The Window That Never FiredKafka · medium · ~12 minThe fraud team's new Flink SQL job computes 5-minute windowed transaction counts per merchant from the payments topic (4 partitions). It deployed cleanly at 09:30. It has now been running for 2 hours.It has emitted zero windows. No errors, no restarts, no consumer lag. Events are visibly flowing — the input topic receives steady traffic. The on-call engineer restarted the job twice ("maybe it's stuck") with no change, and is now proposing to rewrite it in processing time "since event time is broken."You pull the watermark diagnostics before the rewrite happens.Evidence

Watermark diagnostics

Topic + job facts

What the team tried

Checkpoints

Partitions 0–2 have healthy watermarks seconds behind now. The OPERATOR watermark is stuck at negative infinity. Why?

The last event on partition 0 arrived stamped 11:30:07, and the DDL delay is 10 seconds. What is partition 0's current watermark? (Answer as seconds after 11:29:00 — e.g. 11:29:45 would be 45.)

The team sets scan.watermark.idle-timeout to 30 seconds. What happens next?

Write the incident wrap-up: the root cause chain, why the processing-time rewrite would have been the wrong fix, the two config/infra remedies, and the monitoring you'd add so this class of failure pages before a human notices missing dashboards.