Apache Flink Best Practices

Core Principles

• Assign stable UIDs to every operator — mandatory for savepoint compatibility
• Use keyed state with TTL — unbounded state is the #1 cause of production failures
• Use event time with watermarks — processing time is non-deterministic on replay
• Use RocksDB state backend with incremental checkpoints for production
• Never do blocking I/O in operators — use AsyncDataStream
• Name every operator for debuggability in the Flink Web UI
• Deploy in Application Mode for production workloads

Application Structure (HIGH)

app-operator-uids

Assign .uid("stable-id") to every operator. This is mandatory.

Without UIDs, Flink cannot map state across job restarts or upgrades from savepoints. Missing UIDs will cause state loss on any job modification.