Beta slice: five scenarios, one interaction phenomenon each. You write the component handlers and the policies — timeouts, retries, caches, workers. The platform supplies the load and the failure. The naive policy runs first, always, because the collapse is the lesson.
One process serves every search. When the ticket queue spikes, the p99 tail is the queue — and there is no queue data structure anywhere.
Every prediction needs a feature lookup. A's contract is composed of B's contract, and A's p99 is B's p99 plus overhead — until A grows a budget.
Payments starts failing for ten seconds. A retry feels helpful, but every attempt adds load to the same sick dependency. Bound the work before the outage spreads.
The payment provider is still sick. A retry budget limits damage, but the gateway keeps paying the cost of discovering the same failure. Open the circuit and degrade deliberately.
A traffic burst is not a reason to let every request enter the dependency. Decide what the system can afford, reject excess work early, and keep the accepted path useful.
queueing in a latency tail → synchronous coupling → retry storms → caching and the herd → backpressure → partitioning → replication and staleness → the incident. Every scenario reuses the same transport: only your policies change between the naive run and the fixed run.