What Worked

Successes

Complete message-driven path

CSV rows become canonical domain events, statistics, signal snapshots, decisions, and portfolio reports.

Same data, different outcomes

The strategy layer makes behavior visible: NYC buys aggressively, LON waits for volatility confirmation, TYO stays flat.

Patterns solve real seams

GoF patterns organize object behavior; EIPs organize message movement and routing responsibilities.

Failure handling stays explicit

Invalid market records are separated from valid flow, making data quality issues visible without blocking processing.

What Was Hard

Pain Points

Correlation had to be precise

engineId:correlationId:symbol prevents statistics from different engines or symbols being aggregated together.

Rolling-window warmup is non-obvious

Early ticks correctly produce no statistic until the window is full, which can look like missing output during a demo.

Topic granularity adds wiring

By-statistic topics make Selective Consumer clear, but they create more route definitions and endpoint mapping.

Asynchronous output needs discipline

Concurrent routes can interleave decisions and reports, so centralized recording is important for clear final summaries.

Statistic strategies run against each valid tick after warmup.

Trading engines consume overlapping but different symbols and stats.

Error queues keep invalid records and dead letters visible.