Streams

A stream is data processed as events arrive, often without a finite end. Begin by clarifying rate, ordering guarantees, out-of-order lateness, required accuracy, and whether the question is algorithmic or systems-oriented. For an online average, maintain count and sum; for a rolling window, keep only events still inside the window and evict old ones; for top-K, maintain a bounded heap. Exact distinct counts may require unbounded memory, so probabilistic structures such as HyperLogLog are a deliberate accuracy tradeoff. In a production pipeline, partition by key, make handlers idempotent, retain offsets or checkpoints, and apply backpressure instead of silently dropping data. Per-event work should be bounded, commonly O(1) or O(log k), with memory tied to the active window or maintained summary rather than total history. State explicitly how duplicate delivery and late events affect correctness.