FlinkUpdated
Checkpoint tuning
Recommended checkpoint interval, mode and state backend settings for jobs with large state.
Background
Checkpoints are what give Flink exactly-once guarantees. Too frequent and they add backpressure; too sparse and recovery takes longer.
Recommended settings
For jobs with more than 100 GB of state, use the RocksDB state backend with incremental checkpoints and an interval of 3 to 5 minutes.
Monitoring
Watch checkpoint duration, alignment time and state size. If alignment time keeps growing, consider unaligned checkpoints.