Chapter 39 worked answers

These answers use baseline cbfc94ade5b57e449d6b73bf1485b1c5b1647022 and the chapter's linked execution records. Procedures remain scoped to disposable resources.

1. Identify the timed work

average_convergence_time averages initial service creation plus synthetic running reports in MemoryService. It excludes engine image pull/start, application health, HTTP/TLS and durable database work. SchedulerIterations counts simulation ticks; ReconciliationIterations counts initial service creations. Names alone do not prove package controller invocations. The observed100-service run's323records include historical failures/replacements rather than323desired live replicas.

2. Repeat bounded samples

Use the exact bounded-benchmarks record: Go1.25.0, Linuxamd64, CPUcount2,100ms, count3, benchmark memory reporting. Scheduler Plan's observed samples were11,991,457/12,012,169/12,050,878ns per operation, each718,664bytes and630allocations. The fixture has100ready nodes and1000pending tasks with fixed requests. Median/range describe these samples only. Shared-host variation and three short samples do not justify a confidence interval or production cluster limit. Correctness properties must remain intact when comparing optimizations.

3. Separate listener does not mean authenticated

NewDebugHandler registers pprof without its own auth middleware. Default loopback reduces exposure, but configurable ORCH_DEBUG_ADDR can broaden it. Restrict management access, bound captures, and treat profiles as sensitive. The chapter captured a synthetic benchmark profile without opening a listener. Prebuilt go tool pprof was absent; source-built go run cmd/pprof succeeded separately, with a short120ms CPU sample total.

4. Build a different production experiment

Pin images/cache state, services/replicas, update rate, constraints, authoritative durable store/schema, network path and readiness criterion. Measure API acceptance, scheduler computation, actual runtime start and application health separately. Include retained task/event history, database errors, partitions/recovery and exact cleanup. The tagged thousand-service test passed only task-count/iteration assertions with fake transitions. Running PostgreSQL beside that command does not make it a persistence benchmark. Record unresolved work rather than extrapolating a safe production maximum.

See the corresponding chapter claim ledger for exact immutable source anchors and limitations.

Return to Chapter 39.