Designing a Streaming Analytics Platform for Sub-Second Insights
How we replaced a 4-hour batch pipeline with a streaming architecture that delivers dashboards in under a second.

## The batch ceiling
Batch pipelines have a hard latency floor. This post walks through the streaming architecture (Kafka → Flink → ClickHouse) we used to break past it, the rollup strategy that kept query cost flat, and the tradeoffs we accepted along the way.
### Key decisions
- Columnar storage for cheap aggregations
- Incremental rollups over raw replays
- Server-streamed queries for perceived speed