Sunday, 18 November 2012

Tracking Prism EventAggregator traffic

Tracking Prism Events

It is possible to track the Prism events that are sent and received through the Prism EventAggregator,
but this is not out of the box of the Prism Framework. Usually, the developer must insert breakpoints
where the CompositePresentationEvent is published and subscribed

This is satisfactory for small scenarios, but as your solution and project(s) grow, it can become an increasingly difficult task. It is in many cases desired to see which events are published and where they are subscribed. The Prism 4.x Framework still do not support this, but luckily there is a way to achieve this. 

First off, it is required to change your Prism Events from using the class CompositePresentationEvent, to instead use a derived class presented next, CustomCompositePresentationEvent. I have chosen to display the Prism Event traffic in the Output of Visual Studio. This is done using Debug.WriteLine. It could be perhaps an alternative to log this event traffic or display it in a separate Window if you are writing a WPF application. 

The custom class CustomCompositePresentationEvent looks like this: 
using System;
using System.ComponentModel;
using System.Diagnostics;
using System.Text;
using Microsoft.Practices.Prism.Events;

namespace EventAggregation.Infrastructure
{
    
    /// <summary>
    /// Custom CompositePresentationEvent class, derived from this base class to allow the Prism Events to be displayed in the Debug Output Window. 
    /// </summary>
    public class CustomCompositePresentationEvent<TEvent> : 
    CompositePresentationEvent<TEvent> where TEvent : class 
    {

        /// <summary>
        /// Overload of the publish method of CompositePresentationEvent 
        /// </summary>
        public override void Publish(TEvent payload)
        {
            var declaringType = string.Empty; 
            var callingMethod = new StackTrace().GetFrame(1).GetMethod(); 
            if (callingMethod != null)
                declaringType = callingMethod.DeclaringType.FullName; 
            System.Diagnostics.Debug.WriteLine(string.Format(@"Publishing event {0}
            at time: {1}. Source type: {2}. Event Payload: {3}", 
            typeof(TEvent).Name,DateTime.Now, declaringType, 
            GetContentAsString(payload))); 
            base.Publish(payload);
        }

        /// <summary>Overload of the Subscribe method of  CompositePresentationEvent</summary>
        /// <remarks>
        /// For this overload to be executed, set keepSubscriberReferenceAlive to 
        /// true in the subscribe setup method on the target. 
        /// </remarks>
        public override SubscriptionToken Subscribe(Action<TEvent> action,
        ThreadOption threadOption, bool keepSubscriberReferenceAlive, 
        Predicate<TEvent> filter)
        {

            var loggingAction = new Action<TEvent>(e =>
            {
                System.Diagnostics.Debug.WriteLine(string.Format(
                "Subscribing event {0} at time: {1}. Target type: {2}", 
                typeof(TEvent).Name,
                    DateTime.Now, action.Target.GetType().Name));

                action.Invoke(e);
            });

            var subscriptionToken = base.Subscribe(loggingAction, threadOption, 
            keepSubscriberReferenceAlive, filter);
            return subscriptionToken;
        }


        private string GetContentAsString(TEvent payload)
        {
            if (payload == null)
                return string.Empty;
            else
            {
                StringBuilder sb = new StringBuilder();
                foreach (PropertyDescriptor property in 
                 TypeDescriptor.GetProperties(payload))
                {                    
                    sb.Append(string.Format("{0}={1} ", property.Name, 
                    property.GetValue(payload))); 
                }
                return sb.ToString();                 
            }
        }

    }

}

As you can see from the source code above, the methods Subscribe and Publish are overriden. These are virtual methods of the CompositePresentationEvent base class. I have tested using this derived class from CompositePresentationEvent in the EventAggregator sample, which can be downloaded in the Prism 4.1 library with source code at the following url: Download Prism 4.1 You will find the EventAggregator sample in the \QuickStarts\EventAggregation folder after you have extracted the Prism 4.1 library with source code. I have done some minor changes to this demo solution. To the project EventAggregation.Infrastructure.Desktop, I added the class I made which was just presented, CustomCompositePresentationEvent. Next up, our Prism Events must use this CustomCompositePresentation event class! This is easy to change, as shown in the Prism Event FundAddedEvent, which I changed to the following, still in the project EventAggregation.Infrastructure.Desktop:

namespace EventAggregation.Infrastructure
{
    public class FundAddedEvent : CustomCompositePresentationEvent
    {
    }
}

As you can see, the FundAddedEvent Prism Event, is now inheriting from the class CustomCompositePresentationEvent. There are no changes needed to do for the way the FundAddedEvent is published, but the Subscribe method must be changed. The boolean flag keepSubscriberReferenceAlive must be set to true. The reason is that the Subscribe override method in the CustomCompositePresentationEvent requires this, as you can see in the source code. We modify the action sent back again to the base class to continue with its Subscribe implementation. I could not make the override method to work, i.e. the method override was not called, without setting the boolean flag keepSubscribeReferenceAlive to true. The change I had to make for the subscribe method of the FundAddedEvent is in the project ModuleB.Desktop, in the class ActivityPresenter. The line 67 was changed to this:

  subscriptionToken = fundAddedEvent.Subscribe(FundAddedEventHandler,  
  ThreadOption.UIThread, true, FundOrderFilter); 
The changes required then, is to change the Prism Events to inherit from the class CompositePresentationEvent to instead inherit from (use) the class CustomCompositePresentationEvent. This custom class is basically the CompositePresentationEvent class, but we have overridden the methods Subscribe and Publish. In addition, we must set the keepSubscriberReferenceAlive boolean flag to true in our Subscribe methods. This will potentially mean that you must change to the Subscribe overload method which has this flag. Let's first see the GUI of the EventAggregation sample:







The output from the Visual Studio is the most important feature of this CustomCompositePresentationEvent class. After all, we want to be able to see which classes publish and subscribe events, and the content of the payload of these events, correct? In this sample, the output looks like the following example:

Publishing event FundOrder at time: 18.11.2012 15:31:30. Source type: ModuleA.AddFundPresenter. Event Payload: CustomerId=Customer1 TickerSymbol=FundB 
Subscribing event FundOrder at time: 18.11.2012 15:31:30. Target type: ActivityPresenter

As we can see from the output above, the event FundOrder was published by the AddFundPresenter class. The PayLoad of this Prism event was CustomerId=Customer1 and the TickerSymbol=FundB. The event FundOrder was subscribed, interestingly at the same millisecond (i.e. the publish-subscribe execution obviously performs almost instantly in the Prism framework). If you think the concept of being able to see Prism traffic is both interesting and will help you as a developer, consider taking into to use this CompositePresentationEvent class. You will have to include the changes which I listed above also, but these changes should be relatively easy to overcome in a Prism application solution. The tracking or tracing of Prism Event Traffic will help you understand the application better and catch possible errors by this more convenient debugging scenario. For security reasons, you should consider if putting a Debug.WriteLine call to output the Event Payload is safe or not. Last, I want to mention the fact that Prism Publish-Subscribe scenarios take in my test 0 ms and occur to be happening almost instantly. Of course they take some CPU cycles to execute, but if you in any case wonder how fast is really a Publish-Subscribe, it will usually be the answer - REALLY FAST. However, it is possible that if multiple classes subscribe to a given Prism Event, it might take some milliseconds to execute. I have not tested this yet. In a MVVM Scenario, it is usually the ViewModels that publish and subscribe Prism Events, usually via an ICommand implementation (DelegateCommand) that is data bound in WPF or Silverlight to a Button or similar GUI object.

Monday, 5 November 2012

User-Defined conversion sample in C#

This article will present a sample of how to implement a User Defined Conversions in C#. The article will use a typical conversion example, where temperature measured in Celsius scale (Centigrades) are converted into the Fahrenheit scale (Fahrenheits). A conversion formula is available at the following urls (Wikipedia):
User Defined Conversions are implemented in C-sharp in a similar manner as operator overloading, actually it is an overload of conversion. These conversions can be either:
  • Implicit - An object is converted into another type without a specific cast, i.e implicit cast.
  • Explicit - An object is converted into another type with a specific cast, i.e.
    explicit cast.
A class can have multiple implicit and/or explicit user defined conversions. The user defined conversions must all be methods that are marked as public and static. In addition, they must be marked as implicit or explicit to specify which cast is allowed. If a conversion is marked as explicit, converting an object into the specified type without an explicit cast will in other words not call the user defined conversion. 

To specify a User-Defined Conversion in C# which is implicit, the following signature must be used for a implicit User-Defined Conversion:
public static implicit operator TReturnType (TInputType objectToConvert) {

 .. //Implementation elided 

A User-Defined Conversion in C# which is explicit, uses the following signature: 

public static explicit operator TReturnType (TInputType objectToConvert) {

 .. //Implementation elided 

The C-sharp programmer who has experience in operator overloading will see the similarities to operator overloading. Operator overloading follows much the same signature as shown here, but without the keywords explicit or implicit. In addition, operator overloading will specify the same keyword operator, but also specify the operator to overload. User-Defined Conversion do not specify an operator to overload, but implicitly or explicitly we are overloading casting (type conversion). 
Demonstration code 
An example of how to implement User-Defined Conversion follows. In this example, a base class called Temperature is used and there are two derived classes, Celsius and Fahrenheit. Both these derived classes defines an explicit user defined conversion which will take an instance of their own type and return an instance of the converted type. This means that the Fahrenheit class will contain an explicit User-Defined Conversion which will receive a Fahrenheit instance and return a Celsius instance. In return, the Celsius class will contain an explicit User-Defined Conversion which will receive a Celsius instance and convert this into a Fahrenheit instance. In the code, the mathematical conversion between the two temperature scales is performed inside the explicit user defined conversion. This will let the programmer control the way a cast is performed and allow the client or consumer of the class to convert the type into another type through a cast and logically fill the expectation that the converted object is correctly converted or casted. 
As with operator overloading, just because C-sharp allows User Defined Conversions, does not mean that the programmer should use this functionality in the C#-language wherever this strategy might fit. User-Defined Conversions are a programming capability many C#-programmers are not aware of, and the coder which adopt this strategy might confuse his fellow programmers or quickly understand that the concepts must be explained to the other coders. But then again, just as operator overloading allows very logical and powerful expressibility in code, user defined conversion will let the programmer control the way explicit and implicit type conversions, that is - casts - are performed. If you expect the class you are writing to be casted into different types or will provide consumers or clients of your class with the benefits of simple expressibility and determined behavior, you should also add explicit user defined conversions. Implicit User-Defined Conversions are sometimes hard to see which User-Defined Conversion will be called. Therefore, if a choice is available between either implicit or explicit User-Defined Conversion, I would suggest to go for explicit. This is the most specific and clear-intentioned casting overload of the two. 
The code for the base class Temperature is shown below:     

    /// Base class for temperatures 
    /// 
    public class Temperature
    {
        protected float _value; 

        public Temperature(float temperature)
        {
            _value = temperature; 
        }

        public float Value
        {
            get { return _value; } 
        }

    }

The derived class Celsius is shown next. Note that this class defines one explicit User-Defined conversion: 
    public class Celsius : Temperature 
    {

        public Celsius(float temperature) : base(temperature)
        {
        }

        public static explicit operator Fahrenheit(Celsius temperature)
        {
            float convertedTemperature = (temperature.Value * (9.0f / 5.0f)) + 32; 
            return new Fahrenheit(convertedTemperature); 
        }

    }

The derived class Fahrenheit is shown below. This class also defines one explicit User-Defined conversion. 
    public class Fahrenheit : Temperature
    {

        public Fahrenheit(float temperature) : base(temperature)
        {
            
        }

        public static explicit operator Celsius(Fahrenheit temperature)
        {
            float convertedTemperature = (5.0f / 9.0f)*(temperature.Value - 32);
            return new Celsius(convertedTemperature); 
        }

    }

The client or consumer class is shown next:
class Program
    {

        static void Main(string[] args)
        {
            Console.WriteLine("Centigrades to Fahrenheits conversion in range [-40,40] Celsius: ");

            Celsius celsius = null;
            Fahrenheit fahrenheit = null;

            for (int i = -40; i <= 40; i++)
            {
                celsius = new Celsius(i);
                fahrenheit = (Fahrenheit)celsius;
                Console.WriteLine("{0} Centigrades is measured into {1:f1} Fahrenheits.", i, fahrenheit.Value);
            }

            Console.WriteLine("Press any key to continue ...");
            Console.ReadKey();
        }

    }

The output in the console window when running the Main method in the consumer class 
above is shown next: 
Centigrades to Fahrenheits conversion in range [-40,40] Celsius:
-40 Centigrades is measured into -40,0 Fahrenheits.
-39 Centigrades is measured into -38,2 Fahrenheits.
-38 Centigrades is measured into -36,4 Fahrenheits.
-37 Centigrades is measured into -34,6 Fahrenheits.
-36 Centigrades is measured into -32,8 Fahrenheits.
-35 Centigrades is measured into -31,0 Fahrenheits.
-34 Centigrades is measured into -29,2 Fahrenheits.
-33 Centigrades is measured into -27,4 Fahrenheits.
-32 Centigrades is measured into -25,6 Fahrenheits.
-31 Centigrades is measured into -23,8 Fahrenheits.
-30 Centigrades is measured into -22,0 Fahrenheits.
-29 Centigrades is measured into -20,2 Fahrenheits.
-28 Centigrades is measured into -18,4 Fahrenheits.
-27 Centigrades is measured into -16,6 Fahrenheits.
-26 Centigrades is measured into -14,8 Fahrenheits.
-25 Centigrades is measured into -13,0 Fahrenheits.
-24 Centigrades is measured into -11,2 Fahrenheits.
-23 Centigrades is measured into -9,4 Fahrenheits.
-22 Centigrades is measured into -7,6 Fahrenheits.
-21 Centigrades is measured into -5,8 Fahrenheits.
-20 Centigrades is measured into -4,0 Fahrenheits.
-19 Centigrades is measured into -2,2 Fahrenheits.
-18 Centigrades is measured into -0,4 Fahrenheits.
-17 Centigrades is measured into 1,4 Fahrenheits.
-16 Centigrades is measured into 3,2 Fahrenheits.
-15 Centigrades is measured into 5,0 Fahrenheits.
-14 Centigrades is measured into 6,8 Fahrenheits.
-13 Centigrades is measured into 8,6 Fahrenheits.
-12 Centigrades is measured into 10,4 Fahrenheits.
-11 Centigrades is measured into 12,2 Fahrenheits.
-10 Centigrades is measured into 14,0 Fahrenheits.
-9 Centigrades is measured into 15,8 Fahrenheits.
-8 Centigrades is measured into 17,6 Fahrenheits.
-7 Centigrades is measured into 19,4 Fahrenheits.
-6 Centigrades is measured into 21,2 Fahrenheits.
-5 Centigrades is measured into 23,0 Fahrenheits.
-4 Centigrades is measured into 24,8 Fahrenheits.
-3 Centigrades is measured into 26,6 Fahrenheits.
-2 Centigrades is measured into 28,4 Fahrenheits.
-1 Centigrades is measured into 30,2 Fahrenheits.
0 Centigrades is measured into 32,0 Fahrenheits.
1 Centigrades is measured into 33,8 Fahrenheits.
2 Centigrades is measured into 35,6 Fahrenheits.
3 Centigrades is measured into 37,4 Fahrenheits.
4 Centigrades is measured into 39,2 Fahrenheits.
5 Centigrades is measured into 41,0 Fahrenheits.
6 Centigrades is measured into 42,8 Fahrenheits.
7 Centigrades is measured into 44,6 Fahrenheits.
8 Centigrades is measured into 46,4 Fahrenheits.
9 Centigrades is measured into 48,2 Fahrenheits.
10 Centigrades is measured into 50,0 Fahrenheits.
11 Centigrades is measured into 51,8 Fahrenheits.
12 Centigrades is measured into 53,6 Fahrenheits.
13 Centigrades is measured into 55,4 Fahrenheits.
14 Centigrades is measured into 57,2 Fahrenheits.
15 Centigrades is measured into 59,0 Fahrenheits.
16 Centigrades is measured into 60,8 Fahrenheits.
17 Centigrades is measured into 62,6 Fahrenheits.
18 Centigrades is measured into 64,4 Fahrenheits.
19 Centigrades is measured into 66,2 Fahrenheits.
20 Centigrades is measured into 68,0 Fahrenheits.
21 Centigrades is measured into 69,8 Fahrenheits.
22 Centigrades is measured into 71,6 Fahrenheits.
23 Centigrades is measured into 73,4 Fahrenheits.
24 Centigrades is measured into 75,2 Fahrenheits.
25 Centigrades is measured into 77,0 Fahrenheits.
26 Centigrades is measured into 78,8 Fahrenheits.
27 Centigrades is measured into 80,6 Fahrenheits.
28 Centigrades is measured into 82,4 Fahrenheits.
29 Centigrades is measured into 84,2 Fahrenheits.
30 Centigrades is measured into 86,0 Fahrenheits.
31 Centigrades is measured into 87,8 Fahrenheits.
32 Centigrades is measured into 89,6 Fahrenheits.
33 Centigrades is measured into 91,4 Fahrenheits.
34 Centigrades is measured into 93,2 Fahrenheits.
35 Centigrades is measured into 95,0 Fahrenheits.
36 Centigrades is measured into 96,8 Fahrenheits.
37 Centigrades is measured into 98,6 Fahrenheits.
38 Centigrades is measured into 100,4 Fahrenheits.
39 Centigrades is measured into 102,2 Fahrenheits.
40 Centigrades is measured into 104,0 Fahrenheits.

Press any key to continue ...

Most Europeans are familiar with the Centigrade system, but US-English citizens use Fahrenheit scale instead for describing the temperature. The freezing point of water, 0 degrees Celsius, is 32 Fahrenheit and about -17,78 Centigrades is 0 Fahrenheit. The boiling point of water, 100 degrees Centigrades, is 212 degrees Fahrenheit. As we can see, the temperature scales grow differently and are shifted somewhat. The User-Defined conversion allows to convert the two inheriting classes of the temperature base class through explicit casts. Some, but not all programmers, believe this is a natural way of performing a conversion. Often, one can see code which actually does the same kind of functionality but does not use User-Defined conversion, i.e. a public static method which receives an instance of type TInput and returns a converted object of type TResult, but without the keywords implicit or explicit and operator. This is of course allowed, but note that this will demand that the user specifically calls this method to perform the conversion routine to get the converted instance. Instead, specifying an implicit and/or expliciting User-Defined Conversion will in many cases be less ambigious and more precise. In the end, judge yourself if User-Defined Conversions are something to use in C#. But be aware that they exist and what they can be used for. More explanation on this topic available on the MSDN page:
User Defined Conversions
If no explicit cast is performed using the cast operator (parentheses), the implicit User Defined Conversion will be called, if both a matching explicit and implicit User Defined Conversion is defined. Note that if the assignment operator = is used without a cast, the implicit conversion is used and therefore the implicit User Defined Conversion will be utilized.

Sunday, 30 September 2012

MEF and memory

For those of you that are coding in .NET today, chances are high that you are doing it in using composition through usually a container and using Inversion of Control (IoC). Many will use the Managed Extensibility Framework or MEF, for achieving IoC. MEF is not entirely a true IoC container, as the framework supports non-specific import and export of interfaces bound to implementation(s) using inheritors of ComposablePartCatalog such as AssemblyCatalog.

To download versions of MEF, visit the site http://mef.codeplex.com However, most of us will use the stable version of MEF available in .NET 4, situated in the System.ComponentModel.Composition namespace.

MEF is very easy to get started with and the concepts of importing and exporting parts are really great.
However, there is big flaw in MEF which is well known to its users and contributors, which is the way MEF handles memory. Without getting too much into detail, MEF will generally keep a reference to all parts in its container as long as the application or system is running. This results in a high and observable memory leak. To take an example, I have put together a downloadable solution at the end of this article which shows a way to avoid the memory leak. The test solution will be explained throughout this article with the most relevant parts underlined.

First off, the base of this test solution is not mine. I needed to get a sample of a WPF/MVVM/MEF/Prism solution where I could test out the concepts at home and googled for some simple demo project. I found one at the top, Prism 4 MEF application. It is available at the following URL:

Prism 4 MEF application:
http://www.codeproject.com/Articles/254120/Prism-4-MEF-Application

Code Project Open License terms:
http://www.codeproject.com/info/cpol10.aspx

I will understate that my sample is based on the solution written by Danny Siew from Wellington, New Zealand. Thanks to Danny Siew I could focus on finding a solution of the memory leak problem in MEF.

To verify that the memory leak was fixed with my solution, I downloaded the ANTS Memory Profiler from Red Gate. A trial can be downloaded at the following URL:
http://www.red-gate.com/products/dotnet-development/ants-memory-profiler/

Using the ANTS Memory Profiler, one can quickly identity the memory leak through MEF by profiling your application that is using MEF and witness that parts that are removed and is no longer used in the application are kept alive (that is, not disposed so memory resources can become available again). My main concern are the parts that are marked as CreationPolicy.NonShared. The other parts are usually the singletons, with CreationPolicy.Shared. These parts will have at most one instance and should live throughout the lifetime of the application and not become disposed before the container is disposed. It is though a huge issue that those parts that are marked as NonShared will not be dispoed before the container is dispose, because this means that the number of instances will be accumulated in memory. Clearly, not a desired behavior of MEF!

My solution is actually based upon the following discussion at the following thread on the MEF Codeplex site:
http://mef.codeplex.com/discussions/285445

Thanks to the user Palomino for this great tip:
http://www.codeplex.com/site/users/view/palomino

The strategy of Palomino is to create a CustomCatalog inheriting from the usual abstract base class ComposablePartCatalog in MEF. This class will detect when a part is requiring disposing. I have modified Palomino's code slightly, as it will deactivate the non-desired caching of the part in the CatalogExportProvider MEF uses if the part is marked as CreationPolicy.NonShared. This is how I mean the correct behavior should be. The singletons or shared parts should be disposed in the usual "MEF-way", that is when the container is disposed. The non-shared parts should be disposed when parts go out of scope for early release of resources. In no circumstances should MEF keep caching non-shared parts, as the chances for memory leak is 100%. The project manager of MEF, Glenn Block has created ExportFactory classes that are able to dispose parts on demand. This will be available in .NET 4.5 and a port has been made available at Glenn's Skydrive for download (consisting of a single DLL file) where the inherited MEF Bootstrapper can be adjusted to allow this dispose on demand. I spent some time testing out this ExportFactory class but using ANTS Memory Profiler I could not see the memory leak go away..

It is time present some code for how to avoid this memory leak problem in ANTS, since any real application or system out there using MEF today will have this huge memory leak problem and cause degraded performance and larger and larger memory footprint as the application is running.

Lets start off with this new custom catalog, for any reason of better naming, I decided to change as little of Palomino's code, so let's just call it CustomCatalog here too:



  public class CustomCatalog : ComposablePartCatalog
    {
        private readonly ComposablePartCatalog _innerCatalog;

        public CustomCatalog(ComposablePartCatalog catalog)
        {
            _innerCatalog = catalog;
        }

        public override IEnumerable<Tuple<ComposablePartDefinition, ExportDefinition>> GetExports(ImportDefinition definition)
        {
            var exports = _innerCatalog.GetExports(definition);
            return exports.Select(e => new Tuple<ComposablePartDefinition, ExportDefinition>(
                new CustomComposablePartDefinition(e.Item1), e.Item2));
        }

        public override IQueryable<ComposablePartDefinition> Parts
        {
            get
            {
                return _innerCatalog.Parts
                    .Select(p => new CustomComposablePartDefinition(p))
                    .AsQueryable();
            }
        }
    }
To explain this class in short terms, we pass in an inheritor of the abstract ComposablePartCatalog, such as AssemblyCatalog, in the constructor. Then we override the method GetExports and readonly property Parts. Instead of just returning the implementation of the passed in ComposablePartCatalog inheritor instance, a wrapping CustomComposablePartDefinition is used. This class will be presented next:
    public class CustomComposablePartDefinition : ComposablePartDefinition
    {
        private readonly ComposablePartDefinition _innerPartDefinition;

        public CustomComposablePartDefinition(ComposablePartDefinition innerPartDefinition)
        {
            _innerPartDefinition = innerPartDefinition;
        }

        public override ComposablePart CreatePart()
        {
            var innerPart = _innerPartDefinition.CreatePart();
            if (ReflectionModelServices.IsDisposalRequired(_innerPartDefinition) && 
                _innerPartDefinition.Metadata.ContainsKey(CompositionConstants.PartCreationPolicyMetadataName) && 
                (CreationPolicy)_innerPartDefinition.Metadata[CompositionConstants.PartCreationPolicyMetadataName] == CreationPolicy.NonShared)
            {
                return new NotDisposableComposablePart(innerPart);
            }
            return innerPart;
        }

        public override IEnumerable<ExportDefinition> ExportDefinitions
        {
            get { return _innerPartDefinition.ExportDefinitions; }
        }

        public override IEnumerable<ImportDefinition> ImportDefinitions
        {
            get { return _innerPartDefinition.ImportDefinitions; }
        }

        public override IDictionary<string, object> Metadata
        {
            get { return _innerPartDefinition.Metadata; }
        }

        public override string ToString()
        {
            return _innerPartDefinition.ToString();
        }
    }

The class CustomComposablePartDefinition shown above inherits from the usual ComposablePartDefinition and provides a custom functionlity for the case where the part is a NonShared part and requires disposing, which usually means the part implements IDisposable. This is the source of the memory leak, i.e. where MEF will keep its cached reference to the part in the CatalogExportProvider and stop the Garbace Collector (GC) from being able to dispose the part on a timely manner. The code for the NotDisposableComposablePart follows:
public class NotDisposableComposablePart : ComposablePart
    {

        private ComposablePart _composablePart; 

        public NotDisposableComposablePart(ComposablePart composablePart)
        {
            _composablePart = composablePart; 
        }

        public override IEnumerable<ExportDefinition> ExportDefinitions
        {
            get { return _composablePart.ExportDefinitions; }
        }

        public override object GetExportedValue(ExportDefinition definition)
        {
            return _composablePart.GetExportedValue(definition); 
        }

        public override IEnumerable<ImportDefinition> ImportDefinitions
        {
            get { return _composablePart.ImportDefinitions; }
        }

        public override void SetImport(ImportDefinition definition, IEnumerable<Export> exports)
        {
            _composablePart.SetImport(definition, exports); 
        }
    }
The NotDisposableComposablePart inherits from the base class ComposablePart. It does not implement IDisposable and will not therefore be cached by the ExportCatalogProvider of MEF. Hence, the memory leak would not occur anymore where the container keeps reference. By using the ANTS Memory Profiler I can verify this behavior. An example how of to adjust an AssemblyCatalog in the MEF Bootstrapper follows, using this CustomCatalog:
           var assemblyCatalog = new AssemblyCatalog(typeof(ItemViewModel).Assembly);
            this.AggregateCatalog.Catalogs.Add(new CustomCatalog(assemblyCatalog)); 
This wrapping of the CustomCatalog can be done not only for AssemblyCatalog and TypeCatalog and so on. Wherever a memory leak is shown in large scale in an application using MEF, the wrapping where one uses CustomCatalog is strongly recommended. The following image shows an image from the sample application, which I have adjusted some from Danny's sample on Code Project:

Here is a screenshot of the memory profiling application I use, ANTS Memory Profiler from Red Gate:

To evaluate this memory leak fix of MEF yourself, the sample solution is available for download at the following URL: Prism 4 MEF Sample with memory leak fix [Zip file, 4 MB download]


It is strange that Microsoft choose to have an instance retention policy where the disposing of parts are postponed until the disposing of the container, which is usually done when the application is exited. This actually allows a huge memory leak, especially for the NonShared parts. It is usual in say a WPF/MVVM/Prism/MEF scenario to add and remove view models and other types of instances and maintaining these parts in memory allows a great access to memory leak in a multitude of applications. If you use MEF today, do not be suprised that there are a lot of memory leaks in your managed coded application. Just because it is .NET and managed code does not mean memory leaks are not present, on the contrary side. Thanks to the great tips of Palomino, user at MEF Codeplex, and Danny Siew for providing an easy and great Prism 4 MEF starting solution. Evaluate the code. Make note that the necessary Prism and MEF DLL-s are added. In addition to .NET 4 (obviously), download Microsoft WPF Ribbon 4.0 from MSDN here:


Microsoft Ribbon 4.0 WPF I have compiled the solution in Visual Studio 2012 RC so you will find binaries in the bin folders also to start off the sample right away. Hopefully, this fix will patch MEF until MEF 2 is released. After reading a bit more about MEF 2, it seems they use ExportFactory and the early release or dispose of objects will be an opt in. It will be easier to wrap the different catalogs used with this CustomCatalog to at least try to fix the massive memory leak that MEF almost provides as default behavior. Feel free to comment this article if you have supplementary information. I have used MEF quite a lot in the last two years, but I have less deep knowledge of MEF. I wrote this article because I had to inspect if there really were workarounds of this central bug of MEF.. Happy coding!