01User outcomes, indicators, targets, demand, and workload models
Provide product analytics, service telemetry, a forecast and a request to prove scale. Ask the person to define the outcome, measurement and workload before selecting a generator.
Confirm: The person begins with correct user or system completion and distinguishes response time, end-to-end latency, queue time, throughput, goodput, freshness, batch duration, availability and resource efficiency; defines the measured population, success criteria, exclusions, time window, percentiles and error policy; separates SLI, SLO, SLA, internal threshold and capacity forecast; avoids universal speed targets; identifies interactive, asynchronous, streaming and batch classes; derives arrival rates, concurrency, pacing, think time, abandonment, session length, journey and operation mix, payload and response size, read-write balance, tenant and key skew, data volume and growth, cache state, geography, device, network, scheduled work, retries and dependencies; models normal, burst, seasonal, growth and failure demand; distinguishes an open arrival model from a closed virtual-user model; records missing or biased telemetry; and obtains owner acceptance of proxies and target tradeoffs.
02Reproducible scenarios, environments, generators, data, and safe execution
Provide an environment with smaller capacity, stale data, warm caches and shared neighbors, plus a script that saturates its own injector. Ask for a test system whose result can be repeated and bounded.
Confirm: The person maps the end-to-end path and differences among local, component, integration, preproduction and production conditions; records release, build, configuration, feature flags, topology, instance types, quotas, autoscaling, regions, network, dependencies and observability; designs smoke, average-load, stress, spike, soak, breakpoint and microbenchmarks only for their specific questions; creates separable safe accounts and representative data distributions; preserves cold, warm and steady states deliberately; validates functionality before load and samples correctness during it; synchronizes useful clocks and correlation identity; versions scripts, images, tools and parameters; sizes and monitors load generators, connections and network; uses distributed generation only when necessary and reconciles nodes; defines ramp, duration, warmup, steady state, repetition, cooldown and cleanup; controls cost and test data; obtains production approval; sets abort conditions for user harm, error, saturation, spend or lost observability; and restores the system after stress.
03Distributions, correctness, bottleneck hypotheses, and causal evidence
Provide a run with a healthy mean, a poor tail, fast failures, dropped arrivals, retry amplification and rising queue depth. Ask what the system actually did and where investigation should go next.
Confirm: The person keeps successful, failed, timed-out, canceled, rejected and wrong work visible; distinguishes offered load, accepted traffic, completed throughput and correct goodput; checks dropped iterations and coordinated-omission risk; reports counts and distributions with median and relevant high percentiles rather than averaging percentiles across workers; understands histogram boundaries, estimation error, sample size and aggregation; segments by journey, operation, status, payload, tenant, region, client and release without creating unusable cardinality; aligns client timing, server timing, traces, logs, profiles and resource metrics; observes latency, traffic, explicit and implicit errors, queues, locks, pools, caches, retries, throttling, garbage collection, CPU, memory, storage, network and dependency saturation; distinguishes symptom, correlation, contention and cause; identifies the capacity knee and nonlinear behavior; changes one useful factor; predicts the result; repeats under controlled conditions; and reports counter-effects on correctness reliability security cost and other workloads.
04Capacity envelope, regression, recovery, decisions, and continuity
Provide a proposed optimization, an autoscaling change, a prior baseline and a release deadline. Ask for a decision record that remains useful after traffic and architecture change.
Confirm: The person states capacity as a bounded curve across demand mix, latency, goodput, errors, saturation and cost rather than one maximum; separates sustained capacity, burst tolerance, breakpoint, safety margin and forecast; tests overload controls including quotas, backpressure, shedding and retry behavior without assuming they preserve correctness; verifies scaling and downscaling lag, state movement and dependency limits; observes recovery, backlog drain, cache refill and lingering resource damage after load stops; compares candidate and baseline on equivalent releases, environments, data, scenarios and repetitions; reports run-to-run variance and material environment drift; treats a threshold as one gate, not universal safety; stores raw and summarized results, queries and evidence; links regression to an owner and decision; updates continuous checks at proportionate frequency; retires stale workloads; and demonstrates that another qualified person can run, interpret and safely stop the suite.
