# Load Testing

Load testing evaluates system behavior under a defined workload to check whether expected demand can meet performance goals.

- Type: method
- Domain: Engineering & architecture
- Level: Intermediate
- Last reviewed: 2026-10-19
- Original: [Soroush’s Product Wiki](<https://wiki.soroush.uk/topic/load-testing/>)

## What it means

Load testing applies representative traffic or work to a system and measures response time, throughput, errors, and resource behavior. The workload includes request mix, arrival pattern, data size, and duration, not just a user count. Stress testing intentionally goes beyond expected capacity; a normal-load test answers a different question.

## Why it matters

Capacity problems often appear only when requests compete for shared resources. Evidence from realistic demand helps teams plan launches, evaluate bottlenecks, and avoid promising responsiveness that a system cannot sustain.

## When to use it

Use it before major demand changes or architecture changes and when investigating capacity risks. Start in a controlled, representative environment with agreed limits. A tiny or unrealistic test system supports narrower conclusions than production-equivalent testing.

## How it works

1. Define expected traffic patterns and measurable success thresholds.
2. Build representative journeys, data volumes, and arrival behavior.
3. Check that the load generator itself is not the bottleneck.
4. Ramp to the target, hold it, and measure latency percentiles, errors, throughput, and resources.
5. Investigate bottlenecks, change one relevant factor, and repeat comparable measurements.

## A practical example

A ticketing team models browsing and reservations at an anticipated launch rate. Average latency looks acceptable, but the slowest reservation requests exceed the agreed target when inventory locks contend. The team fixes that bottleneck before increasing marketing traffic.

## Where it breaks

Synthetic traffic cannot capture every production behavior. Caches, network paths, data distributions, and third-party quotas affect validity. Successful testing at one load does not establish safe behavior above it or during long-running resource leaks.

## Common mistakes

- Reporting only average latency.
- Generating unrealistic requests that never exercise the costly path.
- Confusing virtual-user count with a guaranteed request arrival rate.

## Sources

1. [Average-load testing](<https://grafana.com/docs/k6/latest/testing-guides/test-types/load-testing/>) — Grafana Labs · Official documentation
2. [Architecture strategies for performance testing](<https://learn.microsoft.com/en-us/azure/well-architected/performance-efficiency/performance-test>) — Microsoft · Official documentation

## Related topics

- [Service-Level Objective](<https://wiki.soroush.uk/topic/service-level-objective/>)
- [Observability](<https://wiki.soroush.uk/topic/observability/>)
- [Database Index](<https://wiki.soroush.uk/topic/database-index/>)
