app.module Bootstrap Smoke Tests (4 NestJS services)

SummaryExtended the Sprint 195 lesson (submission CacheModule circular dependency throwing 26ms into bootstrap → CrashLoopBackOff, undetected because the 14 unit specs assembled providers manually) from a single module to the entire AppModule. Introduced Test.createTestingModule({imports:[AppModule]}).compile() bootstrap smoke tests across 4 NestJS services (gateway·submission·problem·identity) to catch DI-graph throws (circular dependencies, missing providers) at bootstrap-equivalent level. TypeORM real DB connection blocked via getDataSourceToken() override of initialize() (problem also overrides the NEW_DB_CONNECTION named connection); Redis blocked via REDIS_CLIENT override (submission/problem) and jest.mock('ioredis') for gateway (many constructors call new Redis() directly); RabbitMQ avoided by calling .compile() only (not .init()), skipping onModuleInit. Single it · zero branches · no uninvoked function literals keep coverage thresholds intact for spec-collected services (problem/identity/gateway). Negative check (empty DataSource → repo factory TypeError) proves the graph is genuinely compiled. Critic (Codex gpt-5.x) ran 2 rounds — R1 P2 (problem DUAL_WRITE_MODE env leak) resolved → clean R2. 4 specs pass standalone + full regression, tsc 0, ESLint 0, CI #347 Passed 412/Failed 0. PR #347 squash → 5d1e5fc.

Date2026-05-22

ImpactCritical

Goal

Introduce bootstrap smoke tests that compile the entire DI graph in one shot, catching Sprint 195-class circular dependency / module-graph throws at bootstrap-equivalent level.
Extend the Sprint 195 hotfix's cache.module.spec.ts (single-module compile regression) pattern to the entire AppModule, making "tests green = bootstrap possible" hold for every NestJS service.

Background

In Sprint 195, new submission code threw circular dependency inside CacheModule 26ms into bootstrap, causing CrashLoopBackOff. Yet the 14 existing unit specs assembled providers manually (providers:[...]), so they never compiled the real module DI graph — bootstrap failed 100% even though "tests were green."
The Sprint 195 hotfix added a Test.createTestingModule({imports:[CacheModule]}).compile() regression test, but it was single-module scope → other cycles/omissions in the full AppModule graph remained undefended.
Target NestJS services: gateway·submission·problem·identity. (github-worker is plain Node.js with no app.module; ai-analysis is FastAPI — both excluded.)
Technical challenge: .compile() runs graph build + useFactory (TypeORM/Throttler, etc.), so providers that attempt real infra (Postgres/Redis/RabbitMQ) connections must be blocked to validate only the DI graph without networking.

Decision

D1. Scope — all 4 NestJS services (user, AskUserQuestion)

Add src/app.module.spec.ts to each of gateway·submission·problem·identity. Every NestJS service guarantees "tests green = bootstrap possible" → Sprint 195-class graph throws blocked across all services.

D2. Call `.compile()` only (`.init()` out of scope)

.compile() builds and instantiates the DI graph but does not call onModuleInit/onApplicationBootstrap (those are .init()-only). Thus MqPublisherService's amqplib.connect (onModuleInit) does not run → RabbitMQ connection avoided. But circular dependencies / missing providers throw during graph build (compile), so they are still caught (same rationale as the Sprint 195 cache.module.spec).
Lifecycle-hook (onModuleInit) verification + amqplib mock are deferred to a separate scope (future sprint).

D3. Infra-blocking mechanism (verified against @nestjs/[email protected] source)

Infra	Blocking method
TypeORM real DB connection	`.overrideProvider(getDataSourceToken()).useValue(mockDataSource)` — replaces TypeOrmCoreModule's dataSourceProvider (whose useFactory calls `dataSource.initialize()`) wholesale → initialize never runs. The mock only needs the members the repository factory references: `entityMetadatas:[]`·`options.type`·`getRepository`.
Redis (`new Redis()`)	submission/problem: `.overrideProvider(REDIS_CLIENT)`. gateway: `jest.mock('ioredis')` — many constructors (~10, including throttler) call `new Redis()` directly at `.compile()` time; this is not via DI, so overrideProvider cannot intercept it and a module-level mock is required.
RabbitMQ (`amqplib.connect`)	`.compile()` only (D2).

D4. problem dual connection — companion named DataSource override

DualWriteModule adds a NEW_DB_CONNECTION='new-problem-db' named connection beyond the default, so getDataSourceToken(NEW_DB_CONNECTION) (='new-problem-dbDataSource') is also overridden. DUAL_WRITE_MODE=off is set explicitly (D6) so only the OFF branch is evaluated → no NEW_DATABASE_* needed.

D5. No coverage impact on spec-collected services (problem/identity/gateway)

Only submission excludes specs from coverage (!**/*.spec.ts in jest.config.ts). problem/identity/gateway collect specs too, so the new specs are written with a single it, zero branches, and no uninvoked function literals (bare jest.fn(); () => ... literals only where actually invoked at compile) to avoid affecting function/branch thresholds (95~98%).

D6. Test env is set explicitly, not preserved via spread (Critic P2)

beforeAll sets process.env; afterAll restores via spread. This satisfies constructor/factory-time getOrThrow (submission: DATABASE_·INTERNAL_KEY_AI_ANALYSIS; gateway: JWT_SECRET·OAUTH_CALLBACK_URL·INTERNAL_KEY_). problem sets DUAL_WRITE_MODE='off' explicitly — a spread alone would preserve an expand·switch-read value from the environment/local .env, leaking into the active branch (caught in Critic R1).

Implementation

Implementation commits (2 commits, PR #347 squash → `5d1e5fc`)

1c5f126 test — add 4 AppModule bootstrap smoke tests
- New services/{gateway,submission,problem,identity}/src/app.module.spec.ts
- submission: override getDataSourceToken()·getEntityManagerToken()·REDIS_CLIENT (./cache/cache.constants)
- problem: default + NEW_DB_CONNECTION (./database/dual-write.config) named DataSource/EntityManager + REDIS_CLIENT (./cache/cache.module) overrides
- identity: DataSource/EntityManager override only (no module-level Redis; NODE_ENV='test' skips TokenEncryptionService key validation)
- gateway: top-of-file jest.mock('ioredis') (default-import constructor mock, chainable .on()) + no TypeORM
- common: single it + one compile() (avoids MetricsModule/prom-client duplicate registration) + expect(moduleRef).toBeDefined() + no close() (on teardown, TypeOrmCoreModule.onApplicationShutdown re-resolves the overridden DataSource and fails; with mocked infra there are no real handles to clean up)
8fcd15e test — [Critic R1 P2] set DUAL_WRITE_MODE='off' explicitly in problem smoke (hermetic env)

Verification

Type/build: tsc --noEmit 0 (4 services). ESLint 0 (4 services; spec override turns off no-explicit-any).
Tests: 4 specs pass standalone + full regression. Per-service coverage thresholds pass (jest --coverage exit 0 — submission/problem/identity/gateway). Zero open handles/hangs (jest exits cleanly; gateway throttler setInterval is .unref()).
Negative check: temporarily swapped submission's DataSource override for an empty object → the repo factory's dataSource.entityMetadatas.find(...) failed compile() with TypeError: Cannot read properties of undefined (reading 'find') → proves the test genuinely compiles the graph and the mock is load-bearing (reverted).
Critic: codex review --base main (gpt-5.x) 2 rounds — Critical/High 0. R1 P2 (problem DUAL_WRITE_MODE env leak) resolved → R2 0 ("no discrete introduced issue that would break existing behavior or tests"). ✅ mergeable.
CI #347: Passed 412 / Failed 0 / MERGEABLE·CLEAN → Squash merge.

Lessons Learned

① .compile() runs graph build + useFactory but not onModuleInit — exploiting this boundary yields a smoke test that avoids lifecycle connections (e.g., RabbitMQ) while still catching circular dependencies / missing providers. A direct application of the Sprint 195 lesson that unit specs which don't compile the module DI graph (manual provider assembly) cannot catch cycles.
② Override only intercepts DI-routed providers — TypeORM DataSource/REDIS_CLIENT are providers and can be blocked via overrideProvider, but when a factory/constructor calls new Redis() directly (gateway), it is not DI-routed, so overrideProvider cannot stop it and a jest.mock('ioredis') (module-level mock) is required.
③ Test env should be set explicitly, not preserved via spread — an ORIGINAL_ENV spread preserves branch-deciding env (DUAL_WRITE_MODE, etc.) from the environment/local .env, so env that determines branch behavior must be set explicitly to keep the test hermetic (Critic catch).
④ For spec-collected services, uninvoked function literals erode function coverage — in services that don't exclude specs from coverage, a mock's jest.fn(() => ...) literal becomes an uncovered function if not invoked at compile. Avoid via bare jest.fn() + a single compile.

New Patterns

AppModule bootstrap smoke test — Test.createTestingModule({imports:[AppModule]}).overrideProvider(getDataSourceToken()).useValue(mockDataSource)...compile(). Override infra providers (DataSource·REDIS_CLIENT), module-mock directly-newed clients (gateway ioredis), avoid RabbitMQ by limiting to .compile(). New NestJS services/modules validate the bootstrap graph once with this smoke test ("tests green = bootstrap possible").
Minimal TypeORM DataSource mock — { entityMetadatas:[], options:{type:'postgres'}, getRepository: jest.fn(()=>mockRepository), manager:{...} }. Just the members the repository provider factory references resolve the graph. Named connections override getDataSourceToken(name) separately.