A stateful DoFn with Beam state and timers fixes the gRPC batch size and maximum wait time instead of leaving the bundle size to the runner.

Batching gRPC calls in a stateless DoFn so one request covers a whole bundle, cutting the time a Beam Python pipeline spends on enrichment.

Data augmentation in Beam Python by calling a gRPC service once per input element, running on a local Flink cluster with Kafka as the source.