Krzysztof Koźmic

In the previous post we established that the usage of default constructor required by NHibernate is theoretically not required. The fact that NHibernate does use them though has to do with technical limitations of CLR. Turns out in most cases there is a workaround, which is not perfect but was a fun experiment to implement.

The code presented here is just an experiment. It is not of production quality – it is merely a proof of concept so if you use it in production and the world blows up, don’t blame me.

Interceptor

The place that NHibernate lets you plug into to instantiate entities and components as they are pulled from database is in NHibernate’s IInterceptor (not to be confused with Castle DynamicProxy IInerceptor) interface and its instantiate method. NHibernate has also a IObjectsFactory interface but that one is used to instantiate everything including NHibernate’s own infrastructure (like IUserTypes for example) and we want to override only activation of entities and components.

public class ExtensibleInterceptor : EmptyInterceptor
{
	private readonly INHibernateActivator activator;

	public ISessionFactory SessionFactory { get; set; }

	public ExtensibleInterceptor(INHibernateActivator activator)
	{
		this.activator = activator;
	}

	public override object Instantiate(string clazz, EntityMode entityMode, object id)
	{
		if (entityMode == EntityMode.Poco)
		{
			var type = Type.GetType(clazz);
			if (type != null && activator.CanInstantiate(type))
			{
				var instance = activator.Instantiate(type);
				SessionFactory.GetClassMetadata(clazz).SetIdentifier(instance, id, entityMode);
				return instance;
			}
		}
		return base.Instantiate(clazz, entityMode, id);
	}
}

The code here is straightforward. The INHibernateActivator is a custom interface that we’ll discuss in details below. That’s where the actual activation happens. We’re also using ISessionFactory so that we properly set the id of retrieved instance.

Proxy validator

Turns out that is not enough. In the interceptor we activate objects we fully obtain from the database. However big part of NHibernate performance optimisation is to ensure we don’t load more data that we need (a.k.a. lazy loading) and to do it transparently NHibernate uses proxies.

For every persistent type NHibernate uses object called proxy validator to ensure object meets all the requirements it has to be able to properly proxy a type and that work is done by IProxyValidator.

Part of the work of proxy validator is to check for default constructor so we need to override that logic to reassure NHibernate we have things under control.

public class CustomProxyValidator : DynProxyTypeValidator
{
	private const bool iDontCare = true;

	protected override bool HasVisibleDefaultConstructor(Type type)
	{
		return iDontCare;
	}
}

Proxy factory

Now that we told NHibernate we can handle proxying types without default constructor it’s time to write code that actually does it. That’s the task of ProxyFactory which uses Castle DynamicProxy under the cover

public class CustomProxyFactory : AbstractProxyFactory
{
	protected static readonly IInternalLogger log = LoggerProvider.LoggerFor(typeof (CustomProxyFactory));
	private static readonly DefaultProxyBuilder proxyBuilder = new DefaultProxyBuilder();
	private readonly INHibernateActivator activator;

	public CustomProxyFactory(INHibernateActivator activator)
	{
		this.activator = activator;
	}

	public override INHibernateProxy GetProxy(object id, ISessionImplementor session)
	{
		try
		{
			var proxyType = IsClassProxy
				                ? proxyBuilder.CreateClassProxyType(
				                	PersistentClass,
				                	Interfaces,
				                	ProxyGenerationOptions.Default)
				                : proxyBuilder.CreateInterfaceProxyTypeWithoutTarget(
				                	Interfaces[0],
				                	Interfaces,
				                	ProxyGenerationOptions.Default);

			var proxy = activator.Instantiate(proxyType);

			var initializer = new LazyInitializer(EntityName, PersistentClass, id, GetIdentifierMethod, SetIdentifierMethod,
				                                    ComponentIdType, session);
			SetInterceptors(proxy, initializer);
			initializer._constructed = true;
			return (INHibernateProxy) proxy;
		}
		catch (Exception e)
		{
			log.Error("Creating a proxy instance failed", e);
			throw new HibernateException("Creating a proxy instance failed", e);
		}
	}

	public override object GetFieldInterceptionProxy()
	{
		var proxyGenerationOptions = new ProxyGenerationOptions();
		var interceptor = new LazyFieldInterceptor();
		proxyGenerationOptions.AddMixinInstance(interceptor);
		var proxyType = proxyBuilder.CreateClassProxyType(PersistentClass, Interfaces, proxyGenerationOptions);
		var proxy = activator.Instantiate(proxyType);
		SetInterceptors(proxy, interceptor);
		SetMixin(proxy, interceptor);

		return proxy;
	}

	private void SetInterceptors(object proxy, params IInterceptor[] interceptors)
	{
		var field = proxy.GetType().GetField("__interceptors");
		field.SetValue(proxy, interceptors);
	}

	private void SetMixin(object proxy, LazyFieldInterceptor interceptor)
	{
		var fields = proxy.GetType().GetFields(BindingFlags.Public | BindingFlags.Instance);
		var mixin = fields.Where(f => f.Name.StartsWith("__mixin")).Single();
		mixin.SetValue(proxy, interceptor);
	}
}

It is pretty standard and not very interesting. We’re again using the INHibernateActivator, we’ve yet to discuss, to create instances. Also since we’re not calling any constructors to do it, we have to set the proxy dependencies that would otherwise by provided via constructor inline. For that we’re using some internal knowledge about how DynamicProxy lays out its proxy types.

Good news is – we’re almost there. Actually we might want to also override NHibernate reflection optimizer. As its name implies replaces some actions that are performed via reflection with faster, compiled versions. The default one works with assumption that default constructor is available and may fail otherwise but we can disable that feature for our simple proof of concept application.

Proxy factory factory

There’s one more step actually – we need to attach out proxy factory and proxy validator to session factory and we do that via ridiculously named IProxyFactoryFactory

public class CustomProxyFactoryFactory : IProxyFactoryFactory
{
	public IProxyFactory BuildProxyFactory()
	{
		return new CustomProxyFactory(new NHibernateActivator());
	}

	public bool IsInstrumented(Type entityClass)
	{
		return true;
	}

	public bool IsProxy(object entity)
	{
		return (entity is INHibernateProxy);
	}

	public IProxyValidator ProxyValidator
	{
		get { return new CustomProxyValidator(); }
	}
}

Putting it all together

Now that we have all these pieces let’s see how we put them together (with some help from FluentNHibernate)

var config = Fluently.Configure()
	.Mappings(c => c.FluentMappings.AddFromAssemblyOf<Program>())
	.Database(MsSqlConfiguration.MsSql2008
				.ConnectionString(s => s.Database("NHExperiment").Server(".").TrustedConnection())
				.ProxyFactoryFactory<CustomProxyFactoryFactory>())
	.BuildConfiguration();
var interceptor = new ExtensibleInterceptor(new NHibernateActivator());
config.Interceptor = interceptor;
NHibernate.Cfg.Environment.UseReflectionOptimizer = false;
var factory = config.BuildSessionFactory();
interceptor.SessionFactory = factory;

We set up our standard stuff – mapping and connection string. Then we need to tell NHibernate to use our custom ProxyFactoryFactory. Then we build configuration object and set our custom interceptor to it. Having done that we can build session factory and give the interceptor reference to the session factory. Certainly not the smoothest ride but gets the job done. Oh – and as I mentioned we disable reflection optimizer so that we don’t have to override yet another class (two actually).

That is all it takes to be able to get rid of all the seemingly superfluous constructors and have a much simpler model.

Well not quite – what about the activator

Right – the activator. How do we make that happen? Here’s all the code in this class.

public class NHibernateActivator : INHibernateActivator
{
	public bool CanInstantiate(Type type)
	{
		return true;
	}

	public object Instantiate(Type type)
	{
		return FormatterServices.GetUninitializedObject(type);
	}
}

Remember how I said that NHibernate can be viewed as fancy serializer? Here we make it obvious by using System.Runtime.Serialization.FormatterServices class to give us an uninitialized instance of given type. Uninitialized means that all the fields will be null or 0 and no code will be ran. This however is precisely what we want for the reasons outlined in previous post. We then throw the object to NHibernate machinery to perform all the necessary initialization for us, so that when the object is returned from the session it is fully created and usable. We could also implement a mechanism similar to one of the options that standard .NET serialization provides to allow the object to initialize its volatile state.

Final words

That is all it takes to make the model a bit more persistence ignorant. Like I said this approach won’t work always. It requires full trust environment and probably if you have complex volatile state the solution will be too simplistic. I would be interested to hear what do you think about the approach in general. Can you spot any severe shortcomings or flaws? Do you like it? Would you use it?

One of the first things you learn about NHibernate is that in order for it to be able to construct your instances and take advantage of lazy loading every persistent class must have the default, parameterless constructor. This leads to having entities looking like this (standard blog with posts and comments example).

 public class Post { private readonly IList<comment> comments = new List<comment>();
    private Blog blog;

    [Obsolete("For NHibernate")]
    protected Post()
    {
       
    }

    public Post(string title, string body, Blog blog)
    {
        Blog = blog;
        Title = title;
        Body = body;
        Published = DateTimeOffset.Now;
    }

    public virtual int Id { get; private set; }
    public virtual DateTimeOffset Published { get; private set; }

    public virtual string Title { get; set; }
    public virtual string Body { get; set; }


    public virtual IEnumerable<comment> Comments
    {
        get { return comments; }
    }

    public virtual Blog Blog
    {
        get { return blog; }
        set
        {
            if (blog != null)
            {
                throw new InvalidOperationException("already set");
            }
            blog = value;
            if (blog != null)
            {
                blog.AddPost(this);
            }
        }
    }

    [EditorBrowsable(EditorBrowsableState.Never)]
    protected internal virtual void AddComment(Comment comment)
    {
        comments.Add(comment);
    }
}

Notice the first constructor. It doesn’t do anything. As the obsolete message (which is out there to get compilation warning in case some developer accidently calls this constructor in their code) points out – this constructor is there only so that NHibernate can do its magic. Some people do put there initialization logic for collections, (especially when you use automatic properties) but I use fields and initialize them inline. In that case the constructor is pretty much useless. I started thinking why is it even there and that perhaps it doesn’t really belong to the class. But let’s start at the beginning.

What is a constructor

As basic as the question may seem, it is useful to remind ourselves why we need constructors at all. The best book about C# and .NET defines them as follows:

Constructors are special methods that allow an instance of a type to be initialized to a good state.

Notice two important things about this definition. First, it doesn’t say that constructor creates the instance or that constructors are the only way to create the instance. Second, constructors initialize newly created object to their initial state so that anything that uses the object afterwards deals with fully constructed, valid object.

Constructors and persistence

The above definition very well applies to the other constructor we have on the Post class. That constructor initializes the Post to a valid state. In this case valid means the following.

• Post is part of a blog – we can’t have a post that lives on its own. Our posts need to be part of a blog and we make this requirement explicit by requiring a Blog instance to be provided when constructing Post.
• Post requires a title and a body and that’s why we also require those two properties to be provided when constructing a post.
• Posts are usually displayed in a inverse chronological order hence we set the Published timestamp.

We do none of the above in the other, “nhibernate” constructor. That means that according to the definition of a constructor it is not really doing what a constructor is supposed to be doing. It is never used to construct an object.

Hydration

Let’s take a step back now. What NHibernate is doing with objects in nutshell is serialization. You create an object in your code and initialize it using constructor, do some stuff with it and then you save the object away, so that it can be retrieved later, after your app has been closed, or perhaps on another server instance. You save away the state of the object so that the state representation of the object can live longer than volatile, in-memory representation of the object. If you follow this path of though the next obvious conclusion is that if you have a load-balanced system and two server instances work with Post#123 they both are dealing with the same object, even though they are two separate machines.

The conclusion of that is that when NHibernate is retrieving an object from the persistent store it is not constructing it. It is recreating an in-memory representation of an object that had been created and persisted previously. Hence we are merely recreating object that already has a well known state and had been initialized and just providing different representation for it. This process is called hydration.

Persistent and Volatile state

The full picture is a bit more complicated than what I painted so far. The database and in-memory object are two representation of the same entity but they don’t have to be fully one to one. Specifically it is possible for the in-memory representation to have state beyond the persistent state. In other words the in-memory object may have some properties that are specific to it, and not relevant to the in-database representation. A convenient example that most people will be able to relate to would be a logger. Please don’t quote me as advocating using logging in your entities but logger is one of the things you may want to have on your in-memory object and use it while executing code in your application but then let them go once you no longer need the object and not persist them. If we had one in the Post class the empty constructor would change to the following:

[Obsolete("For NHibernate")]
protected Post()
{
    logger = LoggerProvider.LoggerFor(typeof(Post));
}

If we don’t use constructor for recreation of the object, how can we get the logger in? How do we make NHibernate hold the contructor semantics and give us fully initialized object? Remember I said one way of looking at NHibernate from the object’s perspective is that’s just a fancy serializer/deserializer. Turns out serialization mechanism in .NET offers us four(that I know of, possibly more) ways of tacking this issue

• you can use serialization surrogate that knows how to recreate the full state of the object
• you can use deserialization callback interface to be notified when the object has been fully deserialized and then initialize the volatile state.
• you can use serialization callback attributes to be notified when various steps in the serialization/deserialization process happen and initialize the volatile state.
• you can use ISerializable interface and implement serialization constructor which is used to deserialize the object.

Notice that only one of those approaches uses special constructor. Since as we discussed NHibernate doesn’t really need the default constructor (in theory that is), can we really get rid of it? Turns one we can (in most cases), and we’ll look at how to do it in the next post.