Skip to content

injection.md

Latest commit

 

History

History
178 lines (134 loc) · 6.48 KB

injection.md

File metadata and controls

178 lines (134 loc) · 6.48 KB
Raw

Injection

When a scope has no provider for a requested key, it builds the value by constructor injection. This page covers the rules, the parameter shapes you can inject, and how lazy injection breaks cycles.

The one-constructor rule

An injectable class must have exactly one public constructor. The container inspects that constructor, resolves each parameter through the scope, and calls it.

public class Service {
    private final Repository repository;

    public Service(Repository repository) {
        this.repository = repository;
    }
}

Records work naturally — their canonical constructor is that single public constructor:

public record Service(Repository repository) {}

Supported parameter shapes

Each constructor parameter may take any of these forms:

TypeX value                                  // resolve a single TypeX
Provider<TypeX> lazyValue                     // resolve lazily, on first get()
Provider<Iterable<Provider<TypeX>>> all       // resolve all providers of TypeX
@Named("main") TypeX namedValue               // resolve a qualified TypeX
@Named("main") Provider<TypeX> namedLazyValue // qualified + lazy

Unsupported cases

The container will refuse (with UnsupportedInjectionException or a resolution error) to build:

  • local or anonymous classes;
  • non-static inner classes;
  • non-concrete generic types such as Provider<List<T>>;
  • a class without a single public constructor;
  • a direct, non-lazy cycle (see below).

Lazy injection

A Provider<T> parameter is resolved lazily: the dependency is not created until provider.get() is called.

record Controller(Provider<Service> service) {
    void handle() {
        service.get().run(); // Service is built here, not at Controller construction
    }
}

Use it to defer an expensive construction, or to break a cycle.

Cycles

A direct cycle cannot be constructed — neither side can be built first:

record C1(C2 c2) {}
record C2(C1 c1) {}

scope.get(C1.class); // BeanResolutionException

A lazy cycle through Provider<T> is allowed, because the provider defers construction of the other side:

record A1(Provider<A2> a2) {}
record A2(Provider<A1> a1) {}

A1 a1 = scope.get(A1.class);
A2 a2 = a1.a2().get();

assert a2.a1().get() == a1; // same scope singletons

The rule of thumb: if two types must reference each other, put a Provider<> on at least one side.

Where an auto-created bean lands

This is the part people miss. When you call get(T) / provider(T) and no provider for T exists anywhere in the visible graph — you never seed, provide or bind it — the container builds T by constructor injection and registers it in the current scope: the exact scope you called get on. The same goes for any of T's dependencies that are also unregistered — each missing one is created locally too. Dependencies that do exist in a parent are reused from there.

record RootScope() {}
record Config(String value) {}
record Service(Config config) {}

Scope<RootScope> root = new Scope<>(new RootScope());
root.seed(Config.class, new Config("prod"));

Service service = root.get(Service.class); // Service is registered nowhere

Mental model — Service materializes inside the scope that asked for it:

root.get(Service.class)

┌─ root : RootScope ─────────────────────────────────────────────────────────────┐
│                                                                                │
│  Config  = Config("prod")        (seeded)                                      │
│                                                                                │
│  Service = Service(config)   ◄── created HERE  the scope that called get()     │
│             └─ needs Config ─┐                                                 │
│                              └─► Config above  (reused, already local)         │
│                                                                                │
└────────────────────────────────────────────────────────────────────────────────┘

Why the current scope matters

The bean is owned by the scope that created it, so it dies when that scope closes — even if all of its dependencies came from a parent. Resolve from a child and the bean lives and dies with the child:

Scope<RootScope> root  = new Scope<>(new RootScope());
root.seed(Config.class, new Config("prod"));

Scope<Player> child = new Scope<>(new Player("Ada"));
child.ownedBy(root);

Service s = child.get(Service.class); // Service unregistered; Config lives in root
┌─ root ────────────────────────────┐
│                                   │
│  Config = Config("prod")          │   ← Config stays in root
│                                   │
└───────────────────────────────────┘
            ▲
            │ visible (ownedBy)
            │  reuses Config
            │
┌─ child : Player ──────────────────┐
│                                   │
│  Service = Service(config)        │   ← created in CHILD (the caller),
│            └─ Config from root    │      owned by child, gone on child.close()
│                                   │
└───────────────────────────────────┘

So where you call get from decides the bean's lifetime. If a type should be a singleton at the root, resolve it from the root (or bind/provide it there) — not from a short-lived child. This is the same shadowing model as everything else: a bean's dependencies see its scope first, then visible parents (Scopes & lifecycle).

Next: Qualifiers & collections.