All stacks · Compute & processing
Stateful stream processing, watermarks, and exactly-once sinks.
Apache FlinkProcessing time lies after a restart. Watermarks drop late events on purpose. Checkpoints recover; savepoints migrate. Exactly-once to Kafka does not upsert Postgres.
Flink is the stateful stream processor: event time, watermarks, checkpoints, and sinks that can actually claim exactly-once. Spark Structured Streaming is micro-batch with a streaming costume; Flink is the engine you pick when state and lateness are the product.
These notes cover the interview and production questions: watermarks, savepoints vs checkpoints, and when Flink is overkill compared with Kafka Streams or a warehouse Dynamic Table.
Low-latency stateful jobs, large keyed state, and true event-time windows with late data. Spark is fine for micro-batch ETL that happens to read Kafka. If you are explaining watermarks in the design review, you probably want Flink.
The Flink checkpoint barrier plus a transactional or idempotent sink. “Flink is exactly-once” without a sink contract is marketing. Dual-write without a key is still at-least-once in practice.
Not for warehouse-latency SQL. You need Flink when the serving path is a stream, a keyed store, or a sub-minute product SLA the warehouse cannot hit. Most analytics teams do not need Flink first.
One email a week with our latest tutorials. No spam.
Newsletter signup is not live yet. Use the contact form if you want to be notified.