Watermark
An event-time progress estimate indicating that events older than a given timestamp are not normally expected to arrive, allowing a stream processor to close windows and handle late data.
Distributed producers do not necessarily deliver records in the order in which their events occurred. A watermark provides an event-time progress boundary so a processor can decide when window results are ready and how to classify records that arrive after that point.
It is an operational estimate rather than an absolute guarantee. An aggressive watermark can discard useful late data; a very conservative one keeps state and result latency high. Concurrency in distributed data systems provides the broader ordering and time context.