Table of contents
- Quick read
- Readiness check
- Define the question the load test answers
- Size the environment before the window
- Book the window and name a stop owner
- Use synthetic identities and masked data
- Watch shared dependencies during the run
- Publish results without overstating them
- Role brief
- Frequently asked questions
- Sources
Is the handoff ready?
Use these checks before access expands or the first ticket begins.
Question
Ready: The test names a threshold and the decision it informs.
Needs work: The team runs a generic stress test and reads the graph later.
Environment
Ready: The target is separate and its differences are recorded.
Needs work: The test points at the production cluster to save time.
Stop control
Ready: One named person can stop the run without approval.
Needs work: Stopping requires a meeting or a vendor ticket.
Data
Ready: Synthetic records and sandbox integrations are verified.
Needs work: The run copies live customer records.
Define the question the load test answers
A load test should answer a specific question, such as whether checkout holds at twice the current peak or whether a new endpoint stays inside its latency target. Write the question and the threshold before the test starts. Do not run a generic stress test and interpret the graph afterward. A test without a question produces a number without a decision.
Name the user journey, the endpoints, and the data each step needs. Separate a capacity test from a soak test and from a failure test. Each one needs a different duration and a different stop condition. Record the choice and the reason in the test ticket.
Size the environment before the window
Size the test environment so that the result says something about production. Match the application version, the database size, and the network path closely enough for the question. Record every difference from production in the test note. A small environment with a large claim is misleading.
Check the quotas on the load generator, the database, and any third-party API before the window. Request an increase early if the test needs one. Confirm that the test cannot exhaust a shared production quota. The environment should fail inside the test, not in a customer account.
Book the window and name a stop owner
Book the window with the client, the on-call engineer, and any vendor whose service the test touches. Publish the start time, the duration, and the expected load. Name one person who can stop the test without asking for permission. A stop that needs a meeting is not a stop.
Agree on the signals that trigger a stop, such as error rate, database latency, or a support queue spike. Write the thresholds and the person who watches each one. Keep a shared channel open for the whole window. Record the time of every stop decision.
Use synthetic identities and masked data
Generate test accounts and records rather than copying production customers. Keep the generation script under version control and record the seed. Check the dataset for real email addresses, phone numbers, and payment values before the run. A load test that uses live data is a data incident with a graph.
Make sure the test cannot send real email, charge a card, or ship a parcel. Disable outbound integrations or point them at a sandbox. Confirm the isolation with a test order before the load starts. Record the check and the person who ran it.
Publish results without overstating them
Report the question, the environment, the load profile, the result, and the differences from production. State what the test did not cover and which dependencies were mocked. Give the raw data to the client technical owner. A summary without the method invites a wrong decision.
Turn each result into a decision, such as a capacity change, a code fix, or a follow-up test. Record the owner and the date. Do not publish a headline number without the conditions that produced it. The next team should be able to repeat the test from the record.
Copy-ready brief
Put the operating agreement in writing.
State the journey, the load profile, the threshold, and the decision the result will inform.
Record the version, data size, network path, and every difference from production.
Name the stop owner, the signals, the thresholds, and the shared channel.
Confirm synthetic records, masked values, and disabled outbound integrations.
Frequently asked questions
Can a load test run against production if the load is small?
Only with explicit client approval, a stop plan, and synthetic accounts, and the test note should say so.
Who decides to stop the test?
One named stop owner with the authority to halt without asking for permission.
What belongs in the report?
The question, environment, load profile, result, limits, and the decision that follows.
Sources
- Grafana k6 documentation
Load test design and threshold practice.
- OWASP ASVS
Sensitive data handling and environment checks.
- NIST SP 800-53 security and privacy controls
Capacity and contingency planning expectations.
- Google SRE Workbook
Capacity planning and incident response practice.
