Performance Engineering

Beyond load test

• Architecture review.
• Code profiling.
• DB optimization.
• Caching strategy.
• Capacity planning.
• Right-sizing infra.

Process

• Define SLO/SLA — what is acceptable?
• Baseline — measure current.
• Identify bottleneck — profile, monitor.
• Optimize — fix bottleneck.
• Validate — re-measure.
• Repeat.

SLI/SLO/SLA

• SLI — indicator (metric).
• SLO — objective (target).
• SLA — agreement (with consequence).

Example:

• SLI: p95 API latency.
• SLO: p95 < 500ms 99% of time.
• SLA: refund customer if SLO miss > 1h.

Common bottlenecks

• DB — slow query, missing index, N+1.
• Network — chatty API, large payload.
• CPU — inefficient algorithm, GC pressure.
• Memory — leak, large object.
• Disk — slow IO.
• External — third-party API slow.

Tools

• APM: Datadog, NewRelic, Dynatrace.
• Profiler: pprof, Async Profiler (Java), pyflame.
• DB: pg_stat_statements, MySQL slow query log.
• Tracing: Jaeger, Honeycomb, Zipkin.

Capacity planning

• Forecast traffic 6-12 month.
• Stress test 2-3× current.
• Identify scaling cliff.
• Auto-scaling rules.

VN context

Tết peak: e-com 10× normal, banking 5×, food delivery 7×. Production capacity buffer always > peak forecast 30%.

Performance budget

Set per page / per request:

• Mobile homepage < 200KB.
• API p95 < 500ms.
• DB query < 100ms.
• Block release if regression > 10%.