Streaming Was Only Half the Fix: Retrofitting a Stable gRPC API - Sohan Kunkerkar, Red Hat

Why watch

A case study from a CRI-O maintainer explaining how adding streaming RPCs to the Kubernetes Container Runtime Interface solved ResourceExhausted errors caused by unary list operations returning thousands of container objects. The talk covers the mixed-version rollout strategy using UNIMPLEMENTED-based fallback, and reveals why initial streaming implementations still cause high server memory usage by materializing full result sets before sending—leaving HTTP/2 flow control ineffective. For Kubernetes maintainers, runtime developers, and gRPC API designers.