NHibernate - why is this static method legal in this domain class? - nhibernate

I have read in Chapter 4 of the NHibernate docs that all of a persistent classes public methods, properties and events must be declared as virtual.
However, whilst a runtime error is generated for any Properties that are not marked as virtual, I have found that static methods are allowed and do not generate a runtime error . As they are static they are of course not marked virtual which seems to break the rule in point 4.1.4 of the documentation (see above). I have checked the resulting sql and it also implements lazy loading correctly when I run a test against the method so is it therefore ok to use static methods?
Here's the basic details of the persistant class:
public class CmsPage
{
public virtual int? Id { get; set; }
public virtual string Title { get; set; }
public virtual void Update()
{
using (ISession session = NHibernateHelper.OpenSession())
{
using (ITransaction transaction = session.BeginTransaction())
{
session.Update(this);
transaction.Commit();
}
}
}
// Note: static and non-virtual and yet it will not cause a problem for Nhibernate
public static IEnumerable<CmsPage> GetList()
{
IList<CmsPage> pageList;
using (ISession session = NHibernateHelper.OpenSession())
{
string hql = "from CmsPage p";
pageList = session.CreateQuery(hql)
.List<CmsPage>();
}
return pageList;
}
}
So my question is why is it ok to use a static method in the persistent domain class when the documentation seems to say it's not?
Please answer from NHibernate's point of view not an OO design point of view; I don't want to get into an OOD/OOP debate if it can be avoided please.

The documentation says: "NHibernate works best if these classes follow some simple rules, ..." It doesn't say it won't work (clearly it does work).
So, really, the discussion boils down to an OO issue.

Actually this applies only to properties. Methods are not persisted, so proxies and lazy-loading does not apply. Ideally you should separate data access (the static methods in your case) from the domain object. But you are correct to point this out, maybe the documentation should have been clearer.
In conclusion your class is perfectly fine but it could be even better if you separated the concerns.

NHibernate needs all your properties to be virtual because it carries out its lazy-loading magic by making proxies of your objects that override everything. So when you write this code:
class Foo {
public virtual Foo[] Neighbors { get; set; }
}
NHibernate secretly generates classes like:
class NHProxy03450843275 : Foo {
public virtual Foo[] Neighbors { /* Godawful lazy-loading magic goes here */ }
}
Actually it's worse than that, but this gives you the idea. Anyway, static methods aren't bound to particular instances of a class, so NH doesn't need proxies to deal with them. Thus they can be non-virtual.

Related

Problem with Include() EntityFramework Core with blazor server side [duplicate]

I had seen some books(e.g programming entity framework code first Julia Lerman) define their domain classes (POCO) with no initialization of the navigation properties like:
public class User
{
public int Id { get; set; }
public string UserName { get; set; }
public virtual ICollection<Address> Address { get; set; }
public virtual License License { get; set; }
}
some other books or tools (e.g Entity Framework Power Tools) when generates POCOs initializes the navigation properties of the the class, like:
public class User
{
public User()
{
this.Addresses = new IList<Address>();
this.License = new License();
}
public int Id { get; set; }
public string UserName { get; set; }
public virtual ICollection<Address> Addresses { get; set; }
public virtual License License { get; set; }
}
Q1: Which one is better? why? Pros and Cons?
Edit:
public class License
{
public License()
{
this.User = new User();
}
public int Id { get; set; }
public string Key { get; set; }
public DateTime Expirtion { get; set; }
public virtual User User { get; set; }
}
Q2: In second approach there would be stack overflow if the `License` class has a reference to `User` class too. It means we should have one-way reference.(?) How we should decide which one of the navigation properties should be removed?
Collections: It doesn't matter.
There is a distinct difference between collections and references as navigation properties. A reference is an entity. A collections contains entities. This means that initializing a collection is meaningless in terms of business logic: it does not define an association between entities. Setting a reference does.
So it's purely a matter of preference whether or not, or how, you initialize embedded lists.
As for the "how", some people prefer lazy initialization:
private ICollection<Address> _addresses;
public virtual ICollection<Address> Addresses
{
get { return this._addresses ?? (this._addresses = new HashSet<Address>());
}
It prevents null reference exceptions, so it facilitates unit testing and manipulating the collection, but it also prevents unnecessary initialization. The latter may make a difference when a class has relatively many collections. The downside is that it takes relatively much plumbing, esp. when compared to auto properties without initialization. Also, the advent of the null-propagation operator in C# has made it less urgent to initialize collection properties.
...unless explicit loading is applied
The only thing is that initializing collections makes it hard to check whether or not a collection was loaded by Entity Framework. If a collection is initialized, a statement like...
var users = context.Users.ToList();
...will create User objects having empty, not-null Addresses collections (lazy loading aside). Checking whether the collection is loaded requires code like...
var user = users.First();
var isLoaded = context.Entry(user).Collection(c => c.Addresses).IsLoaded;
If the collection is not initialized a simple null check will do. So when selective explicit loading is an important part of your coding practice, i.e. ...
if (/*check collection isn't loaded*/)
context.Entry(user).Collection(c => c.Addresses).Load();
...it may be more convenient not to initialize collection properties.
Reference properties: Don't
Reference properties are entities, so assigning an empty object to them is meaningful.
Worse, if you initiate them in the constructor, EF won't overwrite them when materializing your object or by lazy loading. They will always have their initial values until you actively replace them. Worse still, you may even end up saving empty entities in the database!
And there's another effect: relationship fixup won't occcur. Relationship fixup is the process by which EF connects all entities in the context by their navigation properties. When a User and a Licence are loaded separately, still User.License will be populated and vice versa. Unless of course, if License was initialized in the constructor. This is also true for 1:n associations. If Address would initialize a User in its constructor, User.Addresses would not be populated!
Entity Framework core
Relationship fixup in Entity Framework core (2.1 at the time of writing) isn't affected by initialized reference navigation properties in constructors. That is, when users and addresses are pulled from the database separately, the navigation properties are populated.
However, lazy loading does not overwrite initialized reference navigation properties.
In EF-core 3, initializing a reference navigation property prevents Include from working properly.
So, in conclusion, also in EF-core, initializing reference navigation properties in constructors may cause trouble. Don't do it. It doesn't make sense anyway.
In all my projects I follow the rule - "Collections should not be null. They are either empty or have values."
First example is possible to have when creation of these entities is responsibility of third-part code (e.g. ORM) and you are working on a short-time project.
Second example is better, since
you are sure that entity has all properties set
you avoid silly NullReferenceException
you make consumers of your code happier
People, who practice Domain-Driven Design, expose collections as read-only and avoid setters on them. (see What is the best practice for readonly lists in NHibernate)
Q1: Which one is better? why? Pros and Cons?
It is better to expose not-null colections since you avoid additional checks in your code (e.g. Addresses). It is a good contract to have in your codebase. But it os OK for me to expose nullable reference to single entity (e.g. License)
Q2: In second approach there would be stack overflow if the License class has a reference to User class too. It means we should have one-way reference.(?) How we should decide which one of the navigation properties should be removed?
When I developed data mapper pattern by myself I tryed to avoid bidirectional references and had reference from child to parent very rarely.
When I use ORMs it is easy to have bidirectional references.
When it is needed to build test-entity for my unit-tests with bidirectional reference set I follow the following steps:
I build parent entity with emty children collection.
Then I add evey child with reference to parent entity into children collection.
Insted of having parameterless constructor in License type I would make user property required.
public class License
{
public License(User user)
{
this.User = user;
}
public int Id { get; set; }
public string Key { get; set; }
public DateTime Expirtion { get; set; }
public virtual User User { get; set; }
}
It's redundant to new the list, since your POCO is depending on Lazy Loading.
Lazy loading is the process whereby an entity or collection of entities is automatically loaded from the database the first time that a property referring to the entity/entities is accessed. When using POCO entity types, lazy loading is achieved by creating instances of derived proxy types and then overriding virtual properties to add the loading hook.
If you would remove the virtual modifier, then you would turn off lazy loading, and in that case your code no longer would work (because nothing would initialize the list).
Note that Lazy Loading is a feature supported by entity framework, if you create the class outside the context of a DbContext, then the depending code would obviously suffer from a NullReferenceException
HTH
The other answers fully answer the question, but I'd like to add something since this question is still relevant and comes up in google searches.
When you use the "code first model from database" wizard in Visual Studio all collections are initialized like so:
public partial class SomeEntity
{
[System.Diagnostics.CodeAnalysis.SuppressMessage("Microsoft.Usage", "CA2214:DoNotCallOverridableMethodsInConstructors")]
public SomeEntity()
{
OtherEntities = new HashSet<OtherEntity>();
}
public int Id { get; set; }
[System.Diagnostics.CodeAnalysis.SuppressMessage("Microsoft.Usage", "CA2227:CollectionPropertiesShouldBeReadOnly")]
public virtual ICollection<OtherEntity> OtherEntities { get; set; }
}
I tend to take wizard output as basically being an official recommendation from Microsoft, hence why I'm adding to this five-year-old question. Therefore, I'd initialize all collections as HashSets.
And personally, I think it'd be pretty slick to tweak the above to take advantage of C# 6.0's auto-property initializers:
public virtual ICollection<OtherEntity> OtherEntities { get; set; } = new HashSet<OtherEntity>();
Q1: Which one is better? why? Pros and Cons?
The second variant when virtual properties are set inside an entity constructor has a definite problem which is called "Virtual member call in a constructor".
As for the first variant with no initialization of navigation properties, there are 2 situations depending on who / what creates an object:
Entity framework creates an object
Code consumer creates an object
The first variant is perfectly valid when Entity Framework creates a object,
but can fail when a code consumer creates an object.
The solution to ensure a code consumer always creates a valid object is to use a static factory method:
Make default constructor protected. Entity Framework is fine to work with protected constructors.
Add a static factory method that creates an empty object, e.g. a User object, sets all properties, e.g. Addresses and License, after creation and returns a fully constructed User object
This way Entity Framework uses a protected default constructor to create a valid object from data obtained from some data source and code consumer uses a static factory method to create a valid object.
I use the answer from this Why is my Entity Framework Code First proxy collection null and why can't I set it?
Had problems with constructor initilization. Only reason I do this is to make test code easier. Making sure collection is never null saves me constantly initialising in tests etc

Repository OO Design - Multiple Specifications

I have a pretty standard repository interface:
public interface IRepository<TDomainEntity>
where TDomainEntity : DomainEntity, IAggregateRoot
{
TDomainEntity Find(Guid id);
void Add(TDomainEntity entity);
void Update(TDomainEntity entity);
}
We can use various infrastructure implementations in order to provide default functionality (e.g. Entity Framework, DocumentDb, Table Storage, etc). This is what the Entity Framework implementation looks like (without any actual EF code, for simplicity sake):
public abstract class EntityFrameworkRepository<TDomainEntity, TDataEntity> : IRepository<TDomainEntity>
where TDomainEntity : DomainEntity, IAggregateRoot
where TDataEntity : class, IDataEntity
{
protected IEntityMapper<TDomainEntity, TDataEntity> EntityMapper { get; private set; }
public TDomainEntity Find(Guid id)
{
// Find, map and return entity using Entity Framework
}
public void Add(TDomainEntity item)
{
var entity = EntityMapper.CreateFrom(item);
// Insert entity using Entity Framework
}
public void Update(TDomainEntity item)
{
var entity = EntityMapper.CreateFrom(item);
// Update entity using Entity Framework
}
}
There is a mapping between the TDomainEntity domain entity (aggregate) and the TDataEntity Entity Framework data entity (database table). I will not go into detail as to why there are separate domain and data entities. This is a philosophy of Domain Driven Design (read about aggregates). What's important to understand here is that the repository will only ever expose the domain entity.
To make a new repository for, let's say, "users", I could define the interface like this:
public interface IUserRepository : IRepository<User>
{
// I can add more methods over and above those in IRepository
}
And then use the Entity Framework implementation to provide the basic Find, Add and Update functionality for the aggregate:
public class UserRepository : EntityFrameworkRepository<Stop, StopEntity>, IUserRepository
{
// I can implement more methods over and above those in IUserRepository
}
The above solution has worked great. But now we want to implement deletion functionality. I have proposed the following interface (which is an IRepository):
public interface IDeleteableRepository<TDomainEntity>
: IRepository<TDomainEntity>
{
void Delete(TDomainEntity item);
}
The Entity Framework implementation class would now look something like this:
public abstract class EntityFrameworkRepository<TDomainEntity, TDataEntity> : IDeleteableRepository<TDomainEntity>
where TDomainEntity : DomainEntity, IAggregateRoot
where TDataEntity : class, IDataEntity, IDeleteableDataEntity
{
protected IEntityMapper<TDomainEntity, TDataEntity> EntityMapper { get; private set; }
// Find(), Add() and Update() ...
public void Delete(TDomainEntity item)
{
var entity = EntityMapper.CreateFrom(item);
entity.IsDeleted = true;
entity.DeletedDate = DateTime.UtcNow;
// Update entity using Entity Framework
// ...
}
}
As defined in the class above, the TDataEntity generic now also needs to be of type IDeleteableDataEntity, which requires the following properties:
public interface IDeleteableDataEntity
{
bool IsDeleted { get; set; }
DateTime DeletedDate { get; set; }
}
These properties are set accordingly in the Delete() implementation.
This means that, IF required, I can define IUserRepository with "deletion" capabilities which would inherently be taken care of by the relevant implementation:
public interface IUserRepository : IDeleteableRepository<User>
{
}
Provided that the relevant Entity Framework data entity is an IDeleteableDataEntity, this would not be an issue.
The great thing about this design is that I can start granualising the repository model even further (IUpdateableRepository, IFindableRepository, IDeleteableRepository, IInsertableRepository) and aggregate repositories can now expose only the relevant functionality as per our specification (perhaps you should be allowed to insert into a UserRepository but NOT into a ClientRepository). Further to this, it specifies a standarised way in which certain repository actions are done (i.e. the updating of IsDeleted and DeletedDate columns will be universal and are not at the hand of the developer).
PROBLEM
A problem with the above design arises when I want to create a repository for some aggregate WITHOUT deletion capabilities, e.g:
public interface IClientRepository : IRepository<Client>
{
}
The EntityFrameworkRepository implementation still requires TDataEntity to be of type IDeleteableDataEntity.
I can ensure that the client data entity model does implement IDeleteableDataEntity, but this is misleading and incorrect. There will be additional fields that are never updated.
The only solution I can think of is to remove the IDeleteableDataEntity generic condition from TDataEntity and then cast to the relevant type in the Delete() method:
public abstract class EntityFrameworkRepository<TDomainEntity, TDataEntity> : IDeleteableRepository<TDomainEntity>
where TDomainEntity : DomainEntity, IAggregateRoot
where TDataEntity : class, IDataEntity
{
protected IEntityMapper<TDomainEntity, TDataEntity> EntityMapper { get; private set; }
// Find() and Update() ...
public void Delete(TDomainEntity item)
{
var entity = EntityMapper.CreateFrom(item);
var deleteableEntity = entity as IDeleteableEntity;
if(deleteableEntity != null)
{
deleteableEntity.IsDeleted = true;
deleteableEntity.DeletedDate = DateTime.UtcNow;
entity = deleteableEntity;
}
// Update entity using Entity Framework
// ...
}
}
Because ClientRepository does not implement IDeleteableRepository, there will be no Delete() method exposed, which is good.
QUESTION
Can anyone advise of a better architecture which leverages the C# typing system and does not involve the hacky cast?
Interestly enough, I could do this if C# supported multiple inheritance (with separate concrete implementation for finding, adding, deleting, updating).
I do think that you're complicating things a bit too much trying to get the most generic solution of them all, however I think there's a pretty easy solution to your current problem.
TDataEntity is a persistence data structure, it has no Domain value and it's not known outside the persistence layer. So it can have fields it won't ever use, the repository is the only one knowing that, it'a persistence detail . You can afford to be 'sloppy' here, things aren't that important at this level.
Even the 'hacky' cast is a good solution because it's in one place and a private detail.
It's good to have clean and maintainable code everywhere, however we can't afford to waste time coming up with 'perfect' solutions at every layer. Personally, for view and persistence models I prefer the quickest and simplest solutions even if they're a bit smelly.
P.S: As a thumb rule, generic repository interfaces are good, generic abstract repositories not so much (you need to be careful) unless you're serializing things or using a doc db.

Using Test Doubles with DbEntityEntry and DbPropertyEntry

I am using the new Test Doubles in EF6 as outlined here from MSDN . VS2013 with Moq & nUnit.
All was good until I had to do something like this:
var myFoo = context.Foos.Find(id);
and then:
myFoo.Name = "Bar";
and then :
context.Entry(myFoo).Property("Name").IsModified = true;
At this point is where I get an error:
Additional information: Member 'IsModified' cannot be called for
property 'Name' because the entity of type
'Foo' does not exist in the context. To add an
entity to the context call the Add or Attach method of
DbSet.
Although, When I examine the 'Foos' in the context with an AddWatch I can see all items I Add'ed before running the test. So they are there.
I have created the FakeDbSet (or TestDbSet) from the article. I am putting each FakeDbSet in the FakeContext at the constructor where each one gets initialized. Like this:
Foos = new FakeDbSet<Foo>();
My question is, is it possible to work with the FakeDbSet and the FakeContext with the test doubles scenario in such a way to have access to DbEntityEntry and DBPropertyEntry from the test double? Thanks!
I can see all items I Add'ed before running the test. So they are there.
Effectively, you've only added items to an ObservableCollection. The context.Entry method reaches much deeper than that. It requires a change tracker to be actively involved in adding, modifying and removing entities. If you want to mock this change tracker, the ObjectStateManager (ignoring the fact that it's not designed to be mocked at all), good luck! It's got over 4000 lines of code.
Frankly, I don't understand all these blogs and articles about mocking EF. Only the numerous differences between LINQ to objects and LINQ to entites should be enough to discourage it. These mock contexts and DbSets build an entirely new universe that's a source of bugs in itself. I've decided to do integrations test only when and wherever EF is involved in my code. A working end-to-end test gives me a solid feeling that things are OK. A unit test (faking EF) doesn't. (Others do, don't get me wrong).
But let's assume you'd still like to venture into mocking DbContext.Entry<T>. Too bad, impossible.
The method is not virtual
It returns a DbEntityEntry<T>, a class with an internal constructor, that is a wrapper around an InternalEntityEntry, which is an internal class. And, by the way, DbEntityEntry doesn't implement an interface.
So, to answer your question
is it possible to (...) have access to DbEntityEntry and DBPropertyEntry from the test double?
No, EF's mocking hooks are only very superficial, you'll never even come close to how EF really works.
Just abstract it. If you are working against an interface, when creating your own doubles, put the modified stuff in a seperate method. My interface and implementation (generated by EF, but I altered the template) look like this:
//------------------------------------------------------------------------------
// <auto-generated>
// This code was generated from a template.
//
// Manual changes to this file may cause unexpected behavior in your application.
// Manual changes to this file will be overwritten if the code is regenerated.
// </auto-generated>
//------------------------------------------------------------------------------
namespace Model
{
using System;
using System.Data.Entity;
using System.Data.Entity.Infrastructure;
public interface IOmt
{
DbSet<DatabaseOmtObjectWhatever> DatabaseOmtObjectWhatever { get; set; }
int SaveChanges();
void SetModified(object entity);
void SetAdded(object entity);
}
public partial class Omt : DbContext, IOmt
{
public Omt()
: base("name=Omt")
{
}
protected override void OnModelCreating(DbModelBuilder modelBuilder)
{
throw new UnintentionalCodeFirstException();
}
public virtual DbSet<DatabaseOmtObjectWhatever> DatabaseOmtObjectWhatever { get; set; }
public void SetModified(object entity)
{
Entry(entity).State = EntityState.Modified;
}
public void SetAdded(object entity)
{
Entry(entity).State = EntityState.Added;
}
}
}

Optimum Way To Restore Domain Object

This is such a simple and common scenario I wonder how did I managed until now and why I have problems now.
I have this object (part of the Infrastructure assembly)
public class Queue {}
public class QueueItem
{
public QueueItem(int blogId,string name,Type command,object data)
{
if (name == null) throw new ArgumentNullException("name");
if (command == null) throw new ArgumentNullException("command");
BlogId = blogId;
CommandType = command;
ParamValue = data;
CommandName = name;
AddedOn = DateTime.UtcNow;
}
public Guid Id { get; internal set; }
public int BlogId { get; private set; }
public string CommandName { get; set; }
public Type CommandType { get; private set; }
public object ParamValue { get; private set; }
public DateTime AddedOn { get; private set; }
public DateTime? ExecutedOn { get; private set; }
public void ExecuteIn(ILifetimeScope ioc)
{
throw new NotImplementedException();
}
}
This will be created in another assembly like this
var qi = new QueueItem(1,"myname",typeof(MyCommand),null);
Nothing unusal here. However, this object will be sent t oa repository where it will be persisted.The Queue object will ask the repository for items. The repository should re-create QueueItem objects.
However, as you see, the QueueItem properties are invariable, the AddedOn property should be set only once when the item is created. The Id property will be set by the Queue object (this is not important).
The question is how should I recreate the QueueItem in the repository? I can have another constructor which will require every value for ALL the properties, but I don't want that constructor available for the assembly that will create the queue item initially. The repository is part of a different assembly so internal won't work.
I thought about providing a factory method
class QueueItem
{
/* ..rest of definitions.. */
public static QueueItem Restore(/* list of params*/){}
}
which at least clears the intent, but I don't know why I don't like this approach. I could also enforce the item creation only by the Queue , but that means to pass the Queue as a dependency to the repo which again isn't something I'd like. To have a specific factory object for this, also seems way overkill.
Basically my question is: what is the optimum way to recreate an object in the repository, without exposing that specific creational functionality to another consumer object.
Update
It's important to note that by repository I mean the pattern itself as an abstraction, not a wrapper over an ORM. It doesn't matter how or where the domain objects are persisted. It matters how can be re-created by the repository. Another important thing is that my domain model is different from the persistence model. I do use a RDBMS but I think this is just an implementation detail which should not bear any importance, since I'm looking for way that doesn't depend on a specific storage access.
While this is a specific scenario, it can applied to basically every object that will be restored by the repo.
Update2
Ok I don't know how I could forget about AutoMapper. I was under the wrong impression it can't map private fields/setter but it can, and I think this is the best solution.
In fact I can say the optimum solutions (IMO) are in order:
Directly deserializing, if available.
Automap.
Factory method on the domain object itself.
The first two don't require the object to do anyting in particular, while the third requires the object to provide functionality for that case (a way to enter valid state data). It has clear intent but it pretty much does a mapper job.
Answer Updated
To answer myself, in this case the optimum way is to use a factory method. Initially I opted for the Automapper but I found myself using the factory method more often. Automapper can be useful sometimes but in quite a lot of cases it's not enough.
An ORM framework would take care of that for you. You just have to tell it to rehydrate an object and a regular instance of the domain class will be served to you (sometimes you only have to declare properties as virtual or protected, in NHibernate for instance). The reason is because under the hood, they usually operate on proxy objects derived from your base classes, allowing you to keep these base classes intact.
If you want to implement your own persistence layer though, it's a whole nother story. Rehydrating an object from the database without breaking the scope constraints originally defined in the object is likely to involve reflection. You also have to think about a lot of side concerns : if your object has a reference to another object, you must rehydrate that one before, etc.
You can have a look at that tutorial : Build Your Own dataAccess Layer although I wouldn't recommend reinventing the wheel in most cases.
You talked about a factory method on the object itself. But DDD states that entities should be created by a factory. So you should have a QueueItemFactory that can create new QueueItems and restore existing QueueItems.
Ok I don't know how I could forget about AutoMapper.
I wish I could forget about AutoMapper. Just looking at the hideous API gives me shivers down my spine.

NHibernate exception: method Add should be 'public/protected virtual' or 'protected internal virtual'

Take this class as example:
public class Category : PersistentObject<int>
{
public virtual string Title { get; set; }
public virtual string Alias { get; set; }
public virtual Category ParentCategory { get; set; }
public virtual ISet<Category> ChildCategories { get; set; }
public /*virtual*/ void Add(Category child)
{
if (child != null)
{
child.ParentCategory = this;
ChildCategories.Add(child);
}
}
}
When running the application without the virtual keyword of add method, I getting this error:
method Add should be 'public/protected virtual' or 'protected internal virtual'
I understand why properties need to declare as virtual, because thay need to be overridden by the lazy loading feature.
But I don't understand why Methods need to be declare as virtual... they need to be overridden for what reason?
This very confusing!
Methods as well need to be virtual because they could access fields. Consider this situation:
class Entity
{
//...
private int a;
public virtual int A
{
get { return a; }
}
public virtual void Foo()
{
// lazy loading is implemented here by the proxy
// to make the value of a available
if (a > 7)
// ...
}
}
I believe this is required for the lazy-loading feature in NHibernate where NHibernate creates proxies of your entity and controls all access to it. This is why every single method and property must be virtual. Basically, if there is a member doing anything with the entity, NH need to know about it and tap into it.
Like mentioned earlier, in order for NHibernate to do the 'magic' it creates proxy classes which inherit from your entities (Category in your case). However, if you make your entities implement an interface, it will use that interface to create a proxy instead of concrete types. This way, you wouldn't have to mark everything virtual.
EDIT: Some corrections... According to this, i am compelled to say that it almost looks like NH doesn't really do anything with virtual methods, after all. And i even read someone saying that they removed this run-time check from the NH core assembly just to get around it. My assumption would be that it is an older requirements which hasn't been removed. The cool thing is that it looks like there is an initiative to use PostSharp for static proxies, so your classes won't have to declare anything virtual for NH to generate proxies. The bad thing is that it looks like it's been stuck in a branch for almost two years.