microdi-js
v1.1.0
Published
A tiny DI container
Downloads
22
Readme
Basic usage | Concepts | Features
Microdi-js
Microdi-js is a tiny but feature rich dependency injection container for JavaScript.
Basic usage
Object Registration
You can register an object using one of three methods:
container.register(identifier, Item)
.This registers either a constructor function OR an object prototype
Item
using the given stringidentifier
.You can chain calls to alter the registration settings:
var identifier = 'itemKey1'; container .register('identifier', Item) .inject('otherDependencyIdentifier1', 'otherDependencyIdentifier1') .singletonPerContainer() .inGroup('mySimilarObjects');
Here we register
Item
using the stringidentifier
. We state that it requires dependenciesotherDependencyIdentifier1
andotherDependencyIdentifier1
. It's lifetime management issingletonPerContainer
. It can be resolved using theidentifier
or as part of the groupmySimilarObjects
.container.registerInstance(identifier, objectInstance)
.You can use
registerInstance
to register an existing instance with the given stringidentifier
.container.registerFactory(identifier, factory)
.This registers a creation factory using the given string
identifier
. The factory is a function that must return a new object instance. The factory will receive the current container as a first parameter. Any additional arguments passed during theresolve
call will be passed to the factory.container .registerFactory('fooId', (container, ...additionalDependencies) => { let fooDependency = container.resolve('fooDependencyId'); return new Foo(fooDependency, ...additionalDependencies); }) .transient(); let foo = container.resolve('fooId', 1, 2, 3);
Here we register a factory that will resolve a
Foo
. We then resolve an instance (foo
) passing some additional arguments. Our creation factory resolves an additional dependency, then passes this dependency, plus the additional dependencies1, 2, 3
toFoo
s constructor and returns it. It's registered using a transient scope so each time you resolve it, the factor will be invoked to resolve a newFoo
instance.
Object Resolution
Object resolution is done via :
container.resolve(identifier [, additionalDependency1, additionalDependency2, etc. ]);
This simply builds the object registered with the given
identifier
. Any dependencies will also be built and injected. Optionally provide additional dependencies which will get passed in after any that were specified viainject()
at registration time.container.resolveGroup(name);
Resolving a group returns an array of objects registered against that group.
Example
Below we have 2 simple classes.
Parent
takes Child
as it's dependency.
Child
is registered with the identifier 'child' and Parent
with the identifier 'parent'.
Additionally the parent registration injects the Child
as a dependency.
class Child {
sayHello() {
console.log('Hello from the child');
}
}
class Parent {
constructor(child) {
this._child = child;
}
sayHello() {
console.log('Hello from the parent');
this._child.sayHello();
}
}
var container = new microdi.Container();
container.register('child', Child);
container.register('parent', Parent).inject('child');
var parent = container.resolve('parent');
parent.sayHello();
Output:
keith@devshop:~/dev/personal/microdi-js/dist/examples$ node readme.js
Hello from the parent
Hello from the child
Concepts
Identifiers
JavaScript doesn't have a type system which makes containers a little cumbersome to use. Typically in typed languages you'd utilise information provided by the type system to aid in dependency resolution and injection. However without such a system all is not lost, we can simply use strings (i.e. 'identifiers') to identify objects to construct.
Dependencies
We need to manually specify objects dependencies at registration time.
This is because there are no means available to inspect arguments required for constructors
or init
methods at object instantiation/initialisation time.
We specify dependencies via inject
:
container.register('a', A)
.inject(dependency1, dependency2, etc.);`
Each item passed to inject
represents a dependency to be injected into the instance that will be build.
During resolve
dependencies will be passed to the new instance in the same order they were provided to inject
.
The arguments to inject
can contain string identifiers
referencing other registrations and/or resolverKey
s.
A resolverKey
references a dependency resolver which adds functionality to how the container will acquire and build the dependency.
Lets look at an example:
class Controller {
constructor(a, factoryThatMakesB) {}
}
container.register('a', A);
container.register('b', B);
container
.register('controller', Controller)
.inject('a', { resolver: 'factory', key : 'b' });
The above registers a Controller
using the identifier controller
.
It specifies Controller
requires 2 dependencies:
- The dependency registered as
a
. - The dependency
{ resolver: 'factory', key : 'b' }
, this is aresolverKey
telling the container to build the dependencyb
using the build in injection factory resolver.
Features
Object creation & dependency injection
A call to resolve
will trigger build up of the object in question.
Any dependencies the object requires will be resolved and injected.
Function constructors
If the type registered is a constructor function (i.e. typeof registeredObject === 'function') it will be initialised accordingly and any dependencies passed in.
Prototypical inheritance
If the type registered is not a constructor function it will be assumed a prototype.
At resolve time new object/s will be created using Object.create(registeredObject).
If the object has an init
method then this will be called passing any dependencies.
Lifetime management
An object’s lifetime can be controlled by the container in a number of ways.
Singleton
A singleton registration specifies the container will hold onto the object instance.
Multiple calls to resolve
with the same key will yield the same instance.
When the container is disposed, the container will, for any registered instances with a dispose
function, invoke that function.
var Foo = {};
container.register('theFoo', Foo).singleton();
var foo1 = container.resolve('theFoo');
var foo2 = container.resolve('theFoo');
console.log(foo1 == foo2); // true
Note this is the default lifetime used and so the call to .singleton()
is optional.
Singleton per container
This is similar to a singleton however you get a single instance per child container.
var Bar = {};
container.register('theBar', Bar).singletonPerContainer();
var bar1 = container.resolve('theBar');
var bar2 = container.resolve('theBar');
var childContainer = container.createChildContainer();
var bar3 = childContainer.resolve('theBar');
var bar4 = childContainer.resolve('theBar');
console.log(bar1 == bar2); // true
console.log(bar2 == bar3); // false
console.log(bar3 == bar4); // true
Transient
This creates a new instance each time a call to resolve(identifier)
or resolveGroup(groupName)
is made.
The container will not hold a reference to the instances created.
var Baz = {};
container.register('theBaz', Baz).transient();
var baz1 = container.resolve('theBaz');
var baz2 = container.resolve('theBaz');
console.log(baz1 == baz2); // false
External
Like a singleton however the container won't dispose the object when it's disposed.
var Disposable = {
init() {
this.isDisposed = false;
return this;
},
dispose() {
this.isDisposed = true;
}
};
// Note registerInstance internally sets the lifetime type as 'external',
// You can pass false as the last argument if you want to change this.
// The container will then manage it as 'singleton' and dispose the instance at disposal time
container.registerInstance('disposable1', Object.create(Disposable).init());
container.register('disposable2', Disposable);
var disposable1 = container.resolve('disposable1');
var disposable2 = container.resolve('disposable2');
container.dispose();
console.log(disposable1.isDisposed); // false
console.log(disposable2.isDisposed); // true
Resolve groups
You can group objects together at registration time then resolve them using resolveGroup(name)
.
Typically this is handy when you're objects share a related abstraction.
var Foo = {
name: "theFoo"
};
var Bar = {
name: "theBar"
};
var container = new microdi.Container();
container.register('foo', Foo).inGroup('group1');
container.register('bar', Bar).inGroup('group1');
var group1 = container.resolveGroup('group1');
for (let i = 0, len = group1.length; i < len; i++) {
console.log(group1[i].name);
}
Output:
theFoo
theBar
Injecting groups
If you want to inject a group simply register the injection using the groupName
.
From example, if you wanted to inject all dependencies in group group1
, from our example above, you'd do this:
class Bazz {
constructor(group1) {
// group1 would be an array with dependencies 'foo' and 'bar'
this._group1 = group1;
}
};
container.register('bazz', Bazz).inject('group1');
var bazz = container.resolveGroup('bazz');
Resolution with additional dependencies
When calling resolve
you can optionally pass additional dependencies.
These will be appended to the dependencies provided to inject
at registration time (if any).
class Foo {
constructor(fizz, bar, bazz) {
console.log('%s %s %s', fizz.name, bar.name, bazz.name);
}
}
var container = new microdi.Container();
container.register('fizz', { name: 'fizz'});
container.register('foo', Foo).inject('fizz');
var foo = container.resolve('foo', { name: 'bar'}, { name: 'bazz'});
Output
fizz bar bazz
Injection factories
Sometimes you want your object to receive a factory, that when called will resolve
and return the dependency in question.
class Item {
constructor() {
console.log("creating an item");
}
}
class Manager{
constructor(itemFactory) {
this._itemFactory = itemFactory;
}
createItem(name) {
return this._itemFactory(name);
}
}
var container = new microdi.Container();
container.register('item', Item).transient();
container.register('manager', Manager).inject({ resolver: 'factory', key: 'item'});
var manager = container.resolve('manager');
var item1 = manager.createItem();
var item2 = manager.createItem();
output:
creating an item
creating an item
Note
Injected factories are different from factories you use to create objects (via
container.registerFactory
). A factory registered viaregisterFactory
is simply used to create your instance. An injected factory lets the container create the instance but it wraps this creation in a factory and injects the factory. It's typically used for lazy resolution of objects or when an object needs to create many other objects but doesn't want to take a dependency on the container itself.
Additional dependencies in factories
You can pass arguments to the factory and they forwarded as discussed above.
Lets modify the sample from above to demonstrate this:
class Item {
constructor(otherDependencyA, name) {
console.log('Hello ' + name + '. Other dependency: ' + otherDependencyA);
}
}
class Manager{
constructor(itemFactory) {
this._itemFactory = itemFactory;
}
createItem(name) {
return this._itemFactory(name);
}
}
var container = new microdi.Container();
container.registerInstance('otherDependencyA', 'look! a string dependency');
container.register('item', Item).inject('otherDependencyA').transient();
container.register('manager', Manager).inject({ resolver: 'factory', key: 'item'});
var manager = container.resolve('manager');
var fooItem = manager.createItem('Foo');
var barItem = manager.createItem('Bar');
output:
Hello Bob
Hello Mick
Child containers
Child containers can be used to manage and scope a set of related objects. They inherit configuration from the parent however can be re-configured. They also inherit instances from their parent depending upon the lifetime management of the objects in question.
Create a child container by calling createChildContainer
on a parent;
var container = new microdi.Container();
var childcontainer = container.createChildContainer();
Lifetime management in child containers
Depending upon registration configurations, objects resolved from a child container will either be owned by the child container or the parent.
var Foo = { };
var container = new microdi.Container();
var childContainer = container.createChildContainer();
container.register('foo', Foo); // defaults to singleton registration
var foo1 = container.resolve('foo');
var foo2 = childContainer.resolve('foo');
console.log(foo1 == foo2); // true, same instance
container.register('fooAgain', Foo).singletonPerContainer();
var foo3 = container.resolve('fooAgain');
var foo4 = childContainer.resolve('fooAgain');
console.log(foo3 == foo4); // false, different instance
var foo5 = childContainer.resolve('fooAgain');
console.log(foo4 == foo5); // true, same instance
Overriding Registrations
The configuration of a child container can be overridden if required.
var Foo = { };
var container = new microdi.Container();
container.register('foo', Foo); // defaults to singleton registration
var childcontainer = container.createChildContainer();
childcontainer.register('foo', Foo).transient();
var foo1 = container.resolve('foo');
var foo2 = container.resolve('foo');
console.log(foo1 == foo2); // true, same instance
var foo3 = childcontainer.resolve('foo');
console.log(foo2 == foo3); // false, different instance
var foo4 = childcontainer.resolve('foo');
console.log(foo3 == foo4); // false, different instance
Output:
true
false
false
Resolving the container
Sometimes you have an object that requires the container to be injected. While often thought of as an anti pattern, there are scenarios where this makes sense. If the object is a 'bootstrapper' of sorts, i.e. it needs fine grained control over child containers and sub object graph resolution, injecting a container makes sense.
When there are child containers at play, you'd expect objects who require a container, to be provide with the same container that build them, i.e. the child container in question.
Lets look at an example:
// a class that takes a container as a dependency
class Bootstrapper {
constructor(container) {
this._container = container;
}
}
Lets look at the manual way of doing this:
import { Container } from 'microdi-js';
var container = new Container();
container.registerInstance('theRootContainer', container);
container.register('bootstrapper', Bootstrapper).inject('theRootContainer');
// bootstrapper will get injected with instance `container`
let bootstrapper = container.resolve('bootstrapper');
This will work, but when there are child container at play it gets a bit messy as you have to register things manually each time a child is created. Typically you strive to keep child container re-configuration to a minimum for code maintainability reasons.
microdi supports a special injection key exposed as MicroDiConsts.owningContainer
.
This key will ensure an object getting resolved get it's owning container.
Lets re-work the above example:
import { Container, MicroDiConsts } from 'microdi-js';
var container = new Container();
container
.register('bootstrapper', Bootstrapper)
.inject(MicroDiConsts.owningContainer);
// bootstrapper will get injected with instance `container`
let bootstrapper = container.resolve('bootstrapper');
Objects injected with MicroDiConsts.owningContainer
and resolved from child containers will get the owning child container injected:
import { Container, MicroDiConsts } from 'microdi-js';
var container1 = new Container();
container1
.register('bootstrapper', Bootstrapper)
.transient() // resolve new `Bootstrapper` instance each time
.inject(MicroDiConsts.owningContainer);
// bootstrapper1 will get injected with `container1`
let bootstrapper1 = container1.resolve('bootstrapper');
// bootstrapper2 will get injected with `container2`
let container2 = container1.createChildContainer();
let bootstrapper2 = container2.resolve('bootstrapper');
Checking what's registered
You can check if a dependency is registered using .isRegistered(name)
.
A group can be checked using .isGroupRegistered(groupName):boolean
.
Both these exist on a container
object and will return a boolean result.
Disposal
When you call .dispose()
on a container, that container will inspect any object it holds and if an object has a dispose
function, it will be called.
However it will not dispose transient-created objects or objects registered via registerInstance('aKey', myInstance)
.
Any child containers created from the disposed parent will also be disposed. These child containers will in turn dispose of their registered instances.
class Foo {
dispose() {
console.log('foo disposed');
}
}
var container = new microdi.Container();
container.register('foo', Foo).singletonPerContainer();
var foo1 = container.resolve('foo');
var childcontainer = container.createChildContainer();
var foo2 = childcontainer.resolve('foo');
container.dispose();
Output:
foo disposed
foo disposed
Dependency Resolvers
Dependency resolvers alters the default way a dependency is created.
A dependency resolver is simply an object with a resolve(container, resolverKey)
method.
You can create your own resolvers and add them to the container.
When registering an object, a resolverKey
can be used as a dependency.
This enables the container to resolve the dependency using the resolver specified.
A resolverKey
can can also be specified as a replacement for the object or construction function that is to be built.
At resolve time the container will call the dependency resolver specified by the resolverKey
to create the object in question.
The container
and the resolverKey
specified via inject()
will be passed to the resolvers resolve
method.
This sounds a bit more complicated than it actually is, it's easier to demonstrate with some code.
Here is an example where rather than registering a concrete object to build, a resolverKey
is used.
The DomResolver
resolver will be used to resolve the object.
class DomResolver {
resolve(container, resolverKey) {
// return a pretend dom element,
return {
get description() {
return 'Fake DOM element - ' + resolverKey.domId ;
}
};
}
}
var container = new microdi.Container();
container.addResolver('domResolver', new DomResolver());
// Note the usage of 'isResolverKey' so the container can distinguish this from a normal object.
// This is only required when you don't register a constructor function or prototype.
container.register('view', { resolver: 'domResolver', domId : 'theDomId', isResolverKey: true });
var view = container.resolve('view');
console.log(view.description);
Output:
Fake DOM element - theDomId
Here is an example where a concrete object is registered and a resolverKey is used to resolve a dependency of the registered item.
class DomResolver {
resolve(container, resolverKey) {
// return a pretend dom element,
return {
get description() {
return 'Fake DOM element - ' + resolverKey.domId ;
}
};
}
}
var container = new microdi.Container();
container.addResolver('domResolver', new DomResolver());
class Controller {
constructor(view) {
console.log(view.description);
}
}
// Note we don't need to specify the 'isResolverKey' property on the resolverKey.
// The container assumes it is as it appears in the dependency list.
container.register('controller', Controller).inject({ resolver: 'domResolver', domId : 'viewId' });
var controller = container.resolve('controller');
Output:
Fake DOM element - viewId
Built in resolvers
There are 3 built in resolvers.
Delegate
The delegate
resolver simply defers object creation to a delegate provided by the resolverKey
.
class Foo {
constructor(bar) {
console.log('bar is : [%s]', bar);
}
}
var container = new microdi.Container();
container.register('foo', Foo)
.inject(
{
resolver: 'delegate',
resolve: (container, resolveKey) => {
return 'barInstance';
}
});
var foo = container.resolve('foo');
Output:
bar is : [barInstance]
Injection factory
Discussed above, injection factories use the factory
dependency resolver to wrap the creation call in a function for later invocation.
External factory
Under the covers registerFactory
uses a built in externalFactory
dependency resolver to invoke the factory registered against an identifier.
Annotation usages and IoC
Often container implementation use decorators (annotations in Java and attributes in C#) to decorate objects with IoC related concerns. For example lifetime management, scoping, dependency information etc. The author believes this to be useful in some instances however often an anti-pattern for the following reasons:
- Objects become dependent on their dependencies as they are explicitly declared together.
- It becomes difficult to configure an object for differing usages unless you decorate it for both, if supported.
- It becomes difficult to gain an understanding of how objects play together at runtime. Runtime configuration is scattered on objects in different code files and across the entire codebase. Runtime configuration of objects is simpler to understand when it's in a single place.
- 90% of the time you'll find you need some runtime configuration. You end up in a place where you have both runtime and declarative (i.e. decorator) configuration. This adds confusion.
- It doesn't work when creating shared code, i.e. in libraries. If you hit this you don't want some code using annotation and some using run time configuration (i.e. the above point).