Overlapping sessions with zero allowed lateness due to window expiry rules #18677
Description
Consider this sequence, with session gap durations of 5:
- element arrives with timestamp 0, assigned to proto-window [0, 5)
- watermark advances to 6, emitting the session and discarding it
- element arrives with timestamp 3, assigned to proto-window [3, 8) so it is not dropped as the window is not expired
- watermark advances to 8****, emitting that session
While "technically correct" according to spec, this seems undesirable. It was introduced when late data dropping was tied to window expiry. I think either dropping the second element or including it and emitting a merged window would be OK.
In the case of sessions, we could just retain the window until it cannot possibly merge with other non-expired data. Even with allowed lateness zero this is double the gap duration. The window would be in an interesting state where it would be expired and ineligible for further output but could still merge and the greater window could be output.
The challenge is that sessions are just one kind of merging window - the merging logic has to be assumed opaque. So we cannot simply reason about how sessions work. The other, more drastic option, is to rethink how late data dropping is defined for merging windows, particularly in the "proto-window" phase.
Imported from Jira BEAM-3568. Original Jira may contain additional context.
Reported by: kenn.