An Apache Flink and Kotlin streaming architecture where online machine learning detects concept drift from machinery wear, controlled by a shadow mode router.

Apache Flink, Kafka and Valkey turn a contextual bandit recommender into an event driven service with stateful training and low latency inference.

Flink Table API in Kotlin states the supplier statistics as a declarative windowed aggregation over a DataStream, with late rows routed by hand.

Flink DataStream API in Kotlin computes the same supplier statistics, using watermarks for event time and side outputs to collect late order events.

Kafka Streams in Kotlin aggregates Avro order events into tumbling window supplier statistics and handles late records with a custom extractor.

Avro and Schema Registry replace hand written JSON codecs in a Kotlin Kafka producer and consumer, with generated classes and graceful shutdown.

A Kotlin Kafka producer and consumer for order events, with custom JSON serialisers, admin helpers and a Gradle build, run on odctl.

Pacing messages are added to the Beam sport activity tracker by comparing short term speed metrics against their long term counterparts.

Late droppable elements are detected by a timer in a stateful DoFn and sent to a Beam side output instead of being discarded without notice.

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.