VWAP Analytic Provider

Team-loop system-design round: ingest trade ticks from 100 exchanges (100K–1M symbols each, ~100 ticks/sec per symbol, ~20 bytes per tick), maintain per-(symbol, window) VWAP in a sharded stateful engine, and serve ~100K screen users via push and ~1K enterprise users via API. Expected beats: per-exchange feed handlers normalizing to a canonical tick, symbol sharding with running sums, snapshot/changelog recovery, and a pub/sub → WebSocket / Redis / query-API delivery tier. Requirements Design a VWAP (Volume-Weighted Average Price) analytic provider. VWAP is computed per symbol over a specified time window from continuously arriving trade ticks, and published values update periodically. Functional requirements: Ingest trade ticks from 100 exchanges in real time. Maintain the VWAP for every (symbol, time window) pair as ticks arrive. Update published VWAP values periodically. Serve the computed values to two distinct user populations. Scale anchors given by the interviewer: 100 exchanges, each listing 100K–1M symbols. Roughly 100 ticks per second per symbol, ~20 bytes per tick. User populations: Screen users (~100K): real-time display, updates can be pushed. Enterprise users (~1K): API-driven queries with stricter data-integrity expectations. Context the interviewer feeds in during the round: each exchange has its own data format, so ticks pass through a per-exchange feed handler / adapter that normalizes them into a canonical tick format. Discussion arc the interviewer drives: VWAP engine design — stateful or not, sharding by symbol, maintaining running sums. Fault recovery for the stateful computation — snapshots and changelog replay. Delivery — pub/sub fan-out, a WebSocket gateway for screen users, a cache (e.g. Redis) holding the latest VWAP per symbol, and a query API for enterprise users. Notes A standard streaming + stateful-computation design in Bloomberg's home domain (market-data pipelines). The graded axes are tick ingestion at volume, per-symbol sharded aggregation, recovery of in-flight state, and serving push and pull consumers from the same computed stream. The interviewer supplies background incrementally rather than all up front — expect the requirements to grow mid-design and keep the architecture open to the normalization tier and the second user population. Preparation Drill windowed streaming aggregation: running sums of price×volume and volume per (symbol, window), window advancement, and periodic emission. Run the back-of-envelope — up to 100 exchanges × 1M symbols × 100 ticks/s × 20 bytes — and use it to justify per-exchange feed handlers, symbol-hash sharding, and why a single aggregator cannot hold the state. Rehearse a stateful-stream recovery story (periodic snapshots plus changelog replay) and how the delivery tier — pub/sub into a WebSocket gateway for screens, cached-latest plus a query API for enterprise — hangs off the engine.