Kafka Streams: Windowed aggregation needs a grace period

Late records arrive after the window closes; grace decides how long you still accept them.

Code

orders.groupByKey()
      .windowedBy(TimeWindows.ofSizeAndGrace(
              Duration.ofMinutes(5), Duration.ofMinutes(1)))
      .count()
      .toStream()
      .to("orders-per-5min");

Output

A record 30s late is counted.
A record 90s late is dropped - and shows up in the
dropped-records metric rather than in your total.

Run this yourself in the Online Java Compiler, spin up a live REST API in the API Sandbox, or practise with Java interview questions.

Published 2026-08-11