Skip to main contentSkip to user menuSkip to navigation
Interactive design lab

Load-Test Design Workbench

What is load testing?

Load testing applies controlled, representative traffic to a system so a team can measure capacity, latency, errors, queueing, and recovery before real users discover the limit.

Step 1

Model

Choose an open-loop arrival rate or closed-loop virtual-user model, set pacing and journey mix, describe tested service and dependency limits, then define acceptance thresholds and automatic stop conditions.

Step 2

Observe

Compare offered load with admitted and successful throughput, inspect concurrency and latency percentiles, find the saturated resource, and decide whether the plan should pass, fail, or abort.

Step 3

Challenge

Start with the planned ramp. Then inject a ramp spike, soak leak, dependency slowdown, retry storm, and capacity wall. Treat every result as a planning estimate until a controlled run produces measured telemetry.

Load-test design workbench

Design the run before generating traffic

Shape arrivals, model service limits, inject failure, and decide whether the planned run should pass, fail, or stop.

Planning estimate, not measured data
Challenge mode

Pressure-test the healthy plan

Configured rampNo fault injection
1

Shape traffic and workload

Choose how load is generated, then define journey cost.

Arrival model

Open-loop generators maintain arrivals even when responses slow.

Closed-loop users wait for a response and then apply pacing.

Used to estimate VU demand in both planning modes.

Excludes setup, data seeding, and cool-down.

Read-heavyWrite-heavy
Reads
80%
1.0 work unit
Writes
20%
2.8 work units
2

Define the system envelope

Use limits measured in a controlled environment when available.

3

Set acceptance and stop limits

Acceptance defines success. Stop limits protect the test system.

Estimated run verdict

Plan passes

The estimate stays inside every acceptance threshold and no stop condition fires.

Planned ramp
1,800 RPS open-loop arrival/Application fleet is the tightest constraint
Offered load
1,800 RPS

1,800 RPS estimated successful

Concurrency
179.5

1,530 VUs estimated with pacing

Saturation
78%

78% app / 52% dependency

Errors
0%

240 of 3,240,000 admitted requests

Traffic profile

Offered load versus estimated ceiling

30 minute plan
Offered trafficEstimated throughput ceiling
6191,2391,8582,4780 min30 min
Request path

Trace pressure through the system

Load generator
1,800 RPS
750 ms pacing
Application fleet
78%
2,294 RPS logical capacity
Required dependency
52%
1.01 attempts / operation
Admitted throughput
1,800 RPS
Estimated queue
0 requests
Memory headroom
35%

Latency percentiles

P50
Target 450 ms
100 ms
P95
Target 450 ms
175 ms
P99
Target 1,500 ms
258 ms

Percentiles are deterministic planning estimates derived from configured service time, queue pressure, retries, and scenario modifiers. Replace them with measured histograms after a real run.

Acceptance criteria

P95 latency
<= 450 ms
175 ms
Error rate
<= 1%
0%
Peak saturation
<= 80%
78%

Automatic stop conditions

A real runner should stop new arrivals, retain diagnostics, and cool down when any safety condition fires.

P99 latency stop
258 ms
Trigger > 1,500 ms
Error-rate stop
0%
Trigger > 5%
Queue-depth stop
0
Trigger > 5,000
Interpret the estimate correctly

This model helps expose contradictory targets and unsafe test plans. It is not telemetry and does not claim to predict production exactly. Calibrate per-instance capacity, service times, failure rates, and journey weights from controlled experiments before approving a real run.

No quiz questions available
Could not load questions file
Spotted an issue or have a better explanation? This page is open source.Edit on GitHub·Suggest an improvement