Speaking Topic: Reliability Engineering & Incident Response
Systems rarely fail from load. They fail when they change state, and most incident response is built to watch the wrong signal.
Most platforms do not collapse because of high traffic volume. They collapse when a transition threshold changes behavior: a cache eviction cascade, a connection pool saturation, or an unindexed failover state shift. Traffic is the trigger. State change is the mechanism.
I build this talk around the London Ambulance Service's 1992 dispatch failure, when LASCAD, a computerized dispatch system designed to replace manual map boards with automated algorithmic dispatch, collapsed under exactly this kind of state-transition failure, not a traffic spike. The lesson is thirty years old and still ignored.
Modern cloud-native architectures, spanning hundreds of Kubernetes pods, serverless workers, and distributed message queues, fail in subtle, pathological ways that traditional load testing simply never surfaces, because load testing tests load, not state transitions.
Key Takeaways
- 01State transitions, not raw traffic, are the actual mechanism behind most platform collapses.
- 02The 1992 LASCAD dispatch failure remains the clearest historical case study in state-transition failure.
- 03Load testing does not surface state-transition failures; it was never designed to.
Frequently Asked
If not traffic volume, what actually causes most platform outages?
State-change transitions such as cache eviction cascades, connection pool saturation, or unindexed failover shifts. Traffic volume is usually the trigger, not the underlying mechanism.
From the Q&A
“If traffic isn't the cause, why do outages always follow traffic spikes?”
Traffic is the trigger that exposes a state-transition failure that was already latent, such as a cache eviction cascade. The spike reveals the mechanism, it does not create it.
“What's a quick way to check if a team has this blind spot?”
Ask whether the last major incident's post-mortem named a structural cause or just a traffic number. Most name the traffic number.