Keyed state

Keyed state - state partitioned by key, the keyed state streaming engines are built on: each key's aggregate lives and updates on its own. Hash the key, and every event for user_42 lands on state that only user_42's events ever touch, which is why a job can be split across cores or machines without two workers ever fighting over one counter.

That is also the honest one-line answer to why streaming scales at all. Aggregations, windows and joins all reduce to per-key state, the engine routes events by key, and parallelism falls out for free. The cost sits on the flip side: anything that has to look across keys, a global top-N, a full sort, breaks the partitioning and gets expensive again, in every engine ever built.

Do you need it? Yes. Learn it, it costs one paragraph and it stays true everywhere, ubik included. A GROUP BY merchant in ubik is keyed state with the plumbing removed: one running value per key, and keys that outgrow the memory budget spill to disk instead of taking the process down. Same idea at every scale.