Load testing answers a different question from unit/integration testing: how does the assembled system behave when many requests overlap? Phase 31 uses autocannon because it is simple, scriptable, Node-friendly, and sufficient for learning repeatable HTTP load experiments. # PowerShell - run against a production-like local/staging build pnpm dlx autocannon@8.0.0 ` -c 20 ` -d 30 ` -m GET ` -H "Authorization=Bearer $env:TASKFLOW_TOKEN" ` "http://localhost:3001/tasks?page=1&pageSize=20"
Record more than requests/second Measure Why
p50 / p95 / p99 latency Tail behavior matters to users
throughput How much useful work completes
non-2xx / errors Fast failures are not success
payload bytes Understand bandwidth/serialization cost
Use a warm-up run, fixed dataset, fixed concurrency/duration, production build, and consistent environment. Change one variable at a time. Keep benchmark results with commit/release information so comparisons are reproducible. Production caution: Do not point uncontrolled load tests at production or shared environments. Obtain authorization, define limits, and prefer isolated staging/performance environments.
Tests are executable behavioral contracts. Good tests protect meaningful invariants and failure cases, not implementation trivia.
Restate the goal of “Load Testing with Autocannon” as one observable behavior. Write the success signal before touching code.
Trace the boundary map: behavior contract → arrange → act → assert → regression signal. Mark the one boundary this lesson changes.
Inspect the current repository state and record the exact files, route, table, or runtime involved. Avoid creating an abstraction before you can name the existing owner.
Implement the smallest vertical slice. Add one coherent change, run the narrowest verification command, and read the output before continuing.
Connect the slice to TaskFlow’s phase artifact: Before/after performance report with measured bottleneck, change, and regression guard.. Verify behavior from the outside as a user or API client would.
Create one controlled failure that is safe to reproduce. Predict the symptom first, capture evidence, locate the earliest broken boundary, then fix only that cause.
Close the reference and rebuild or explain the core change from memory. Record what you would monitor or test before trusting the same change in production.
it("rejects deleting another member's task", async () => {
const response = await asUser(bob).delete(`/tasks/${aliceTask.id}`);
expect(response.status).toBe(403);
expect(await taskStillExists(aliceTask.id)).toBe(true);
});Do not copy this as a finished answer. Predict what each boundary does, type the smallest useful slice, run it, then compare the result with your prediction.
Look for shared database state, timers, globals, unordered execution, or missing cleanup before adding retries.
Move the highest-risk boundary to an integration test with the real adapter/database.
Without copying the final source, create a second small scenario that applies “Load Testing with Autocannon” to a different TaskFlow resource or input. Write the acceptance criteria first, implement it, and provide the evidence that each criterion is satisfied.
Measure API endpoint latency with enough context to separate application and database time.
Use the smallest change that demonstrates this lesson. Stop after it works and capture evidence.Optimizing from intuition only
Predict the symptom before causing the failure. Then locate the earliest boundary where reality diverges from expectation.Why does this responsibility live here, what input does it trust, what output proves success, and what would go wrong if another layer owned it?
If your explanation is vague, repeat the boundary trace before moving on.Close the reference. Reproduce the core idea from memory or explain the exact request/data flow aloud. Then check your diff or notes for accidental shortcuts that bypass validation, authorization, error handling, or verification earned in earlier phases.