Size the surviving path before relying on recovery
What This Concept Is
Static capacity analysis asks whether already available resources can handle the defined demand after a failure. Dynamic recovery adds resources or changes configuration later. Both can be useful, but they answer different questions and require different dependencies to work.
Why It Matters Here
Semester 08 estimated replica counts. Here placement matters: losing one zone may remove several replicas at once. Autoscaling also depends on successful provisioning, quotas and capacity availability. A plan to recover eventually must not be presented as uninterrupted service.
Concrete Example
Three zones each have 50 requests/second of usable capacity. At demand 90, losing any one leaves 100; at demand 120 it leaves only 100. Total healthy capacity 150 is not enough evidence. With capacities 80, 40 and 30, losing the first leaves 70, so the worst single-zone failure differs from the average.
Use capacity measured at the required quality target, not a peak throughput number from an overloaded benchmark. A model with instantaneous redistribution can say capacity is sufficient while real clients still see errors during detection, connection draining and retries. Preserve that limitation.
Common Confusion / Misconception
Adding a zone is not automatically safer if it adds a mandatory dependency or changes quorum behavior. A capacity sum does not model data availability, election, retry amplification or request locality. Zero demand is also not proof that a deployed service is healthy; this teaching lab requires positive demand for every service.
How To Use It
List each named failure scenario and sum only surviving replicas. Compare with demand separately for every tier, accounting for fan-out when setting tier demand. Then require all mandatory dependencies to be ready. Record capacity units and how you obtained them. Keep the recovery-time experiment separate from the static calculation.
Check Yourself
Why does 150 total fail the 120-demand case? Which failure is worst in the asymmetric case? Does an available quota guarantee a successful launch? What assumptions permit adding capacities from two replicas?
Mini Drill or Application
Predict all single-zone outcomes before implementing the archive. Add a correlated two-zone loss and a tier whose demand is twice the API demand. Explain why adding API replicas cannot repair a missing database. Repeat the original network/failure exercise and contrast observed behavior with the prediction.
Read This Only If Stuck
AWS Well-Architected: static stability. Revisit the Semester 08 capacity and overload evidence before selecting model inputs.
The Linux Command Line is selective prerequisite support for inspecting local processes, network paths and resource inventories. Provider documentation above remains authoritative for cloud behavior.