Performance Testing Engineer Interview Questions

Option B is strongest: it names four test types (not just two) with distinct objectives, load profiles, and success criteria for each, explains what the stress test produces (a capacity envelope), introduces soak test failure modes that are invisible in shorter tests (memory leaks, connection pool exhaustion), and names the common mistake (treating staging results as production guarantees) with the specific reasons staging differs. Performance test vocabulary: Load test — validates performance meets SLOs at expected production load. Stress test — finds the system's capacity limit and failure mode. Soak test — exposes time-dependent issues at steady load over hours or days. Spike test — tests autoscaling and recovery under a sudden load surge. Capacity envelope — the operating range within which the system meets performance requirements. Graceful degradation — the system slows under stress but remains functional and recovers. Options C and D are accurate but lack the soak test details and the staging caveat explanation.

Option B is strongest: it explains WHY averages are misleading with a concrete numeric example (999 × 100ms + 1 × 10000ms = ~110ms average), explains tail latency amplification in microservices with the mathematical consequence (3-service chain degrades combined p99), provides usage guidance for each percentile level, and ends with the SLO principle (never average alone). Latency vocabulary: p50 (median) — 50% of requests complete faster than this value. p99 — 99% of requests complete faster; 1% are slower. Tail latency — the slowest portion of the latency distribution. Tail latency amplification — in a microservices fan-out, tail events from multiple services compound, increasing the combined tail latency. Arithmetic mean — the average; sensitive to extreme values (outliers). Options C and D are accurate but lack the numeric example showing average masking and the mathematical explanation of tail amplification.

Option B is strongest: it names four distinct problems with CI/CD performance testing (not just the "add a stage" view), explains WHY tiered testing is necessary (full tests don't fit in PRs), introduces relative baseline comparison vs. absolute thresholds with the reason (different services have different absolute targets), names the environment fidelity gap explicitly, and introduces noise reduction techniques (3-run median, Mann-Whitney). CI/CD performance vocabulary: Smoke performance test — a minimal load test that verifies basic endpoint responsiveness, run on every PR. Baseline — the stored performance metrics from the last passing build, used for regression comparison. Regression threshold — the maximum acceptable percentage degradation from baseline before the pipeline fails. Environment fidelity — the degree to which the test environment matches production. Mann-Whitney U test — a non-parametric statistical test used to compare performance distributions for significance. Options C and D are accurate but lack the baseline-vs-absolute-threshold rationale and the Mann-Whitney reference.

Option B is strongest: it structures diagnosis as a four-step flow with dependencies between steps, introduces distributed tracing as the tool to confirm the bottleneck is in the DB (not the application), applies the USE method with specific resource combination interpretations (CPU high + disk low = compute-bound), names the sort order for slow query analysis (total_exec_time DESC for highest-leverage), and introduces lock contention as a load-specific failure mode with the symptom (p99 spikes correlating with transaction conflicts). DB performance vocabulary: USE method — a framework for resource analysis: Utilisation, Saturation, Errors. pg_stat_statements — a PostgreSQL extension that records query execution statistics. EXPLAIN ANALYZE — a query plan command that shows actual execution time and row counts. Sequential scan — a full table scan; indicates a missing or unused index. N+1 query pattern — executing N+1 database queries to fetch N related records. pg_locks — a PostgreSQL view showing active lock information. Options C and D are accurate but lack the USE method resource combination interpretations and the lock contention mechanism.

Option B is strongest: it distinguishes saturation from utilisation (the key insight — 60% CPU with a run queue is more saturated than 70% CPU without), gives the concrete metric for each resource type with the specific tool (vmstat for CPU), introduces the USL (Universal Scalability Law) as the theoretical framework for throughput decline under saturation, names page faults as memory saturation (not just high RAM usage), and introduces the leading vs. lagging indicator distinction. Saturation vocabulary: Saturation — a resource has more demand than capacity; additional work must queue. Run queue length — the number of processes waiting for CPU; the primary CPU saturation metric. Knee of the curve — the inflection point where latency begins to increase steeply as load increases. Universal Scalability Law (USL) — a model describing throughput growth as concurrency increases: growth slows, then reverses due to coherence overhead. Page faults per second — a memory saturation signal indicating OS swapping. Leading indicator — a metric that signals saturation before performance degrades (run queue, queue depth). Options C and D are accurate but lack the CPU run-queue-vs-utilisation example and the USL reference.