/distributed-systems
Distributed systems that stay correct at peak traffic
High-concurrency systems fail in the gaps: a global lock that starves writers, a race between two services, a shard that runs hot. I design the data models and coordination patterns that remove those gaps, then prove them under real traffic.
Architectural invariants
- A global lock on the hot path is a scaling ceiling, not a safety feature.
- Shared state changes by appending ordered events, not by read-modify-write.
- Every piece of state has exactly one owning service.
- Correctness is proven at peak load, not at average traffic.
Proof in production
- Replaced a single global lock with an append-only event log in a high-concurrency service, taking lock starvation from 100% to 0% at peak.
- Led the design of a lock-free data pipeline for real-time model ingestion.
- Designed a sharding and hybrid storage migration (object storage + relational) for high-frequency event data.
- Decoupled message delivery across 30+ microservices with zone-aware routing, validated through fault injection.
- Led monolith decompositions, including replacing a monolithic request flow with a plug-in API gateway.