months.