Enterprise software
Six-fold latency reduction by decomposition
A legacy monolithic pricing engine taken apart into twelve resilient services, cutting median latency six-fold and maintenance cost by well over half.
- 12s → 2s
- median response time
- 6×
- median latency improvement
- 60%
- reduction in system maintenance cost
- 12
- independently deployable services
A pricing engine that had accreted for a decade. Every quote ran the whole monolith; a twelve-second median response was accepted as the cost of the domain being complicated.
Diagnosis
It was not complicated. It was coupled. The slow path touched subsystems it had no business touching, because there was no boundary preventing it. Latency was a structural property, so no amount of profiling was going to fix it — profiling finds hot code, not bad seams.
Approach
Decomposition along domain boundaries rather than technical layers:
- Event-storming the pricing domain to find where the real invariants sit, before a line of code moved.
- Strangler-fig migration — each extracted service took live traffic behind the existing interface, so there was never a big-bang cutover to be afraid of.
- Twelve services, each owning its data, each deployable without a coordinated release.
Outcome
Median response time went from 12 seconds to 2 — a six-fold improvement — and overall system maintenance cost fell 60%. The second number is the one that funded the work: the architecture stopped charging rent.