🛢️ Data Engineering · Streaming & Modeling
Choose windowing and watermarks correctly
Event-time vs processing-time decisions, late-data handling, and exactly-once semantics.
advanced~40 minData EngineersAnalytics EngineersPlatform Data Teams
Steps
- 1Timestamp on event time extracted from payload, not arrival
- 2Set watermark delay from measured lateness distribution p99, not guesswork
- 3Decide allowed-lateness policy: update vs side-output for stragglers
- 4Key by cardinality you can hold state for; watch idle-key eviction
- 5Verify exactly-once end-to-end through failure injection tests
- 6Monitor watermark lag as the primary stream health metric
Common Pitfalls
- ▲Processing-time windows rewriting history on replays
- ▲Unbounded state growth from high-cardinality keys
Commands
Install with skills CLI
$ npx skills add aniruddhaadak80/skills --skill streaming-modeling-stream-processing-windowsInstall globally
$ npx skills add aniruddhaadak80/skills --skill streaming-modeling-stream-processing-windows -gTags
#streaming#kafka#flink#data-engineering#streaming-modeling