Systems often fail not at launch but after initial success when early architectural assumptions break under sustained growth.
Increasing database connection pool size can paradoxically decrease throughput due to context switching, following a non-monotonic inverted U-curve.
The `MERGE INTO` operation in copy-on-write data lakes causes write amplification, where performance scales with files rewritten, not rows changed.
In event-driven microservices, using thin events creates hidden read amplification as consumers make synchronous callbacks, compounding load on the producer.
Failures often share a pattern where a surface abstraction hides a physical cost, leading to a scale-triggered failure after a delayed feedback loop.
To survive growth, systems must prioritize visibility of costs, understand how abstractions hide costs, and treat architectural decisions as falsifiable hypotheses.
When a stable system degrades, the root cause is often a foundational assumption that no longer holds true, not a simple code bug.
Jobs that call for the skills explored in this talk.