Window assignment is a rule over time and identity

A fixed time window includes its start and excludes its end. For zero-aligned tumbling windows of width W, timestamp t belongs to the interval beginning at floor(t/W)W. Timestamp 10 therefore belongs to [10,20), not [0,10), when W=10 seconds. Offsets and time zones can change alignment, so state those choices before calculating. Discrete timestamp implementations use a largest included timestamp below the exclusive end; do not confuse an event's membership with the exact implementation rule that fires a timer or cleans up retained state.

Sliding windows have a width and a slide. If width exceeds slide, one event can contribute to several overlapping windows, increasing state and output work. Membership means a start s on the chosen alignment grid satisfies s≤t<s+W. Count the matching starts rather than dividing timestamp by width and assuming one answer. A downstream consumer also needs window identity if it will combine, replace, or audit outputs. A bare aggregate value does not announce whether it belongs to one window, an updated pane, or a different key.

A session groups activity separated by less than the configured inactivity gap under its assigner and merge rules. In a simple event-time construction, each event proposes an interval extending from its timestamp through a gap, and overlapping proposed intervals merge. A late event can bridge two earlier groups, so sessions are not fixed bins whose identities can always be finalized on first output. The exact equality-at-gap behavior follows the implementation's interval and merge rules; choose examples with strict overlap when the lesson does not need that boundary detail.

For a session aggregate, a merge must combine the relevant state and coordinate triggers and output corrections. Taking the sum of two already emitted counts and then adding a bridging event can work only if the downstream result protocol retracts or replaces the older results consistently. Otherwise the same activity can appear in old sessions and the new merged session. Flink 2.1 distinguishes window assignment, triggers, and allowed lateness; custom behavior needs compatibility with merging windows. Do not assume that a trigger appropriate for nonmerging tumbling windows automatically has a valid merge contract.

An event-time trigger can emit an initial answer and later emit a correction while state remains retained. Processing-time triggers can offer earlier provisional answers without proving event-time completeness. Allowed lateness affects how long eligible late events can modify a window, while cleanup determines retained resources. Describe the output mode—replacement, retraction, or accumulated delta—so a sink does not count every firing as a new independent business result. Test boundary timestamps and bridging arrivals in different orders, and inspect final keyed results rather than only the number of output messages.

The modular lab illustrates why repeating phase is not full time identity. Reset, explicitly set BOTH m=5 and a=1, predict residues zero through four, then five Steps: n=floor(2t) reaches five and the residue returns to zero. That does not make event-time windows [0,5) and [5,10) the same window. Edits retain n through elapsed time; Reset is needed for a fresh orbit. This finite additive cycle has no timestamps, window assignment, sessions, triggers, or lateness. Use it only to separate phase within a cycle from the unbounded count of completed cycles.