Why it matters
Flink is often the right answer when Spark Streaming isn't fast enough. Understanding when to reach for it matters for streaming design.
Advertisement
The architecture
DataStream API: streams as first-class. Map, filter, join, windowing operators.
State: managed keyed state (per-key values), operator state (per-operator).
Advertisement
How it works end to end
Event time: process by when event occurred, not when arrived. Watermarks handle out-of-order.
Checkpoints: periodic distributed state snapshots. Recovery to consistent state.
State backends: memory, RocksDB (disk-spilling for large state), FsStateBackend.
Deployment: standalone, YARN, Kubernetes.