Real-Time Analytics Platform · View 14 of 21 · Runtime

Query and Serving Architecture

How a query reaches an answer in under 2 seconds while dashboards, ad-hoc analysis and exports run concurrently without competing.

Editable source SVG draw.io All views

Predictable latency

Standard dashboard queries resolve against materialized views, not raw scans
Redis absorbs roughly 55% of dashboard traffic on a 30-second TTL
Workload groups cap concurrency and memory per class so one query cannot starve the rest

Workload separation

Ingestion writes to the leader; dashboards and ad-hoc each read a dedicated follower
Followers can be scaled independently of ingestion capacity
Exports and notebooks are contractually second-class and may be throttled first

Trade-off accepted

Follower attach lag adds a small freshness penalty on read replicas
That penalty is inside the 5-second budget and is measured as part of the freshness SLI
A truly freshest-read path exists against the leader, restricted to operational views

◀ Event to Insight — Critical Path All views Replay, Backfill and Correction ▶