Testing architecture and interview lab

Case study: offline checkout

1. Pure unit tests cover money rounding, validation, state reducer, idempotency policy, conflict resolution, and retry classification.
2. ViewModel component tests use fake repositories/clock/dispatchers and verify state/event/cancellation/recreation.
3. Room device tests cover transactions, tenant scoping, ordering, Flow invalidation, and every supported migration path.
4. MockWebServer/contract tests cover HTTP headers/schema/errors/timeouts and ambiguous post-commit disconnect.
5. Worker tests cover input/result/backoff; scheduler tests cover unique work/constraints/tags.
6. Compose screen tests render state matrix and semantic user intents; screenshot/accessibility tests cover representative configurations.
7. Few application/E2E tests prove DI/navigation/auth and one sandbox checkout.
8. Macrobenchmarks protect startup/checkout; production metrics and staged rollout cover real-world gaps.

Failure-injection checklist

• duplicate tap and server success with lost response;
• malformed/unknown JSON and 401/403/429/500;
• rotating refresh concurrency;
• database full/transaction rollback/migration from oldest supported version;
• process death during pending checkout;
• account switch while work is queued;
• Flow subscriber absent/late and stale result after flatMapLatest;
• permission denied/revoked and offline/online transition;
• RTL/large font/dark/expanded window/TalkBack;
• slow emulator, animation, locale/time/random ordering;
• external sandbox unavailable;
• fake diverges from real adapter contract.

Interview questions

1. Pyramid or portfolio? Pyramid guides cost distribution; portfolio maps specific product risks to cheapest sufficient fidelity, including hardware/performance/security/production.
2. Unit versus integration? Isolated functional unit versus multiple real collaborators/boundary behavior; define team terminology explicitly.
3. Why not E2E everything? Slow, brittle, difficult diagnosis/data/environment; move decisions to narrow tests and keep critical wiring journeys.
4. When fake versus mock? Stateful reusable behavior versus verifying an interaction contract; prefer observable outcomes unless call itself matters.
5. Why can fake be dangerous? It may omit SQL/server/scheduler semantics; run shared contracts and real adapter tests.
6. What makes concurrency testable? injected scheduler/dispatcher/clock, structured scopes, explicit state/events, no sleeps/global work.
7. runTest benefit? structured test scope, shared scheduler, virtual delays, child exception/leak handling.
8. Why can StateFlow test hang? Infinite collection is a child of runTest; use finite operators or cancellable backgroundScope collector.
9. Why subscribe to stateIn(WhileSubscribed) in test? Upstream may remain inactive without collector even when reading current value.
10. How test Room migration? Export schemas, create each supported old version with representative data, migrate/validate schema and semantic data/indexes.
11. MockWebServer value? Verifies actual HTTP request/serialization/error/timeout behavior without an external environment.
12. Espresso versus UI Automator? In-app synchronized View interaction versus cross-app/system UI automation.
13. Robolectric trade-off? Fast host framework simulation; not full device/OEM/render/Binder/hardware fidelity.
14. Why semantics-based Compose tests? Stable user meaning and accessibility contract rather than layout implementation.
15. Screenshot limitation? Pixel fidelity depends on environment and cannot prove behavior/accessibility.
16. How fix flakes? Preserve evidence, classify nondeterminism, control time/state/environment, semantic waits, cleanup; retries/quarantine are managed temporary tools.
17. Coverage limitation? Execution does not prove assertions or requirements; use for gaps, combined with risk/quality evidence.
18. Mutation testing? Deliberately changes code to see whether tests detect behavioral faults; useful selectively for high-risk pure logic.

Review checklist

• Risks map to explicit test layers/environments and no layer duplicates blindly.
• Time, IDs, dispatchers, external effects, and state are controllable.
• Doubles model important semantics and share contracts with real adapters.
• Coroutines/Flow tests use one scheduler, virtual time, finite/background collection, and cancellation assertions.
• Room/API/Worker/migration tests cover real boundary failure semantics.
• UI tests use semantics/observable idleness; system UI uses appropriate cross-app tool.
• Compose state/configuration/accessibility/screenshot matrices are intentional.
• Flakes/quarantines have evidence, owner, expiry, and metric.
• Coverage/mutation/performance/security reports complement—not replace—requirements.
• CI tiers are fast, diagnosable, sharded, artifact-rich, and production-connected.

Exercises

1. Produce a test strategy ADR for a payments app with suite budgets and release gates.
2. Convert a sleep-based UI test to deterministic fake server + semantic idling.
3. Add a mutation run to the money module and analyze five meaningful survivors.