Processing infinite data streams requires partitioning time into finite buckets to calculate aggregates. While basic tumbling windows suffice for simple monitoring, accurate stream processing depends on handling the temporal disconnect between when an event occurs and when it arrives at the operator. This discrepancy, caused by network latency or distributed system lag, creates the need for precise time attributes and windowing logic.

This chapter examines the mechanisms Flink uses to track time progress and manage state within windows. We focus on the distinction between event time and processing time, and how to maintain data correctness when events arrive out of order. You will analyze the role of watermarks as the primary signal for time progression. A watermark $W(t)$ flows through the stream, asserting that no events with a timestamp $t' < t$ should arrive subsequently.

Beyond standard clock-aligned windows, we will implement dynamic windowing strategies. User behavior often occurs in bursts rather than fixed intervals, making session windows, which group events based on activity gaps, a more accurate model for analytics. We will also configure custom triggers and evictors. These components allow you to define complex conditions for when a window should evaluate its contents or purge data, offering control beyond the default behavior.

Handling late data is a requirement for resilient pipelines. Dropping events that arrive after a watermark allows for lower latency, but often violates strict consistency requirements. We will configure allowed lateness parameters and route late-arriving events to side outputs, ensuring that all data is accounted for even when it falls outside the expected time boundaries. By the end of this section, you will be able to implement windowing logic that balances latency, completeness, and correctness.