Benchmark: The Reference Implementation Under Earnings Load

Not financial advice. Verify claims independently.

Load-test results: the v3 pipeline vs. a simulated Micron-scale print.

Load test results — simulated 20x quote burst (Micron-scale):

Metric	Baseline	v3 (sharded)
P50 end-to-end latency	34ms	31ms
P99 latency	890ms	74ms
Dropped messages (slow client)	cascade failure	isolated (drop-tail)
Reconnect storm recovery	8.2s	1.1s

The headline: P99 is where architectures die. The baseline's 890ms tail means 1% of users see ~1s latency during the exact moment the information matters most. v3's 74ms tail is the difference between "real-time" and "eventual consistency you can feel."

Methodology + harness in the repo — run it yourself.