Модули с ленивой загрузкой и переключатели функций

By default, NestJS instantiates every module in your imports graph at bootstrap. For an enterprise API with dozens of optional features — a PDF exporter, a payment gateway, an AI scoring engine — that means paying the full provider-instantiation and connection-warmup cost before the app even accepts a request.

• Heavy SDKs (Stripe, AWS, gRPC clients) run their constructors eagerly.
• Modules a given deployment never uses still load.
• Cold-start latency grows linearly with the module graph.

The fix: load select feature modules lazily, only when first invoked.

Nest ships a built-in LazyModuleLoader (from @nestjs/core). You inject it like any provider, then call load() with a factory that returns the module. Nest registers the module's providers on demand and caches the resulting reference for subsequent calls.

Key traits:

• Lazy modules are not listed in any imports array.
• They do not register controllers, resolvers, or enhancers — only providers.
• The first load() instantiates; later calls return the cached ModuleRef.

The object returned by load() is a ModuleRef. Use its get() method to pull a concrete provider out of the freshly-loaded module. Because lazy modules are isolated, request a provider only after the module is loaded.

For request-scoped or transient providers use moduleRef.resolve() instead of get().

The real startup win comes from pairing LazyModuleLoader.load() with a dynamic import(). A static top-level import pulls the module — and its heavy transitive deps — into the bootstrap bundle. A dynamic import() defers that file evaluation until the call runs.

• Static import { PdfModule } at the top of the file = loaded at startup.
• await import('./pdf/pdf.module') inside the method = loaded on first use.

So always import the lazy module's file dynamically, never statically.

The lazy module itself is an ordinary @Module — there is nothing special in its decorator. What makes it lazy is purely how it is consumed (via LazyModuleLoader rather than an imports array).

Keep its providers self-contained so loading it does not drag in the whole app.

Nest internally keeps a registry of already-loaded lazy modules keyed by the factory result. So calling load() repeatedly with the same module class is effectively free after the first hit — no double instantiation, no duplicate connections.

This means you can safely call load() inline in a hot handler without guarding it yourself; the framework deduplicates. The one-time cost is paid on the first request that needs the feature.

Feature toggles and lazy loading are a natural pair. Instead of conditionally registering modules at compile time, you check a flag at runtime and only load() the module when the flag is on. A disabled feature then costs nothing — not even its constructor.

• Flag from env, config service, or a remote flag provider (LaunchDarkly, Unleash).
• If the toggle is off, short-circuit before importing.

A feature-flag check is just a deterministic lookup. Here is a framework-free version you can reason about and test in isolation — the same logic a Nest FeatureFlagService would wrap. It reads a flag map and falls back to a default when the key is unknown.

A critical limitation: when a module is loaded lazily, Nest registers its providers only. It deliberately skips:

• controllers — no new HTTP routes appear.
• Global guards, interceptors, pipes, filters declared in the module.
• GraphQL resolvers.

So a lazy module cannot add endpoints. Expose the feature through a controller in an eagerly-loaded module that delegates to the lazily-loaded provider.

Lazy loading trades a one-time first-request latency spike for a faster, lighter startup. That is the right deal for rarely used, expensive features. For features on the hot path, eager loading (or an explicit warm-up on onApplicationBootstrap) avoids penalizing the first user.

Decision guide:

• Lazy: heavy SDK, used by <X% of requests, optional per deployment.
• Eager: core domain, every request, latency-sensitive.
• Lazy + warm-up: heavy but predictably needed soon after boot.

Quantify before and after. Wrap bootstrap timing and the first lazy load() to confirm the trade-off is real for your workload.

• Startup time should drop by the cumulative constructor + connection cost of the deferred modules.
• First-call latency for the lazy feature absorbs that cost once.
• Watch P99 of the first request after deploy — that is where the deferred cost surfaces.

If the lazy feature is hit on nearly every request, the numbers will tell you to switch it back to eager.

You learned how to shrink NestJS startup cost with on-demand modules:

• LazyModuleLoader.load(() => SomeModule) instantiates a module's providers on first use and caches the result.
• Pair it with a dynamic import() so the module's file (and heavy deps) is never evaluated at bootstrap.
• Resolve providers via moduleRef.get() (or resolve() for scoped providers).
• Feature toggles gate the load: a disabled feature costs nothing, not even a constructor.
• Lazy modules register no controllers, resolvers, or enhancers — delegate from an eager controller.
• Choose lazy for heavy, rarely-used, optional features; eager (or lazy + warm-up) for hot-path code. Measure startup and first-call latency to confirm the trade-off.