Load, Replicas & Resilience
Model load honestly, run bounded real .NET10 experiments, and reason about queues, bottlenecks, replicas and operational limits.
- Workload models: open arrivals and closed workers
Define offered work before interpreting a throughput number.
Workload and admission · 28 min - Latency distributions and Little’s Law
Keep percentile definitions and steady-state units explicit.
Workload and admission · 28 min - Bounded queues and backpressure
Reject excess admission instead of concealing unbounded waiting.
Workload and admission · 28 min - Concurrency budgets and downstream pools
A replica’s worker limit cannot remove a smaller shared-resource limit.
Workload and admission · 28 min - Deadlines, queued work and cancellation
Budget the whole accepted operation instead of just one downstream wait.
Workload and admission · 28 min - Retries, amplification, jitter and idempotency
Bound repeated attempts and establish permission to repeat.
Workload and admission · 28 min - Stateless replicas and process-local session state
Two real processes do not automatically share application memory.
Replicas and resilience · 28 min - Cross-replica idempotency with a local SQLite ledger
Use a real uniqueness boundary while stating its coordination limits.
Replicas and resilience · 28 min - Circuit breaking and controlled recovery
Stop repeated calls during a known failure window.
Replicas and resilience · 28 min - Bulkheads and independent work lanes
Isolate admission budgets while recognizing shared process resources.
Replicas and resilience · 28 min - Readiness, draining and already accepted work
Stop new admission without pretending existing work vanishes.
Replicas and resilience · 28 min - Autoscaling lag and a real HPA manifest
A desired replica count is not immediately available capacity.
Capacity and operations · 28 min - Disruption budgets and rolling availability
Separate voluntary eviction limits from rollout and failure behavior.
Capacity and operations · 28 min - Partitioning and hot-key concentration
More owners do not redistribute repeated work for the same key.
Capacity and operations · 28 min - Single-instance threads, CPU and async waiting
Diagnose blocking without treating a tiny benchmark as proof of starvation.
Capacity and operations · 28 min - Capstone: measure one process versus two
Compare real replicas while preserving workload and uncertainty.
Capacity and operations · 28 min