I'm considering migrating to Dagger 2 some libraries. This library expose a configurable client, each configuration can be named and later retrieved in a singleton fashion.
Let me show a pseudo-code of how the library works from the user perspective:
// initialization
ClientSDK clientA = new ClientSDK.Builder()
.configuration().attributes().here()
.apiKey("someString") // set api key / credentials
.build();
LibraryAuthenticationManager customAuthManager = new MyCustomAuthenticationManager();
ClientSDK clientB = new ClientSDK.Builder()
.configuration().attributes().here()
.apiKey("someStringElse")
.customAuthManager(customAuthManager) // override some default
.baseApiUrl("https://custom.domain.com/and/path") // override some default setting
.build();
ClientSDK.setSingleton("clientA", clientA);
ClientSDK.setSingleton("clientB", clientB);
And when I need an instance elsewhere:
// usage everywhere else
ClientSDK clientB = ClientSDK.singleton("clientB");
clientB.userManager(); // "singleton" using the configuration of clientB
clientB.subscriptionsManager(); // "singleton" using the configuration of clientB
clientB.currentCachedUser(); // for clientB
clientB.doSomething(); // action on this instance of the ClientSDK
ClientSDK instances are created by the user of the library and the ClientSDK statically keep a map of singletons associated to the name.
(The actual behavior of the SDK is slightly different: the naming is automatic and based on a mandatory configuration parameter.)
It's like I have lot of singleton classes with a single point of entry (the ClientSDK) but since I can have multiple configuration of the ClientSDK each with his own singletons instances this are not really singletons.
If I would try write a library like that with Dagger 2 I would do something like:
class ClientSDK {
#Inject SDKConfiguration configuration;
#Inject LibraryAuthenticationManager authManager;
...
}
The problem is that I need each instance of the ClientSDK to have its own configuration and authManager (and many other services) injected. And they need to be definable (the configuration) and overridable (the actual implementation) from the library user.
Can I do something like this with Dagger 2? How?
I've seen I can create custom Scopes but they are defined at compile time and the library user should be the one defining them.
(the library is an Android Library, but this shouldn't be relevant)
Thanks
It sounds like you should be creating stateful/configurable Module instances and then generating separate Components or Subcomponents for each ClientSDK you build.
public class ClientSDK {
#Inject SDKConfiguration configuration;
#Inject LibraryAuthenticationManager authManager;
// ...
public static class Builder {
// ...
public ClientSDK build() {
return DaggerClientSDKComponent.builder()
.configurationModule(new ConfigurationModule(
apiKey, customAuthManager, baseApiUrl)
.build()
.getClientSdk();
}
}
}
...where your ConfigurationModule is a #Module you write that takes all of those configuration parameters and makes them accessible through properly-qualified #Provides methods, your ClientSDKComponent is a #Component you define that refers to the ConfigurationModule (among others) and defines a #Component.Builder inner interface. The Builder is important because you're telling Dagger it can no longer use its modules statically, or through instances it creates itself: You have to call a constructor or otherwise procure an instance, which the Component can then consume to provide instances.
Dagger won't get into the business of saving your named singletons, but it doesn't need to: you can save them yourself in a static Map, or save the ClientSDKComponent instance as an entry point. For that matter, if you're comfortable letting go of some of the control of ClientSDK, you could even make ClientSDK itself the Component; however, I'd advise against it, because you'll have less control of the static methods you want, and will lose the opportunity to write arbitrary methods or throw exceptions as needed.
You don't have to worry yourself about scopes, unless you want to: Dagger 2 tracks scope lifetime via component instance lifetime, so scopes are very easy to add for clarity but are not strictly necessary if you're comfortable with "unscoped" objects. If you have an object graph of true singleton objects, you can also store that component as a conventional (static final field) singleton and generate your ClientSDKComponent as a subcomponent of that longer-lived component. If it's important to your build dependency graph, you can also phrase it the other way, and have your ClientSDKComponent as a standalone component that depends on another #Component.
Related
I'm having problems finding proper solution for my problem, namely:
Let's consider workflow:
Application starts
Main components are registered in Autofac
Application loads plugin assembly and registers modules within it
Container is being build
Plugin handling logic is run
Plugin can add its own controllers. To properly handle that I had to prepare interface which will provide me types of custom controllers:
interface ICustomControllerProvider
{
IEnumerable<Type> GetControllerTypes();
}
Based on the above my app knows how to integrate specified types as controllers.
All controllers are also defined as services, so Autofac deals with their creation, and so...
Problem:
I want to avoid specifying custom controller type twice
public PluginControllerProvider : ICustomControllerProvider
{
public IEnumerable<Type> GetControllerTypes()
{
// 1st type specification
// controller types are specified here, so they could be integrated with app
yield return typeof(ControllerX);
yield return typeof(ControllerY);
}
}
public class PluginModule : Module
{
protected override void Load(ContainerBuilder builder)
{
builder.RegisterType<PluginControllerProvider>().As<ICustomControllerProvider>();
// 2nd type specification
// controllers have to be register in module as well
builder.RegisterType<ControllerX>();
builder.RegisterType<ControllerY>();
}
}
Is there any way how ControllerX and ControllerY could be managed by Autofac, where I specified them only in PluginControllerProvider?
I tried achieving that by providing custom registration source and resolving ICustomControllerProvider, however I cannot resolve ICustomControllerProvider based on arguments provided by IEnumerable<IComponentRegistration> RegistrationsFor(Service service, Func<Service, IEnumerable<ServiceRegistration>> registrationAccessor) from IRegistrationSource
Autofac does not offer the ability to inject a list of types registered with Autofac. You'll have to get that yourself by working with the lifetime scope registry.
Before I get into how you might solve this, I should probably note:
Listing the types registered is not a normal thing from a DI perspective. It'd be like listing all the controllers in your MVC application - you don't normally need to do that, and if you did, you'd likely need to build a whole metadata structure on top of it like the ApiExplorer that was built to do that on top of ASP.NET Core. That structure wouldn't be supported or involved with the DI system because you're querying about the system, not injecting live instances into the system.
If you are relying on DI to resolve controllers, you probably don't need a whole separate controller provider. Once you know what type you need for the request, you'd just resolve it.
All that's to say, while I'll answer the question, the way the design here is posed may be something you'd want to look at. What you're doing, trying to involve DI with listing metadata about the app... seems somewhat backwards (you'd feed the DI container based on the list of types, not get the list of types from the DI container).
But let's just go with it. Given:
There are controllers registered with Autofac
The controllers have no common base class or interface
There's no attributes on the controllers you could query
What I'd do is register all the controllers with some metadata. You need something to be able to locate the controllers in the list of all the types in the container.
builder.RegisterType<ControllerX>().WithMetadata("controller", true);
builder.RegisterType<ControllerY>().WithMetadata("controller", true);
Now in the plugin controller, you need to inject an ILifetimeScope because you have to query the list of stuff registered. The ILifetimeScope that gets injected into the controller will be the same scope from which the plugin controller itself was resolved.
You can use the injected scope to query things in the component registry tagged with your metadata.
public class PluginControllerProvider : ICustomControllerProvider
{
private readonly Type[] _controllerTypes;
public PluginController(ILifetimeScope scope)
{
_controllerTypes = scope
.ComponentRegistry
.Registrations
.Where(r => r.Metadata.ContainsKey("controller"))
.Select(r => r.Activator.LimitType)
.ToArray();
}
public IEnumerable<Type> GetControllerTypes()
{
return _controllerTypes;
}
}
Now, disclaimers:
If you are registering more controllers in child lifetime scopes (e.g., during a BeginLifetimeScope() call), you will need a controller provider from that scope or it won't get all the controller types. The provider needs to come from the scope that has all the registrations.
If you're using registration sources (like the AnyConcreteTypeNotAlreadyRegisteredSource), this won't capture things that come from the registration sources. It'll only capture things that come from direct registrations of a type (or lambda) on a ContainerBuilder.
But it should work.
When you want to provide scoped dependency, Dagger requires to annotate Component with this Scope too. What is the reason for that?
#Singleton // <- for what?
#Component(modules = [MyModule::class])
interface MyComponent {
//...
}
#Module
interface MyModule {
#Binds
#Singleton
fun provideDependency(impl: DepImpl): Dep
}
Scoping in Dagger means "give this binding the same lifetime as the component that contains it", so the component itself holds onto the instance so it can provide the same instance repeatedly.
Components are also hierarchical: through subcomponents and component dependencies you can have multiple components coexisting in the same application. In a typical Android app, you might have an Application component that provides bindings for several Activity components, and and each of those Activity components provides bindings for several Fragment components.
Though Dagger could theoretically infer scopes based on which module installs it, instead of requiring you to specify which scope applies to which component, this makes it more difficult to work with classes that declare scope and #Inject annotations without an explicit Dagger binding. If you're in an Activity component that inherits bindings from an Application component (often "singleton"), and an Activity binding relies on an #ApplicationScope-annotated class with an #Inject constructor, how would Dagger know to put it in the Application component instead of the Activity component? This might also make your code harder for humans to understand—I know I'd have trouble following it.
I am new to Dagger (version 2.16 on Android) and based on my reading so far, I understand that for a component, there should be a provider (#Provides or #Binds) encapsulated in a module (#Module). Going through a lot of samples, I see code which has some objects which are not offered in any Module, nor are they being instantiated using new.
It is also my understanding that in order to access the dependencies in a module, a consumer class needs to inject itself in the component graph (components usually offer a method to inject classes). The code examples are not doing this either.
Here's some code demonstrating both my concerns. RecipePresenter is not being provided in any module, but still RecipeActivity is using it.
A possible explanation I could think of is that #Inject, in addition to requesting the dependency also adds/injects the requesting class (RecipePresenter in the linked code) into the component graph. But assuming there are multiple components/subcomponents, which component does the class using #Inject constructor gets attached to? If my understanding is correct, why do activities and fragments have to inject themselves manually in the component - shouldn't they be auto-injected if they declare a variable annotated with #Inject?
RecipePresenter has an #Inject-annotated constructor, which allows it to be provided. The #Inject annotation on the recipePresenter field within RecipeActivity does not help, just the #Inject-annotated constructor.
class RecipePresenter #Inject constructor(
private val useCase: RecipeUseCase,
private val res: StringRetriever) :
RecipeUseCase.Callback {
From the Dagger User's Guide:
Use #Inject to annotate the constructor that Dagger should use to create instances of a class. When a new instance is requested, Dagger will obtain the required parameters values and invoke this constructor.
If the class with the #Inject-annotated constructor also has a defined scope, then the binding will only affect components with the same scope annotation: An #ActivityScope class wouldn't be accessible from a #Singleton component, for instance. If the class has no scope defined, then the binding could appear on any and all components where it is needed, which is fine: The implementation is always the same, defined by the constructor itself.
#Inject has different meanings based on what it's annotating:
When you annotate a field with #Inject, it indicates that the DI system should set that field based on values from the DI system when it injects that object.
When you annotate a method with #Inject, it indicates that the DI system should call that method with parameters based on values from the DI system when it injects that object.
When you annotate a constructor with #Inject, it indicates that the DI system is allowed to call that constructor in order to create that object (which triggers the field and method injection above). In that sense it does act like a built-in Provider.
In Dagger specifically, #Provides methods take precedence over #Inject constructors (if they both exist for the same class), and unlike in the JSR-330 spec Dagger requires an #Inject-annotated constructor even when no other constructors are present. From the Dagger User's Guide:
If your class has #Inject-annotated fields but no #Inject-annotated constructor, Dagger will inject those fields if requested, but will not create new instances. Add a no-argument constructor with the #Inject annotation to indicate that Dagger may create instances as well.
Classes that lack #Inject annotations cannot be constructed by Dagger.
Finally, to answer your question about why activities and fragments need to inject themselves: Android is allowed to reflectively create Activity/Fragment/View objects without Dagger's help or participation. This means that nothing triggers the field and method injection described above, until you call a members-injection method on the component that instructs Dagger to populate those fields and call those methods. Consequently, you should never see an #Inject-annotated constructor on Application, Activity, Service, Fragment, View, ContentProvider, and BroadcastReceiver subclasses in Android: Android will ignore the #Inject annotation, so you might as well take control over injection yourself, manually or through dagger.android.
I never worked with such a confusing DI-framework like dagger! - However, I try to wrap my head around it.
I have two scopes: ActivityScope and FragmentScope
On some of the samples provided StatisticsFragment.java you see e.g. the fragment annotated with the scope
#ActivityScoped
public class StatisticsFragment extends DaggerFragment implements
StatisticsContract.View {
...
}
Question 1:
Is this just documentation or not? In my app it makes no difference if I annotate the concrete fragment or not.
Question 2: Where in the generated code can I see which scope is used? My fragment injects a Presenter and an AuthProvider. The AuthProvider is annotated with Singleton (in AppModule), the Presenter is defined in UIModule -> LoginModule
looks like this:
UIModule.java:
#Module(includes = AndroidSupportInjectionModule.class)
public abstract class UIModule {
#ActivityScope
#ContributesAndroidInjector(modules = LoginModule.class)
abstract LoginActivity loginActivity();
#ChildFragmentScope
#ContributesAndroidInjector(modules = LoginModule.class)
abstract LoginFragment loginFragment();
#Binds
//#ChildFragmentScope
public abstract LoginContract.View loginView(final LoginFragment fragment);
}
LoginModule.java
#Module
public abstract class LoginModule {
#Provides
//#ChildFragmentScope
static LoginContract.Presenter provideLoginPresenter(final LoginContract.View view, final BaseStore store) {
return new LoginPresenter(view,store);
}
}
LoginFragemt.java
public class LoginFragment extends DaggerFragment {
#Inject
LoginContract.Presenter presenter;
#Inject
Provider<MyAuthClass> myAuthClass;
...
}
presenter is created every time the Fragment gets created, myAuthClass gets created only once and is singleton.
Perfect - but I have no idea HOW this works!!!
DaggerFragment#onAttach must somehow know that Presenter is a "local" singleton and MyAuthClass is a global-singleton ...
Scope is one of two ways you can tell Dagger to always bind the same object, rather than returning a newly-created one on each injection request. (The other way is the manual way: Just return the same object in a #Provides method.)
First, a scope overview: Let's say you have a component, FooComponent, which has a #FooScope annotation. You define a subcomponent, BarComponent, which has a #BarScope annotation. That means that using a single FooComponent instance, you can create as many BarComponent instances as you want.
#FooScoped
#Component(modules = /*...*/)
public interface FooComponent {
BarComponent createBarComponent(/* ... */); // Subcomponent factory method
YourObject1 getYourObject1(); // no scope
YourObject2 getYourObject2(); // FooScoped
}
#BarScoped
#Subcomponent(modules = /*...*/)
public interface BarComponent {
YourObject3 getYourObject3(); // no scope
YourObject4 getYourObject4(); // BarScoped
YourObject5 getYourObject5(); // FooScoped
}
When you call fooComponent.getYourObject1(), YourObject1 is unscoped, so Dagger does its default: create a brand new one. When you call fooComponent.getYourObject2(), though, if you've configured that YourObject2 to be #FooScoped, Dagger will return exactly one instance for the entire lifetime of that FooComponent. Of course, you could create two FooComponent instances, but you'll never see multiple instances of a #FooScoped object from the same #FooScoped component (FooComponent).
Now onto BarComponent: getYourObject3() is unscoped, so it returns a new instance every time; getYourObject4() is #BarScoped, so it returns a new instance for each instance of BarComponent; and getYourObject5() is #FooScoped, so you'll get the same instance along the instance of FooComponent from which the BarComponent was created.
Now to your questions:
Question 1: Is this just documentation or not? In my app it makes no difference if I annotate the concrete fragment or not.
In classes that have an #Inject-annotated constructor like StatisticsFragment does, adding a scope annotation is not simply documentation: Without the scope annotation, any requests to inject a StatisticsFragment will generate a brand new one. If you only expect there to be a single instance of StatisticsFragment per Activity, this may be surprising behavior, but it might be hard to notice the difference.
However, adding an #Inject annotation to a Fragment may be something of a controversial move, because the Android infrastructure is able to create and destroy Fragment instances itself. The object that Android recreates will not the scoped one, and it will have its members reinjected onAttach due to DaggerFragment's superclass behavior. I think a better practice is to drop the #Inject annotation from the constructor and stick with field injection for your Fragment. At that point you can drop the scope, because Dagger will never create your Fragment, so it'll never decide whether to create a new one or return an existing one.
Question 2: Where in the generated code can I see which scope is used? My fragment injects a Presenter and an AuthProvider. The AuthProvider is annotated with Singleton (in AppModule), the Presenter is defined in UIModule -> LoginModule
The generated code and scoping is always generated in the Component; subcomponents will have their implementations generated as an inner class of the Component. For each scoped binding, there will be a place in the initialize method where the Provider (e.g. AuthProvider) is wrapped in an instance of DoubleCheck that manages the double-checked locking for singleton components. If nobody asks Dagger to create an object (like StatisticsFragment), Dagger can determine the lack of component factory methods or injections in the graph, and can avoid adding any code generation for it at all—which might be why you're not seeing any.
In the strange, strange dependency injection world of Jersey, you can include an AbstractBinder (but maybe not just a Binder) as one of the objects in the return value of your Application's getSingletons() method.
That AbstractBinder can then call various bind() methods from within its configure() method, which Jersey, but no other JAX-RS implementation, is guaranteed to call—and hence you can link implementations to interfaces, which lets you do a semblance of dependency injection from that point forward in your application. That is, once you've done that, then injection points within, say, your resource classes will be "filled" by the implementations you've bound.
(You can also do this with factory classes, so that you bind a particular factory method's return value to a contract it implements.)
OK, fine.
Jersey also lets you place a Class that implements Feature into the return value of your Application's getClasses() method. Fine. This Feature class will be instantiated by HK2 under the covers—make a mental note of that!—and its configure(FeatureContext) method will be called at some point. At that point, the Feature may register some additional stuff by calling FeatureContext#register() and handing it whatever it wants to register.
(Presumably this is all a fairly complicated façade on top of HK2's DynamicConfiguration machinery.)
Anyway, since a Feature is instantiated by HK2 (remember?), it follows that you can #Inject things into its constructor, or have injection points elsewhere in the class. That's cool! Then HK2 can draw upon all the services it knows of to fill those injection points. Very neat.
Ah, but the question is: what is the state of the HK2 world at this point? What services can be injected into a Feature implementation instantiated as part of A JAX-RS Application's startup sequence?
If your Application's getSingletons() method returns an AbstractBinder implementation that binds a FooImpl to Foo in Singleton scope in its configure() method, can your Feature—"registered" by including its class in the return value of your Application's getClasses() method—then inject a Foo?
I think it is important to have your Foo interface binding proxied, i.e:
new AbstractBinder() {
#Override
protected void configure() {
bind(Foo.class)
.proxy(true)
.to(FooImpl.class)
.in(Singleton.class);
}
}
then dependency injection will be insensitive to the instatination order.