Latency on a financial event mesh

Millions of financial events a day. The p99 was the product. Stream layout and backpressure were the work.

2024 — 2025

−34%

p99 latency

Millions of events / day

Volume

The situation

High-throughput services sat on a Kafka-centric event mesh and had to keep up with millions of financial events daily. Tail latency was not a dashboard curiosity — it was whether the next consumer stayed caught up.

The work

I owned the distributed services on that path and went after p99 through stream repartitioning and backpressure tuning.
I designed resilience in as defaults: circuit breakers, bulkhead isolation, and idempotent consumers.
I mentored two mid-level engineers through architecture reviews and production incident retrospectives.