Testing Practices

Philosophy

Test your software, or your users will.

• Test against contracts, not implementations assert what it should do, not how it does it. Tests that break on every refactor are coupling to internals.
• State coverage > line coverage exercise meaningful paths and edge cases, not just lines. We intentionally use no coverage tools; percentage targets create false confidence.
• Tests are the first users of your API if tests are hard to write, the design is wrong. Refactor the interface, not the test.
• Property-based testing finds edges you didn’t think of complement example-based tests with fuzz and property tests where the input space is large.
• Tests should be boring a test that’s hard to read is a test nobody trusts. Inline data, obvious assertions, no clever abstractions.

See review-design for the underlying Pragmatic Programmer principles (design by contract, pragmatic paranoia).

Test Types

Golden File Pattern (Go)

var update = flag.Bool("update", false, "update golden files")

func TestGolden(t *testing.T) {
    got := generate(input)
    golden := filepath.Join("testdata", "golden", name)
    if *update {
        os.WriteFile(golden, got, 0644)
        return
    }
    want, _ := os.ReadFile(golden)
    if diff := cmp.Diff(string(want), string(got)); diff != "" {
        t.Errorf("mismatch (-want +got):\n%s", diff)
    }
}

Run go test -update ./... to regenerate, then commit the diffs.

Fuzz Pattern (Go)

func FuzzParseInput(f *testing.F) {
    // Seed corpus: valid, empty, edge cases
    f.Add([]byte(`{"name": "Jane"}`))
    f.Add([]byte(`{}`))
    f.Add([]byte(``))

    f.Fuzz(func(t *testing.T, data []byte) {
        // Should never panic; errors are fine
        _, _ = ParseInput(data)
    })
}

CI: go test -fuzz=FuzzParseInput -fuzztime=10s -timeout=60s ./...

Benchmark Pattern (Rust)

use criterion::{criterion_group, criterion_main, Criterion};

fn bench_transform(c: &mut Criterion) {
    let input = load_fixture();
    c.bench_function("transform", |b| {
        b.iter(|| transform(&input))
    });
}

criterion_group!(benches, bench_transform);
criterion_main!(benches);

Place in benches/ directory. Run with cargo bench.

Per-Language Conventions

Rust

Fixtures: tempfile::TempDir for filesystem tests (drops on scope exit). include_str!() for static test data.

Go

Fixtures: t.TempDir() for temp directories (auto-cleanup). testdata/ for static files (ignored by Go toolchain).

Python

[tool.pytest.ini_options]
testpaths = ["tests"]
pythonpath = ["src"]

TypeScript

File Organization

Test helpers go in the test file, unexported. Do not create shared testutils/ packages. The duplication cost is lower than the coupling cost.

Fixtures & Mocks

Fixtures

Mocking Rules

• Prefer real implementations. Use TempDir and real git commands over git mocks. Use real HTTP servers over fetch mocks when practical.
• Mock at boundaries. Only mock external services (APIs, databases) and only at the interface boundary.
• Never mock what you own. If you need to mock your own code, the design needs refactoring. Extract an interface.
• Go: Use interfaces for test doubles. No mocking framework needed.
• Rust: Use trait objects or generic type parameters for test substitution.
• TypeScript: vi.fn() and vi.mock() for external dependencies only.

CI Strategy

All test types run with just check locally. CI mirrors just check exactly. No CI-only test logic.

What NOT to Test

• Third-party behavior don’t test that serde serializes correctly or that os.MkdirAll creates directories
• Private implementation details if you need to export something just for testing, the boundary is wrong
• Generated code oag generates TypeScript clients; test the generator, not the output
• Trivial accessors a getter that returns a field does not need a test
• Implementation mirrors if your test duplicates the logic it tests, it proves nothing
• Exact error messages test error types or categories, not wording (it changes)

Gotchas

• Float comparison: Never assert_eq!(f64, f64). Use an epsilon: assert!((a - b).abs() < 1e-10)
• Go parallel + shared state: t.Parallel() runs subtests concurrently. Shared fixtures must be immutable or use sync.Mutex.
• Python src/ discovery: Without pythonpath = ["src"] in pytest config, imports fail. Always configure this in pyproject.toml.
• Rust integration tests: Each file in tests/ compiles as a separate binary. Group related tests in one file to reduce compile time.
• Go golden file diffs: Use go-cmp for readable diffs instead of reflect.DeepEqual. The error messages are vastly better.
• Flaky tests: If a test fails intermittently, it’s a design problem (shared state, timing, network). Fix the root cause; do not retry.