Asp.Net Core: creating a new scoped IServiceProvider - asp.net-core

Objective
Create a asp.net core based solution that permits plugins loaded in runtime, way after IServiceCollection/IServiceProvider have been locked down to change.
Issue
IServiceCollection is configured at startup, from which IServiceProvider is developed, then both are locked for change before run is started.
I'm sure there are great reasons to do this....but I rue the day they came up with it being the only way to do things... so:
Attempt #1
Was based on using Autofac's ability to make child containers, falling back to parent containers for whatever is not specific to the child container,
where, right after uploading the new plugin, I create a new ILifetimeScope so that I can add Services given its containerBuilder:
moduleLifetimeScope = _lifetimeScope.BeginLifetimeScope(autoFacContainerBuilder =>
{
//can add services now
autoFacContainerBuilder.AddSingleton(serviceType, tInterface);
}
save the scope and its Container in a dictionary, against controllerTypes found in the dll, so that:
later can use a custom implementation of IControllerActivator to first try with the default IServiceProvider before falling back to try in the child plugin's child container.
The upside was, Holy cow, with a bit of hacking around, slowly got Controllers to work, then DI into Controllers, then OData....
The downside was that its custom to a specific DI library, and the Startup extensions (AddDbContext, AddOData) were not available as autoFacContainerBuilder doesn't implement IServiceCollection, so it became a huge foray into innards...that sooner or later couldn't keep on being pushed uphill (eg: couldn't figure out how to port AddDbContext)
Attempts #2
At startup, save a singleton copy of the original ISourceCollectionin theISourceCollection` (to easily re-get it later)
Later, upon loading a new plugin,
Clone the original ISourceCollection
Add to the clonedServiceCollection new Plugin Services/Controllers found in by Reflection
Use standard extension methods to AddDbContext and AddOData, etc.
Use a custom implementation of IControllerActivator as per above, falling back to the child IServiceProvider
Holy cow. Controllers work, OData works, DbContext works...
Hum...it's not working perfectly. Whereas the Controllers and being created new on every request, it's the same DbContext every time, because it's not being disposed, because it's not scoped by some form of scopefactory.
Attempt #3
Same thing as #2, but instead of making the IServiceProvider when the module is loaded, now -- in the custom IControllerActivator making a new IServiceProvider on each request.
No idea how much memory/time this is wasting, but I'm guessing its ...not brilliant
But sure...but I've really just pushed the problem a bit further along, not gotten rid of it:
A new IServiceProvider is being created...but nothing is actually disposing of it either.
backed by the fact that I'm watching memory usage increase slowly but surely....
Attempt #4
Same as above, but instead of creating a new IServiceProvider on every request, I'm keeping the IServiceProvider that i first built when I uploaded the module, but
using it to built a new Scope, and get its nested IServiceProvider,
hold on to the scope for later disposal.
It's a hack as follows:
public class AppServiceBasedControllerActivator : IControllerActivator {
public object Create(ControllerContext actionContext)
{
...
find the cached (ControllerType->module Service Provider)
...
var scope = scopeDictionaryEntry.ServiceProvider.CreateScope();
httpController = serviceProvider.GetService(controllerType);
actionContext.HttpContext.Items["SAVEMEFROMME"] = scope;
return httpController;
}
public virtual void Release(ControllerContext context, object controller)
{
var scope = context.HttpContext.Items["SAVEMEFROMME"] as IServiceScope;
if (scope == null){return;}
context.HttpContext.Items.Remove("SAVEMEFROMME");
scope.Dispose(); //Memory should go back down..but doesn't.
}
}
}
Attempt #5
No idea. Hence this Question.
I feel like I'm a little further along...but just not closing the chasm to success.
What would you suggest to permit this, in a memory safe way?
Background Musings/Questions in case it helps?
As I understand it, the default IServiceProvider doesn't have a notion of child lifespan/containers, like Autofac can create.
I see a IServiceScopeFactory makes a new IServiceProvider.
I understand there is some middleware (what name?) that invokes IServiceScopeFactory to make a IServiceProvider on every single request (correct?)
are these per-request IServiceProviders really separate/duplicate, and don't 'descend' from a parent one and falls back to parent if a asked for a singleton?
What is the Middleware doing different to dispose/reduce memory at the end of the call?
Should I be thinking about replacing the middleware? But even if it could -- it's so early that I only would have an url, not yet a Controller Type, therefore don't know what Plugin Assembly the Controller came from, therefore don't know what IServiceProvider to use for it...therefore too early to be of use?
Thank you
Getting a real grip on adding plugin sourced scoped services/controllers/DbContexts would be...wow. Been looking for this capability for several months now.
Thanks.
Other Posts
some similarity to:
Use custom IServiceProvider implementation in asp.net core
but I don't see how his disposing is any different to what I'm doing, so are they too having memory issues?

Related

User scoped dependencies in a custom ASP.NET Core Action Filter?

According to the official documentation here:
https://learn.microsoft.com/en-us/aspnet/core/mvc/controllers/filters#authorization-filters
To implement a custom ActionFilter in ASP.NET Core I have three choices:
SeviceFilterAttribute
TypeFilterAttribute
IFilterFactory
But for all three it is stated that:
Shouldn't be used with a filter that depends on services with a lifetime other than singleton.
So how can I inject scoped services in my custom ActionFilter? I can easily get a scoped service from the current HttpContext like this:
public override void OnActionExecuting(ActionExecutingContext actionContext)
{
ISubscriptionHelper subscriptionHelper =
actionContext.HttpContext.RequestServices
.GetRequiredService<ISubscriptionHelper>();
}
But then I am wondering if I am doing something wrong? What is the correct way to depend on scoped services in a custom ActionFilterAttribute?
Resolving services from the HttpContext.RequestServices will correctly resolve scoped and transient instances without causing any problems such as Captive Dependencies. In case resolved components implement IDisposable, they will be disposed of when the request ends. ASP.NET Core passes on the current HttpContext object to filter's OnActionExecuting method and that HttpContext gives access to the DI Container.
This is completely different from injecting those services into the constructor, because the action filter will be cached for the lifetime of the application. Any dependencies stored in private fields will, therefore, live as long as that filter. This leads to the so called Captive Dependency problem.
Code that accesses the DI Container (the HttpContext.RequestServices is your gateway into the DI Container) should be centralized in the infrastructure code of the startup path of the application—the so called Composition Root. Accessing your DI Container outside the Composition Root inevitably leads to the Service Locator anti-pattern—this should not be taken lightly.
To prevent this, it is advised to keep the amount of code inside the action filter as small as possible and implement the filter as a Humble Object. This means that preferably, the only line of code inside the filter is the following:
actionContext.HttpContext.RequestServices
.GetRequiredService<ISomeService>() // resolve service
.DoSomeOperation(); // delegate work to service
This means all (application) logic is moved to the ISomeService implementation, allowing the action filter to become a Humble Object.

Can I create an object from the DI container/Lamar in .NET 6.0 minimal hosting, preserving singletons?

We have migrated from a windows Framework 4.7 application to .NET 6.0. Lamar is added for Dependency Injection. We are trying to finalize a refactor to the latest "one-file" program.cs but are getting unexpected System.ObjectDisposedException: 'Cannot access a disposed object'. In all cases, the error is against a Func<T> during object creation.
All our tests are running correctly using the same environment, except to start the tests we (a) create the DI container and (b) use the container to create an object that loads the singletons (from MongoDB):
Container = new Container(registry);
var start = Container.GetInstance<HomeService>();
In the program.cs, we configure the container, but do not get to see it created, or access it inside program.cs. Instead we create HomeService from IServiceProvider during the first use of a controller. Here we were trying to limit the lifecyle scope during creation:
using (var scope = _container.CreateScope())
{
scope.ServiceProvider.GetService<INewHomeService>();
}
For test, we use the same loading steps, except for adding controllers/mvc, of course (i.e. NOT using builder.Services.AddControllers(); and builder.Services.AddMvc() for (integration) testing).
We have tried a lot of different things, like creating our object independently to the startup, but that did not align the singletons. We can get functionality by using static instead, but then we lose dynamic change access.
Some great tips like Resolving instances with ASP.NET Core DI from within ConfigureServices and https://andrewlock.net/exploring-dotnet-6-part-10-new-dependency-injection-features-in-dotnet-6/ but I can't see the specific example to get the live container just after initial creation.
Is it possible that the issue is just the difference between the lifecycle management of the new .NET DI implementation? As this is configuration at the composition root, if we can configure as per our testing approach, it should solve our problem. Other solutions welcome!
The problem 'Cannot access a disposed object' was being caused by a lifecycle mismatch between retained context and the controller access. The code retained a handle on the state object, that had a handle on the factory using FUNC. As we did not configure the Func as anything, it was transient during the controller graph creation, and so was disposed when the controller request ended.
To solve, we tried registering ALL of the FUNC, per How to use Func<T> in built-in dependency injection which was a large task as we had a few factories throughout an old codebase.
The better solution was to create a factory in the composition root, and use an injected IserviceProvider (or with Lamar an IContainer). This is a simple workaround.
With our creation concern, the creation of our object after the completion of the startup process is working correctly as a lazy validation of the first controller access.

Autofac Multitenant Database Configuration

I have a base abstract context which has a couple hundred shared objects, and then 2 "implementation" contexts which both inherit from the base and are designed to be used by different tenants in a .net core application. A tenant object is injected into the constructor for OnConfiguring to pick up which connection string to use.
public abstract class BaseContext : DbContext
{
protected readonly AppTenant Tenant;
protected BaseContext (AppTenant tenant)
{
Tenant = tenant;
}
}
public TenantOneContext : BaseContext
{
public TenantOneContext(AppTenant tenant)
: base(tenant)
{
}
}
In startup.cs, I register the DbContexts like this:
services.AddDbContext<TenantOneContext>();
services.AddDbContext<TenantTwoContext>();
Then using the autofac container and th Multitenant package, I register tenant specific contexts like this:
IContainer container = builder.Build();
MultitenantContainer mtc = new MultitenantContainer(container.Resolve<ITenantIdentificationStrategy>(), container);
mtc.ConfigureTenant("1", config =>
{
config.RegisterType<TenantOneContext>().AsSelf().As<BaseContext>();
});
mtc.ConfigureTenant("2", config =>
{
config.RegisterType<TenantTwoContext>().AsSelf().As<BaseContext>();
});
Startup.ApplicationContainer = mtc;
return new AutofacServiceProvider(mtc);
My service layers are designed around the BaseContext being injected for reuse where possible, and then services which require specific functionality use the TenantContexts.
public BusinessService
{
private readonly BaseContext _baseContext;
public BusinessService(BaseContext context)
{
_baseContext = context;
}
}
In the above service at runtime, I get an exception "No constructors on type 'BaseContext' can be found with the constructor finder 'Autofac.Core.Activators.Reflection.DefaultConstructorFinder'". I'm not sure why this is broken....the AppTenant is definitely created as I can inject it other places successfully. I can make it work if I add an extra registration:
builder.RegisterType<TenantOneContext>().AsSelf().As<BaseContext>();
I don't understand why the above registration is required for the tenant container registrations to work. This seems broken to me; in structuremap (Saaskit) I was able to do this without adding an extra registration, and I assumed using the built in AddDbContext registrations would take care of creating a default registration for the containers to overwrite. Am I missing something here or is this possibly a bug in the multitenat functionality of autofac?
UPDATE:
Here is fully runable repo of the question: https://github.com/danjohnso/testapp
Why is line 66 of Startup.cs needed if I have lines 53/54 and lines 82-90?
As I expected your problem has nothing to do with multitenancy as such. You've implemented it almost entirely correctly, and you're right, you do not need that additional registration, and, btw, these two (below) too because you register them in tenant's scopes a bit later:
services.AddDbContext<TenantOneContext>();
services.AddDbContext<TenantTwoContext>();
So, you've made only one very small but very important mistake in TenantIdentitifcationStrategy implementation. Let's walk through how you create container - this is mainly for other people who may run into this problem as well. I'll mention only relevant parts.
First, TenantIdentitifcationStrategy gets registered in a container along with other stuff. Since there's no explicit specification of lifetime scope it is registered as InstancePerDependency() by default - but that does not really matter as you'll see. Next, "standard" IContainer gets created by autofac's buider.Build(). Next step in this process is to create MultitenantContainer, which takes an instance of ITenantIdentitifcationStrategy. This means that MultitenantContainer and its captive dependency - ITenantIdentitifcationStrategy - will be singletons regardless of how ITenantIdentitifcationStrategy is registered in container. In your case it gets resolved from that standard "root" container in order to manage its dependencies - well, this is what autofac is for anyways. Everything is fine with this approach in general, but this is where your problem actually begins. When autofac resolves this instance it does exactly what it is expected to do - injects all the dependencies into TenantIdentitifcationStrategy's constructor including IHttpContextAccessor. So, right there in the constructor you grab an instance of IHttpContext from that context accessor and store it for using in tenant resolution process - and this is a fatal mistake: there's no http request at this time, and since TenantIdentitifcationStrategy is a singleton it means that there will not ever be one for it! So, it gets null request context for the whole application lifespan. This effectively means that TenantIdentitifcationStrategy will not be able to resolve tenant identifier based on http requests - because it does not actually analyze them. Consequently, MultitenantContainer will not be able to resolve any tenant-specific services.
Now when the problem is clear, its solution is obvious and trivial - just move fetching of request context context = _httpContextAccessor.HttpContext to TryIdentifyTenant() method. It gets called in the proper context and will be able to access request context and analyze it.
PS. This digging has been highly educational for me since I had absolutely no idea about autofac's multi-tenant concept, so thank you very much for such an interesting question! :)
PPS. And one more thing: this question is just a perfect example of how important well prepared example is. You provided very good example. Without it no one would be able to figure out what the problem is since the most important part of it was not presented in the question - and sometimes you just don't know where this part actually is...

RoleProvider injection with Ninject

First off I want to say there is a ton of answers on SO and google searches surrounding this, however I'm running into an issue that prevents those solutions from working. The answer here seems to be the way to go. (kernel.Inject(Roles.Provider);)
The issue that I'm having is that when I'm trying to inject the RoleProvider Roles.Provider is null, however my custom provider is found in the list within Roles.Providers. I am thinking that Ninject is trying to access the role provider too soon.
In my NinjectWebCommon class it appears that it's using WebActivator to start itself. [assembly: WebActivator.PreApplicationStartMethod(typeof(Admin.App_Start.NinjectWebCommon), "Start")]
It appears that all of the articles I've come across are using older versions of Ninject and are doing a lot of the heavy lifting in the Global.asax Application_Start... Using my implementation how can I get DI working for a custom role provider?
I'm using the [Inject] attribute within my custom provider.
The WebActivator pipeline runs way before even the standard ASP.NET stack is created. It means that you won't have access to anything created by ASP.NET during bootstrap in NinjectWebCommon.
Use that file only to declare bindings that do not depend on ASP.NET stuff to be up.
In order to get around this issue, you should use Global.asax Application_Start handler to load any additional modules/bindings that are dependend on ASP.NET stuff such as Roles.Provider.
Here is a suggestion that may solve your problem:
public void Application_Start()
{
var kernel = (new Bootstrapper()).Kernel;
kernel.Inject(Roles.Provider);
//Other initialization stuff
}
The Bootstrapper class is a lazy singleton that has a static IKernel initialized within your NinjectWebCommon.cs. So this is the proper way of retrieving the configured kernel instance from outside your NinjectWebCommon.
Give it a try.

Managing NHibernate ISession with Autofac

Does anyone have any tips or best practices regarding how Autofac can help manage the NHibernate ISession Instance (in the case of an ASP.NET MVC application)?
I'm not overly familiar with how NHibernate sessions should be handled. That said, Autofac have excellent instance lifetime handling (scoping and deterministic disposal). Some related resources are this article and this question. Since you're in ASP.Net MVC land make sure you also look into the MVC integration stuff.
To illustrate the point, here's a quick sample on how you can use Autofac factory delegates and the Owned generic to get full control over instance lifetime:
public class SomeController
{
private readonly Func<Owned<ISession>> _sessionFactory;
public SomeController(Func<Owned<ISession>> sessionFactory)
{
_sessionFactory = sessionFactory;
}
public void DoSomeWork()
{
using (var session = _sessionFactory())
{
var transaction = session.Value.BeginTransaction();
....
}
}
}
The container setup to get this to work is quite simple. Notice that we don't have to do anything to get the Func<> and Owned<> types, these are made available automatically by Autofac:
builder.Register(c => cfg.BuildSessionFactory())
.As<ISessionFactory>()
.SingleInstance();
builder.Register(c => c.Resolve<ISessionFactory>().OpenSession());
Update: my reasoning here is that, according to this NHibernate tutorial, the lifetime of the session instance should be that of the "unit of work". Thus we need some way of controlling both when the session instance is created and when the session is disposed.
With Autofac we get this control by requesting a Func<> instead of the type directly. Not using Func<> would require that the session instance be created upfront before the controller instance is created.
Next, the default in Autofac is that instances have the lifetime of their container. Since we know that we need the power to dispose this instance as soon as the unit of work is done, we request an Owned instance. Disposing the owned instance will in this case immediately dispose the underlying session.
Edit: Sounds like Autofac and probably other containers can scope the lifetime correctly. If that's the case, go for it.
It isn't a good idea to use your IoC container to manage sessions directly. The lifetime of your session should correspond to your unit of work (transaction boundary). In the case of a web application, that should almost certainly be the lifetime of a web request.
The most common way to achieve this is with an HttpModule that both creates your session and starts your transaction when a request begins, then commits when the request has finished. I would have the HttpModule register the session in the HttpContext.Items collection.
In your IoC container, you could register something like HttpContextSessionLocator against ISessionLocator.
I should mention that your generic error handling should locate the current session and roll back the transaction automatically, or you could end up committing half a unit of work.