Designs and interprets load tests - smoke, load, stress, soak, and spike - with realistic ramp profiles, percentile-based pass/fail thresholds, and a bottleneck-reading procedure, producing a runnable test plan. Use when someone asks "how many users can we handle", "how do I write a k6 or Gatling test", "will this survive the launch", "why did the site die at 500 users", or is preparing for a traffic event. Do NOT use for diagnosing frontend page-speed or Core Web Vitals - use web-performance instead - and do NOT use for tuning database connection pools found as the bottleneck - use connection-pool-tuner instead.
Click to play with sound.
---
name: Load Testing
description: Designs and interprets load tests - smoke, load, stress, soak, and spike - with realistic ramp profiles, percentile-based pass/fail thresholds, and a bottleneck-reading procedure, producing a runnable test plan. Use when someone asks "how many users can we handle", "how do I write a k6 or Gatling test", "will this survive the launch", "why did the site die at 500 users", or is preparing for a traffic event. Do NOT use for diagnosing frontend page-speed or Core Web Vitals - use web-performance instead - and do NOT use for tuning database connection pools found as the bottleneck - use connection-pool-tuner instead.
---
# Load Testing
The point of a load test is to find the capacity limit and the failure mode before users do, and to turn "will it hold?" into a number with a date on it. The costly mistake this skill prevents is the vanity test: hammering one cached endpoint from one laptop, reading the average latency, and declaring readiness - a test that passes while production dies at one-third the "tested" load.
## Operating procedure
### Step 1: Gather inputs
Collect before writing any script; label estimates as guesses.
1. Target load: expected peak in requests/sec or concurrent users, and where the number came from (analytics, marketing forecast, guess).
2. SLOs: the p95/p99 latency and error-rate targets that define pass/fail. If none exist, propose p95 < 500ms, p99 < 1500ms, error rate < 1% as defaults and get sign-off.
3. Top 3-5 user journeys with their real traffic mix (e.g. 80% browse, 15% search, 5% checkout) - from production access logs, not intuition.
4. Environment: production-like staging (same instance sizes, same data volume) or coordinated production test. Toy data lies: a 1,000-row table performs nothing like a 100M-row one.
5. The traffic event or deadline driving the test.
### Step 2: Pick the test types - each answers a different question
| Type | Question it answers | Profile |
|---|---|---|… load the full skill through Skill Me