Résolution du locataire via le middleware et AsyncLocalStorage

In a multi-tenant backend, one running instance serves many customers (tenants). Every request belongs to exactly one tenant, and almost every layer needs to know which one: the database connection, query filters, caches, audit logs, even outbound emails.

• Passing a tenantId argument down through every service method is noisy and error-prone.
• Forget it once in a query and you leak tenant A's data to tenant B — a critical security bug.

We need to resolve the tenant once at the edge of the request and propagate it implicitly to everything that runs during that request.

Resolution strategy depends on your routing model. Common sources, in order of how early they are available:

• Subdomain: acme.api.example.com → tenant acme.
• Header: X-Tenant-Id: acme (common for internal/service-to-service calls).
• JWT claim: a tid claim inside the access token.
• Path prefix: /t/acme/orders.

Whatever the source, the goal is to turn raw request data into a validated tenant identifier as early as possible — ideally in middleware, which runs before guards, interceptors, and controllers.

A naive approach is to attach the tenant to req and read req.tenantId everywhere. This works in controllers and guards but breaks down quickly:

• Deep services would need the Request injected (@Inject(REQUEST)), forcing them to be request-scoped, which is contagious and hurts performance.
• Code with no access to req — a repository helper, a logger formatter, a TypeORM subscriber — has no clean way to read it.

We want a way to ask "what tenant am I serving right now?" from anywhere in the async call stack, without threading req through. That is exactly what AsyncLocalStorage provides.

AsyncLocalStorage (from Node's built-in async_hooks) gives you a per-request store that survives across await boundaries, timers, and callbacks. You call als.run(store, callback) once, and any code executed inside that callback — however deep, however async — can call als.getStore() to read the same store.

It is conceptually thread-local storage for Node's single-threaded async model. Each request gets its own isolated store; concurrent requests never see each other's data.

Wrap AsyncLocalStorage in a small, typed service so the rest of the app never touches Node internals directly. Keep the store shape minimal but extensible — at least the tenant id, plus anything else you want request-scoped (request id, user id).

• run(store, cb) — enter a new context for one request.
• get() — read the current tenant, throwing if called outside any context (fail loud, not silently global).

Making get() throw on a missing context is deliberate: a silent undefined tenant is how cross-tenant leaks happen.

Middleware is the right place to resolve and open the context because it runs at the very start of the request, before guards and route handlers. The middleware:

• Extracts the raw tenant hint (header, subdomain, etc.).
• Validates it and rejects unknown tenants with 400/404.
• Calls tenantContext.run(...) and invokes next() inside the callback so the entire downstream pipeline executes within the context.

The critical detail: next() must be called inside run(). If you call next() after run() returns, the context is already closed and every getStore() downstream returns undefined.

Register the middleware in a module's configure method and apply it to all routes. Because TenantContext is injected into the middleware, both must be provided/exported from a module the app imports.

Order matters: the tenant middleware should run before anything that depends on the context (logging middleware, etc.). NestJS applies middleware in the order you chain .apply() calls and in module import order.

Now any provider — at any depth, request-scoped or not — can inject TenantContext and read the current tenant. No @Inject(REQUEST), no request-scoped contagion, no passing tenantId through method signatures.

This is the payoff: a singleton service can safely ask for the current tenant because getStore() resolves to the right per-request store at call time.

The biggest win is enforcing tenant isolation at the data layer instead of trusting every developer to add where tenantId = .... With the context available globally, you can centralize it:

• Connection-per-tenant: resolve a tenant-specific datasource/connection from a pool using tenantContext.tenantId.
• Shared schema + filter: a TypeORM/Prisma global filter or query subscriber that injects the tenant predicate.

Both approaches read the tenant from AsyncLocalStorage, so the resolver logic lives in exactly one place.

Why middleware and not a guard or interceptor for opening the context?

• Middleware runs first and wraps the entire remaining pipeline inside its next() callback — guards, interceptors, pipes, and the handler all execute within run(). This is what we want.
• Guards/interceptors run later. An interceptor wraps the handler but not the guards that ran before it, so context opened there is partially missing.

One nuance: if you must resolve the tenant from a verified JWT, the token is validated in a guard, which runs after middleware. A common pattern is middleware to open the store and a guard/early step to populate the tenant once the JWT is verified.

AsyncLocalStorage is robust across async/await and setTimeout, but a few things silently drop the context:

• Detached work: tasks pushed to a queue or run after the response (fire-and-forget jobs, cron triggered work) run outside the request and have no store. Capture tenantId explicitly before handing off.
• Some pools/native callbacks that use non-promise scheduling can lose context — verify, and re-bind with als.run if needed.
• Calling get() in app bootstrap or in a health check that bypasses the middleware throws — guard such paths or skip the middleware for them.

Rule of thumb: anything that outlives the HTTP request must be given the tenant explicitly, not via the store.

You implement tenant resolution in a NestJS middleware that calls tenantContext.run(store, ...). Deep services using a singleton (default-scoped) provider read the tenant via getStore().

You built implicit, leak-resistant tenant propagation:

• Resolve early: middleware extracts and validates the tenant from header/subdomain/JWT at the request edge.
• Propagate implicitly: wrap AsyncLocalStorage in a typed TenantContext service; call run(store, () => next()) so the whole pipeline executes inside the context.
• Consume anywhere: any singleton provider injects TenantContext and reads the tenant — no request-scoped contagion, no threading tenantId through signatures.
• Enforce isolation centrally at the data layer (connection-per-tenant or a global query filter).
• Mind the boundaries: detached/queued/scheduled work loses the store — capture tenantId explicitly, and make get() throw outside a request to fail loud.