If a stream operator crashes, it can restart from its most recent checkpoint and discard any output generated between the last checkpoint and the crash. The checkpoints are triggered by barriers in the message stream, similar to the boundaries between microbatches, but without forcing a particular window size.
