아키텍처

In plain words. Mimi turns YouTube clips into English speaking practice. The interesting engineering is in the backend: when the AI analyzes a clip — which takes several seconds — that work is handed to a background worker so it never ties up the database, like a restaurant taking your order at the counter and sending the slow cooking to the back so the line keeps moving. It also tries a free AI provider first and automatically falls back to paid ones only if that one fails, and every user's data is walled off so one account can never see another's.

Extracted from the shadow-ai repo source on 2026-06-20, checked against the code and cross-checked by a second adversarial pass. Numbers are measured, not estimated.

The shape of the system

Mimi is a YouTube-to-SRS English-shadowing app. The interesting part is the backend: a single Spring Boot / Java 21 monolith organized into per-feature DDD modules (analysis, auth, billing, clip, deck, library, practice, recording, review, video, plus common), backed by PostgreSQL with Flyway migrations and deployed to AWS ECS Fargate behind an ALB via Terraform. The Next.js frontend and React Native mobile app are separate workspaces; this writeup covers the backend only, where the work that's worth explaining lives — an async LLM analysis pipeline, repository-layer multi-tenant isolation, and env-gated provider/storage swaps.

Surface	Size
Production Java	8,892 LOC, 172 files
Test code	3,009 LOC, 35 files, 121 @Test methods
REST surface	15 controllers, 49 endpoints (20 GET / 20 POST / 5 DELETE / 4 PATCH)
Data model	14 JPA @Entity classes, 13 CREATE TABLE, 20 Flyway migrations (226 LOC)
Multi-tenant scoping	25 declared userId-scoped repo finder/count/exists/delete signatures
Transactions	51 @Transactional methods
Terraform infra	1,125 LOC; ECS cluster + service + task def, ALB, 2-subnet VPC, 8 Secrets Manager secrets, ECR, GitHub OIDC

Request path

Every authenticated request enters through JwtAuthenticationFilter, which parses the bearer token, validates the tv (token-version) claim against the DB, and lets CurrentUserArgumentResolver inject the AuthenticatedUser into the controller via the @CurrentUser parameter. From there the service issues a userId-scoped query. The async branch is separate: committing a clip fires an event that runs the LLM pipeline off-request on a bounded pool.

The async LLM pipeline is pinned off the DB connection

A clip commit fires a ClipCreatedEvent. The listener onClipCreated in ClipAnalysisService.java is annotated @TransactionalEventListener(AFTER_COMMIT) plus @Async, so it runs only after the writing transaction commits, on a separate thread. The pipeline then deliberately splits its work across two short transactions with the network call in the gap: prepareAnalysis writes a PENDING row in a short @Transactional method, the seconds-long provider HTTP call happens OUTSIDE any transaction, and completeAsReady / completeAsFailed write the terminal READY / FAILED state in another short transaction.

Two details make this work. Self-invocation is routed through an ObjectProvider<ClipAnalysisService> self-proxy (self()) so the @Async / @Transactional advice actually fires — a plain this.method() call would bypass the proxy and silently run synchronously in the same transaction. The provider call is also wrapped in a Micrometer Timer.Sample that stops on the timer tubeshadow.ai.analysis, tagged by model and outcome, exposed at /actuator/prometheus.

The trade-off being bought: an LLM call can take seconds, and holding a Hikari connection for that whole window would exhaust the pool under any real load. Bookending the network call with two short transactions keeps connection-hold time minimal while still durably recording each state transition.

Bounded async pool with CallerRuns back-pressure

AsyncConfig.java defines a ThreadPoolTaskExecutor named taskExecutor with core 2, max 4, queue 50, a CallerRunsPolicy rejection handler, and waitForTasksToCompleteOnShutdown=true with awaitTerminationSeconds=30, paired with server.shutdown: graceful in application.yml. AsyncConfig implements AsyncConfigurer and overrides getAsyncExecutor() to return that same singleton, so framework-managed @Async work and explicitly-named executor work share one pool rather than silently diverging.

Spring's default @Async executor is SimpleAsyncTaskExecutor, which spawns a fresh unbounded thread per task — a burst of clip imports would walk memory off a cliff. The bounded pool with CallerRunsPolicy degrades differently: once the queue fills, the submitting thread runs the task itself, which slows intake instead of dropping work or OOMing. The graceful-shutdown settings drain in-flight analyses on SIGTERM, which matters for clean rolling ECS deploys. The cost of this choice shows up below in the warts: max 4 threads is a hard ceiling.

Multi-provider AI fallback chain

CompositeAiClient.java is the @Primary AiAnalysisClient. It injects every other AiAnalysisClient bean (excluding itself), orders them by the tubeshadow.ai.order property (default gemini,openai,claude), filters to providers whose API key is set via isConfigured(), and tries each in turn — falling through to the next on any RuntimeException, rethrowing the last if no fallback remains. Each concrete provider wraps its own HTTP call in AiRetry.withRetry from AiRetry.java: 3 attempts max, linear backoff of 500ms * attempt, retrying ONLY HttpClientErrorException.TooManyRequests (429), HttpServerErrorException (5xx), and ResourceAccessException (IO timeout), and failing fast on any other 4xx. There is no Resilience4j or Spring Retry dependency; both the fallback ordering and the retry loop are hand-rolled.

The shape this is buying: a free-tier Gemini key can serve everyday traffic and spill over to paid OpenAI / Claude only on quota exhaustion or outage. Distinguishing transient failures (retry) from permanent ones (fail fast on a 401/400) means a bad key fails immediately instead of burning 3 retries plus the fallthrough.

Repository-layer multi-tenant isolation as the chokepoint

Auth is stateless HS256 JWT. After JwtAuthenticationFilter validates the token and CurrentUserArgumentResolver injects the AuthenticatedUser (used across 11 controllers via @CurrentUser), isolation is enforced one layer down, in the repositories. Tenant-scoped finders carry the userId in their signature — 25 declared finder/count/exists/delete signatures such as findByIdAndUserId, findByUserId, countByUserId, including a raw @Query search in ClipRepository.java with WHERE c.user_id = :userId.

The reason to push the tenant predicate into the query rather than doing a global findById followed by an ownership check is failure mode. A cross-tenant id passed to findByIdAndUserId returns empty and the caller 404s — there's no row to leak. A forgotten ownership check on a global find, by contrast, leaks. Making isolation the default shape of the query means the fail-safe state is "no data" rather than "wrong tenant's data."

JWT revocation via token-version, with a fail-fast secret guard

Tokens carry a tv (token-version) claim. On every authenticated request JwtAuthenticationFilter does one indexed PK lookup (findTokenVersionById, line 41) and rejects the token if the stored version no longer matches the claim. A password change bumps the version — the column was added in migration V14__user_token_version.sql with default 0 — which invalidates every previously-issued token at once. JwtTokenProvider.java additionally refuses to boot under the prod profile if the signing secret still starts with the public dev placeholder dev-only-secret, and requires at least 32 bytes for HS256.

Stateless JWTs normally can't be revoked before they expire; a single indexed version column buys logout-everywhere and password-change revocation cheaply. The prod secret guard catches a failure the length check alone would miss: a 32-byte-but-repo-public key would pass a length check and still be forgeable. The cost is one DB round-trip per request — covered in the warts.

Env-gated storage and provider swaps

RecordingStorage is an interface with two implementations, S3RecordingStorage.java and LocalRecordingStorage.java, selected by @ConditionalOnProperty(tubeshadow.recording.storage = s3 | local, matchIfMissing local). The S3 path returns a ResponseInputStream straight from s3.getObject rather than buffering the whole object onto the heap, and mirrors the local backend's userId/clipId/uuid-name key layout so the two are addressing-compatible. The same @ConditionalOnProperty gating covers the auth rate limiter (AuthRateLimitFilter) and the curated-video seeder (CuratedCollectionSeeder).

The point is that one artifact runs on a laptop against local disk and in ECS against S3 by flipping a single property — no code branch, no AWS credentials needed for local dev, and a prod storage-backend swap with no rebuild.

SM-2 spaced repetition kept pure

Sm2Calculator.java is a Clock-injected SuperMemo-2 implementation: easiness floor MIN_EASINESS = 1.3, an Anki-style 0..5 quality mapping, repetitions and interval reset on a quality < 3 lapse, returning an immutable Sm2Result(easiness, intervalDays, repetitions, nextDueDate). It is deliberately separated from the JPA ReviewItem entity so Sm2CalculatorTest.java can exercise it without Spring or a database.

Keeping the scheduling math referentially transparent and Clock-driven makes it deterministically testable — feed a fixed clock and a quality score, assert the next due date — and decoupled from persistence concerns entirely.

Constant-time billing webhook and OIDC deploy

The server-to-server billing webhook (/api/billing/webhook) is permit-listed past JWT in SecurityConfig.java (permitAll), but BillingController.java verifies an X-Billing-Secret header with MessageDigest.isEqual (constant-time) and fails CLOSED — 503 BILLING_NOT_CONFIGURED — if no secret is configured. Constant-time compare avoids a timing oracle on the entitlement secret; failing closed prevents an unconfigured deploy from silently accepting plan writes. Infra in iam.tf and ecs.tf runs the Fargate service behind an ALB in a 2-subnet VPC with 8 Secrets Manager secrets, and a GitHub OIDC provider (aws_iam_openid_connect_provider) lets CI assume a deploy role without long-lived AWS keys. CORS in the app fails fast with an IllegalStateException if CORS_ALLOWED_ORIGINS is empty under the prod profile.

What this architecture doesn't do

• The auth rate limiter is in-memory and single-instance. AuthRateLimitFilter keeps a plain ConcurrentHashMap<ip, Bucket> (line 44, populated via computeIfAbsent at line 74) with no eviction path at all — no @Scheduled, no remove, clear, or evict. The map grows unbounded with distinct client IPs, and because the limit isn't shared across tasks, running more than one ECS task multiplies the effective rate limit.
• Stateless JWT here isn't actually zero-DB. The tv revocation check means JwtAuthenticationFilter does a findTokenVersionById on every authenticated request (line 41). It's an indexed PK lookup, but there's no cache layer, so the usual "JWT = no DB hit" property doesn't hold.
• Fallback is exception-only and can stack up. CompositeAiClient only advances to the next provider on a thrown RuntimeException, and the inner AiRetry can burn up to 3 attempts × read-timeout per provider first. A multi-provider outage can occupy one async-pool thread for the summed retry time of every provider — and the pool max is only 4 threads, so a few simultaneous outages can stall the pipeline.
• Degraded outputs aren't distinguishable at a glance. completeWithEmptyTranscript (line 130) marks an analysis READY with a Transcript unavailable. placeholder body when the transcript is blank, while completeAsFailed (line 94) marks FAILED when no provider key is configured. Telling a real result from a degraded one means inspecting model/status, not reading the storage layer.
• Leftover scaffolding text. backend/build.gradle.kts line 31 still carries the comment // Persistence (will be activated when DB tasks land) directly above active spring-boot-starter-data-jpa and Flyway dependencies, with 20 migrations already in db/migration — the comment predates the work and was never removed.