NHibernate collections are considered to be a logical part of their owning entity; never of the contained entities. This is a crucial distinction! It has the following consequences:

Instead, the default behavior is that adding an entity to a collection merely creates a link between the two entities, while removing it removes the link. This is very appropriate for all sorts of cases. Where it is not appropriate at all is the case of a parent / child relationship, where the life of the child is bound to the lifecycle of the parent.

Suppose we start with a simple <one-to-many> association from Parent to Child.

<set name="Children">
    <key column="parent_id" />
    <one-to-many class="Child" />
</set>

If we were to execute the following code

Parent p = .....;
Child c = new Child();
p.Children.Add(c);
session.Save(c);
session.Flush();

NHibernate would issue two SQL statements:

This is not only inefficient, but also violates any NOT NULL constraint on the parent_id column.

The underlying cause is that the link (the foreign key parent_id) from p to c is not considered part of the state of the Child object and is therefore not created in the INSERT. So the solution is to make the link part of the Child mapping.

<many-to-one name="Parent" column="parent_id" not-null="true"/>

(We also need to add the Parent property to the Child class.)

Now that the Child entity is managing the state of the link, we tell the collection not to update the link. We use the inverse attribute.

<set name="Children" inverse="true">
    <key column="parent_id"/>
    <one-to-many class="Child"/>
</set>

The following code would be used to add a new Child.

Parent p = (Parent) session.Load(typeof(Parent), pid);
Child c = new Child();
c.Parent = p;
p.Children.Add(c);
session.Save(c);
session.Flush();

And now, only one SQL INSERT would be issued!

To tighten things up a bit, we could create an AddChild() method of Parent.

public void AddChild(Child c)
{
    c.Parent = this;
    children.Add(c);
}

Now, the code to add a Child looks like

Parent p = (Parent) session.Load(typeof(Parent), pid);
Child c = new Child();
p.AddChild(c);
session.Save(c);
session.Flush();

The explicit call to Save() is still annoying. We will address this by using cascades.

<set name="Children" inverse="true" cascade="all">
    <key column="parent_id"/>
    <one-to-many class="Child"/>
</set>

This simplifies the code above to

Parent p = (Parent) session.Load(typeof(Parent), pid);
Child c = new Child();
p.AddChild(c);
session.Flush();

Similarly, we don't need to iterate over the children when saving or deleting a Parent. The following removes p and all its children from the database.

Parent p = (Parent) session.Load(typeof(Parent), pid);
session.Delete(p);
session.Flush();

However, this code

Parent p = (Parent) session.Load(typeof(Parent), pid);
// Get one child out of the set
IEnumerator childEnumerator = p.Children.GetEnumerator();
childEnumerator.MoveNext();
Child c = (Child) childEnumerator.Current;

p.Children.Remove(c);
c.Parent = null;
session.Flush();

will not remove c from the database; it will only remove the link to p (and cause a NOT NULL constraint violation, in this case). You need to explicitly Delete() the Child.

Parent p = (Parent) session.Load(typeof(Parent), pid);
// Get one child out of the set
IEnumerator childEnumerator = p.Children.GetEnumerator();
childEnumerator.MoveNext();
Child c = (Child) childEnumerator.Current;

p.Children.Remove(c);
session.Delete(c);
session.Flush();

Now, in our case, a Child can't really exist without its parent. So if we remove a Child from the collection, we really do want it to be deleted. For this, we must use cascade="all-delete-orphan".

<set name="Children" inverse="true" cascade="all-delete-orphan">
    <key column="parent_id"/>
    <one-to-many class="Child"/>
</set>

Note: even though the collection mapping specifies inverse="true", cascades are still processed by iterating the collection elements. So if you require that an object be saved, deleted or updated by cascade, you must add it to the collection. It is not enough to simply set its parent.

Suppose we loaded up a Parent in one ISession, made some changes in a UI action and wish to persist these changes in a new ISession (by calling Update()). The Parent will contain a collection of children and, since cascading update is enabled, NHibernate needs to know which children are newly instantiated and which represent existing rows in the database. Let's assume that both Parent and Child have (synthetic) identifier properties of type long. NHibernate will use the identifier property value to determine which of the children are new. (You may also use the version or timestamp property, see Section 9.4.2, “Updating detached objects”.)

The unsaved-value attribute is used to specify the identifier value of a newly instantiated instance. In NHibernate it is not necessary to specify unsaved-value explicitly.

The following code will update parent and child and insert newChild.

//parent and child were both loaded in a previous session
parent.AddChild(child);
Child newChild = new Child();
parent.AddChild(newChild);
session.Update(parent);
session.Flush();

Well, thats all very well for the case of a generated identifier, but what about assigned identifiers and composite identifiers? This is more difficult, since unsaved-value can't distinguish between a newly instantiated object (with an identifier assigned by the user) and an object loaded in a previous session. In these cases, you will probably need to give NHibernate a hint; either

null is the default unsaved-value for assigned identifiers, none is the default unsaved-value for composite identifiers.

There is one further possibility. There is a new IInterceptor method named IsUnsaved() which lets the application implement its own strategy for distinguishing newly instantiated objects. For example, you could define a base class for your persistent classes.

public class Persistent
{
    private bool _saved = false;
    
    public void OnSave()
    {
        _saved=true;
    }
    
    public void OnLoad()
    {
        _saved=true;
    }
    
    ......
    
    public bool IsSaved
    {
        get { return _saved; }
    }
}

(The saved property is non-persistent.) Now implement IsUnsaved(), along with OnLoad() and OnSave() as follows.

public object IsUnsaved(object entity)
{
    if (entity is Persistent)
    {
        return !( (Persistent) entity ).IsSaved;
    }
    else
    {
        return null;
    }
}

public bool OnLoad(object entity, 
    object id,
    object[] state,
    string[] propertyNames,
    IType[] types)
{
    if (entity is Persistent) ( (Persistent) entity ).OnLoad();
    return false;
}

public boolean OnSave(object entity,
    object id,
    object[] state,
    string[] propertyNames,
    IType[] types)
{
    if (entity is Persistent) ( (Persistent) entity ).OnSave();
    return false;
}

There is quite a bit to digest here and it might look confusing first time around. However, in practice, it all works out quite nicely. Most NHibernate applications use the parent / child pattern in many places.

We mentioned an alternative in the first paragraph. None of the above issues exist in the case of <composite-element> mappings, which have exactly the semantics of a parent / child relationship. Unfortunately, there are two big limitations to composite element classes: composite elements may not own collections, and they should not be the child of any entity other than the unique parent. (However, they may have a surrogate primary key, using an <idbag> mapping.)