Sub-second queries on fresh events

Real-time analytics combines two demands that pull against each other. New events, usually read from a stream such as Kafka, have to be queryable within seconds of arriving, not after a nightly batch load. And queries have to come back fast enough to sit behind a web page or an API, sometimes at thousands of queries per second when the dashboard is shown to customers rather than to a few analysts.

Systems built for it ingest from streams directly, keep data in columns sorted or partitioned by time, and use indexes, pre-aggregation (rollups or materialized views) and caching so a typical query reads little data. To tell whether one really handles it, measure the time from an event being produced to it showing up in a query, and query latency at the concurrency you expect, not one query at a time. Check also how it handles late or corrected events, since many of these systems append well but update poorly.