Best for
- The launch or campaign you suspect will triple your traffic
- Finding the first bottleneck before customers do
- Replacing 'we think it scales' with numbers
What you give it
- The system, its real traffic patterns (or your best estimates), and the event you fear
What you get back
- Scenarios shaped like reality: the user mix, the arrival pattern, the data spread
- The targets stated upfront: what must hold at what load for the test to pass
- A results readout: where it broke first, what degraded before failing, and what to fix in which order
How it works
- Models traffic from reality: the user-journey mix, arrival shapes (spike, ramp, soak), and data variety that defeats unrealistic caching.
- Sets pass criteria before running: latency percentiles and error rates per journey at the target load.
- Runs in escalating stages, watching the whole system — the bottleneck is usually not where the graph is pointed.
- Separates degrades-gracefully from falls-over, because the difference is the user's experience of the bad day.
Example
You: The TV spot airs in three weeks. Can we take 20x our normal Saturday?
Result: Three scenarios: the spike (0 to 20x in 90 seconds, the TV shape), the sustained plateau, and the soak (4 hours for the leaks). Found: the system holds 14x — the first wall is the connection pool (fixed: 20x clears), the second is the search service at 25x (documented, not fixed — beyond the need). Checkout p95 stayed under the 2-second target throughout; the readout says what to watch on the night.
Limits — please read
- A test environment is not production; it states the differences and their direction of error.
- Third-party dependencies cannot be load-tested without permission; they are stubbed at realistic latencies, noted.
- Numbers age as code changes; the plan is rerunnable on purpose.