Checkpoint barrier
A checkpoint barrier is the marker a job floats through its graph so every operator snapshots the same logical instant.
Also called: barrier alignment
Checkpoint barrier - the marker that flows through a job's graph so every operator snapshots the same logical instant: the checkpoint barrier Flink injects at each source travels down the same channels as the data, and each operator snapshots its state when the barrier reaches it, so the snapshot covers exactly the records before the barrier and none after.
The subtlety is alignment. An operator with two inputs must hold the snapshot until barrier n has arrived on both, pausing the faster channel in the meantime, and under backpressure that pause can stretch a checkpoint from seconds to minutes; Flink grew unaligned checkpoints specifically to soften it. To be fair, it is an elegant piece of distributed-systems engineering, descended from Chandy and Lamport's 1985 snapshot algorithm.
Do you need it? Only if your state is spread across machines that must agree on an instant. The barrier solves a coordination problem the cluster itself created. In one process there is nothing to align: ubik's checkpoint is one atomic file, state and offsets captured together, no marker in flight anywhere. The algorithm stays in the paper.