NHibernate requires that persistent collection-valued fields be declared as a generic interface type, for example:

public class Product
{
    public ISet<Part> Parts { get; set; } = new HashSet<Part>();

    public string SerialNumber { get; set; }
}

The actual interface might be System.Collections.Generic.ICollection<T>, System.Collections.Generic.IList<T>, System.Collections.Generic.IDictionary<K, V>, System.Collections.Generic.ISet<T> or ... anything you like! (Where "anything you like" means you will have to write an implementation of NHibernate.UserType.IUserCollectionType.)

Notice how we initialized the instance variable with an instance of HashSet<T>. This is the best way to initialize collection valued properties of newly instantiated (non-persistent) instances. When you make the instance persistent - by calling Save(), for example - NHibernate will actually replace the HashSet<T> with an instance of NHibernate's own implementation of ISet<T>. Watch out for errors like this:

Cat cat = new DomesticCat();
Cat kitten = new DomesticCat();
....
ISet<Cat> kittens = new HashSet<Cat>();
kittens.Add(kitten);
cat.Kittens = kittens;
session.Save(cat);
kittens = cat.Kittens; //Okay, kittens collection is an ISet
HashSet<Cat> hs = (HashSet<Cat>) cat.Kittens; //Error!

Collection instances have the usual behavior of value types. They are automatically persisted when referenced by a persistent object and automatically deleted when unreferenced. If a collection is passed from one persistent object to another, its elements might be moved from one table to another. Two entities may not share a reference to the same collection instance. Due to the underlying relational model, collection-valued properties do not support null value semantics; NHibernate does not distinguish between a null collection reference and an empty collection.

You shouldn't have to worry much about any of this. Just use NHibernate's collections the same way you use ordinary .NET collections, but make sure you understand the semantics of bidirectional associations (discussed later) before using them.

Collection instances are distinguished in the database by a foreign key to the owning entity. This foreign key is referred to as the collection key . The collection key is mapped by the <key> element.

Collections may contain almost any other NHibernate type, including all basic types, custom types, entity types and components. This is an important definition: An object in a collection can either be handled with "pass by value" semantics (it therefore fully depends on the collection owner) or it can be a reference to another entity with an own lifecycle. Collections may not contain other collections. The contained type is referred to as the collection element type. Collection elements are mapped by <element>, <composite-element>, <one-to-many>, <many-to-many> or <many-to-any>. The first two map elements with value semantics, the other three are used to map entity associations.

All collection types except ISet and bag have an index column - a column that maps to an array or IList index or IDictionary key. The index of an IDictionary may be of any basic type, an entity type or even a composite type (it may not be a collection). The index of an array or list is always of type Int32. Indexes are mapped using <index>, <index-many-to-many>, <composite-index> or <index-many-to-any>.

There are quite a range of mappings that can be generated for collections, covering many common relational models. We suggest you experiment with the schema generation tool to get a feeling for how various mapping declarations translate to database tables.

Collections are declared by the <set>, <list>, <map>, <bag>, <array> and <primitive-array> elements. <map> is representative:

<map
    name="propertyName"                                         (1)
    table="table_name"                                          (2)
    schema="schema_name"                                        (3)
    lazy="true|false|extra"                                     (4)
    inverse="true|false"                                        (5)
    cascade="all|none|save-update|delete|all-delete-orphan"     (6)
    sort="unsorted|natural|comparatorClass"                     (7)
    order-by="column_name asc|desc"                             (8)
    where="arbitrary sql where condition"                       (9)
    fetch="select|join"                                         (10)
    batch-size="N"                                              (11)
    access="field|property|ClassName"                           (12)
    optimistic-lock="true|false"                                (13)
    generic="true|false"                                        (14)
>

    <key .... />
    <index .... />
    <element .... />
</map>

The mapping of an IList or array requires a separate table column holding the array or list index (the i in foo[i]). If your relational model doesn't have an index column, e.g. if you're working with legacy data, use an unordered ISet instead. This seems to put people off who assume that IList should just be a more convenient way of accessing an unordered collection. NHibernate collections strictly obey the actual semantics attached to the ISet, IList and IDictionary interfaces. IList elements don't just spontaneously rearrange themselves!

On the other hand, people who planned to use the IList to emulate bag semantics have a legitimate grievance here. A bag is an unordered, unindexed collection which may contain the same element multiple times. The .NET collections framework lacks an IBag interface, hence you have to emulate it with an IList. NHibernate lets you map properties of type IList or ICollection with the <bag> element. Note that bag semantics are not really part of the ICollection contract and they actually conflict with the semantics of the IList contract (however, you can sort the bag arbitrarily, discussed later in this chapter).

Note: Large NHibernate bags mapped with inverse="false" are inefficient and should be avoided; NHibernate can't create, delete or update rows individually, because there is no key that may be used to identify an individual row.

A collection table is required for any collection of values and any collection of references to other entities mapped as a many-to-many association (the natural semantics for a .NET collection). The table requires (foreign) key column(s), element column(s) and possibly index column(s).

The foreign key from the collection table to the table of the owning class is declared using a <key> element.

<key column="column_name"/>

For indexed collections like maps and lists, we require an <index> element. For lists, this column contains sequential integers numbered from zero. Make sure that your index really starts from zero if you have to deal with legacy data. For maps, the column may contain any values of any NHibernate type.

<index
        column="column_name"                (1)
        type="typename"                     (2)
/>

Alternatively, a map may be indexed by objects of entity type. We use the <index-many-to-many> element.

<index-many-to-many
        column="column_name"                (1)
        class="ClassName"                   (2)
/>

For a collection of values, we use the <element> tag.

<element
        column="column_name"                (1)
        type="typename"                     (2)
/>

A collection of entities with its own table corresponds to the relational notion of many-to-many association. A many to many association is the most natural mapping of a .NET collection but is not usually the best relational model.

<many-to-many
        column="column_name"                               (1)
        class="ClassName"                                  (2)
        fetch="join|select"                                (3)
        not-found="ignore|exception"                       (4)
    />

Some examples, first, a set of strings:

<set name="Names" table="NAMES">
    <key column="GROUPID"/>
    <element column="NAME" type="String"/>
</set>

A bag containing integers (with an iteration order determined by the order-by attribute):

<bag name="Sizes" table="SIZES" order-by="SIZE ASC">
    <key column="OWNER"/>
    <element column="SIZE" type="Int32"/>
</bag>

An array of entities - in this case, a many to many association (note that the entities are lifecycle objects, cascade="all"):

<array name="Foos" table="BAR_FOOS" cascade="all">
    <key column="BAR_ID"/>
    <index column="I"/>
    <many-to-many column="FOO_ID" class="Eg.Foo, Eg"/>
</array>

A map from string indices to dates:

<map name="Holidays" table="holidays" schema="dbo" order-by="hol_name asc">
    <key column="id"/>
    <index column="hol_name" type="String"/>
    <element column="hol_date" type="Date"/>
</map>

A list of components (discussed in the next chapter):

<list name="CarComponents" table="car_components">
    <key column="car_id"/>
    <index column="posn"/>
    <composite-element class="Eg.Car.CarComponent">
            <property name="Price" type="float"/>
            <property name="Type" type="Eg.Car.ComponentType, Eg"/>
            <property name="SerialNumber" column="serial_no" type="String"/>
    </composite-element>
</list>

A one to many association links the tables of two classes directly, with no intervening collection table. (This implements a one-to-many relational model.) This relational model loses some of the semantics of .NET collections:

An association from Foo to Bar requires the addition of a key column and possibly an index column to the table of the contained entity class, Bar. These columns are mapped using the <key> and <index> elements described above.

The <one-to-many> tag indicates a one to many association.

<one-to-many
        class="ClassName"                                  (1)
        not-found="ignore|exception"                       (2)
    />

Example:

<set name="Bars">
    <key column="foo_id"/>
    <one-to-many class="Eg.Bar, Eg"/>
</set>

Notice that the <one-to-many> element does not need to declare any columns. Nor is it necessary to specify the table name anywhere.

Very Important Note: If the <key> column of a <one-to-many> association is declared NOT NULL, NHibernate may cause constraint violations when it creates or updates the association. To prevent this problem, you must use a bidirectional association with the many valued end (the set or bag) marked as inverse="true". See the discussion of bidirectional associations later in this chapter.

Collections (other than arrays) may be lazily initialized, meaning they load their state from the database only when the application needs to access it. Initialization happens transparently to the user so the application would not normally need to worry about this (in fact, transparent lazy initialization is the main reason why NHibernate needs its own collection implementations). However, if the application tries something like this:

IDictionary<string, int> permissions;
using (s = sessions.OpenSession())
using (ITransaction tx = sessions.BeginTransaction())
{
    var u = s.Load<User>(userId);
    permissions = u.Permissions;
    tx.Commit();
}

int accessLevel = permissions["accounts"];  // Error!

It could be in for a nasty surprise. Since the permissions collection was not initialized when the ISession was committed, the collection will never be able to load its state. The fix is to move the line that reads from the collection to just before the commit. (There are other more advanced ways to solve this problem, however.)

Alternatively, use a non-lazy collection. Since lazy initialization can lead to bugs like that above, non-laziness is the default. However, it is intended that lazy initialization be used for almost all collections, especially for collections of entities (for reasons of efficiency).

Exceptions that occur while lazily initializing a collection are wrapped in a LazyInitializationException.

Declare a lazy collection using the optional lazy attribute:

<set name="Names" table="NAMES" lazy="true">
    <key column="group_id"/>
    <element column="NAME" type="String"/>
</set>

In some application architectures, particularly where the code that accesses data using NHibernate, and the code that uses it are in different application layers, it can be a problem to ensure that the ISession is open when a collection is initialized. There are two basic ways to deal with this issue:

You can use the CreateFilter() method of the NHibernate ISession API to get the size of a collection without initializing it:

var count = s
    .CreateFilter(collection, "select count(*)")
    .UniqueResult<long>();

CreateFilter() is also used to efficiently retrieve subsets of a collection without needing to initialize the whole collection.

NHibernate supports collections implemented by System.Collections.Generic.SortedList<T> and System.Collections.Generic.SortedSet<T>. You must specify a comparer in the mapping file:

<set name="Aliases" table="person_aliases" sort="natural">
    <key column="person"/>
    <element column="name" type="String"/>
</set>

<map name="Holidays" sort="My.Custom.HolidayComparer, MyAssembly" lazy="true">
    <key column="year_id"/>
    <index column="hol_name" type="String"/>
    <element column="hol_date" type="Date"/>
</map>

Allowed values of the sort attribute are unsorted, natural and the name of a class implementing System.Collections.Generic.IComparer<T>.

If you want the database itself to order the collection elements use the order-by attribute of set, bag or map mappings. This performs the ordering in the SQL query, not in memory.

Setting the order-by attribute tells NHibernate to use Iesi.Collections.Generic.LinkedHashSet class internally for sets, maintaining the order of the elements. It is not supported on maps.

<set name="Aliases" table="person_aliases" order-by="name asc">
    <key column="person"/>
    <element column="name" type="String"/>
</set>

<map name="Holidays" order-by="hol_date, hol_name" lazy="true">
    <key column="year_id"/>
    <index column="hol_name" type="String"/>
    <element column="hol_date type="Date"/>
</map>

Note that the value of the order-by attribute is an SQL ordering, not a HQL ordering!

Associations may even be sorted by some arbitrary criteria at runtime using a CreateFilter().

sortedUsers = s
    .CreateFilter(group.Users, "order by this.Name")
    .List<User>();

If you've fully embraced our view that composite keys are a bad thing and that entities should have synthetic identifiers (surrogate keys), then you might find it a bit odd that the many to many associations and collections of values that we've shown so far all map to tables with composite keys! Now, this point is quite arguable; a pure association table doesn't seem to benefit much from a surrogate key (though a collection of composite values might). Nevertheless, NHibernate provides a feature that allows you to map many to many associations and collections of values to a table with a surrogate key.

The <idbag> element lets you map a List (or Collection) with bag semantics.

<idbag name="Lovers" table="LOVERS" lazy="true">
    <collection-id column="ID" type="Int64">
        <generator class="hilo"/>
    </collection-id>
    <key column="PERSON1"/>
    <many-to-many column="PERSON2" class="Eg.Person" fetch="join"/>
</idbag>

As you can see, an <idbag> has a synthetic id generator, just like an entity class! A different surrogate key is assigned to each collection row. NHibernate does not provide any mechanism to discover the surrogate key value of a particular row, however.

Note that the update performance of an <idbag> is much better than a regular <bag>! NHibernate can locate individual rows efficiently and update or delete them individually, just like a list, map or set.

As of version 2.0, the native identifier generation strategy is supported for <idbag> collection identifiers.

A bidirectional association allows navigation from both "ends" of the association. Two kinds of bidirectional association are supported:

You may specify a bidirectional many-to-many association simply by mapping two many-to-many associations to the same database table and declaring one end as inverse (which one is your choice). Here's an example of a bidirectional many-to-many association from a class back to itself (each category can have many items and each item can be in many categories):

<class name="NHibernate.Auction.Category, NHibernate.Auction">
  <id name="Id" column="ID"/>
  ...
  <bag name="Items" table="CATEGORY_ITEM" lazy="true">
    <key column="CATEGORY_ID"/>
    <many-to-many class="NHibernate.Auction.Item, NHibernate.Auction" column="ITEM_ID"/>
  </bag>
</class>

<class name="NHibernate.Auction.Item, NHibernate.Auction">
  <id name="id" column="ID"/>
  ...

  <!-- inverse end -->
  <bag name="categories" table="CATEGORY_ITEM" inverse="true" lazy="true">
    <key column="ITEM_ID"/>
    <many-to-many class="NHibernate.Auction.Category, NHibernate.Auction"
        column="CATEGORY_ID"/>
  </bag>
</class>

Changes made only to the inverse end of the association are not persisted. This means that NHibernate has two representations in memory for every bidirectional association, one link from A to B and another link from B to A. This is easier to understand if you think about the .NET object model and how we create a many-to-many relationship in C#:

category.Items.Add(item);          // The category now "knows" about the relationship
item.Categories.Add(category);     // The item now "knows" about the relationship

session.Update(item);                     // No effect, nothing will be saved!
session.Update(category);                 // The relationship will be saved

The non-inverse side is used to save the in-memory representation to the database. We would get an unnecessary INSERT/UPDATE and probably even a foreign key violation if both would trigger changes! The same is of course also true for bidirectional one-to-many associations.

You may map a bidirectional one-to-many association by mapping a one-to-many association to the same table column(s) as a many-to-one association and declaring the many-valued end inverse="true".

<class name="Eg.Parent, Eg">
    <id name="Id" column="id"/>
    ....
    <set name="Children" inverse="true" lazy="true">
        <key column="parent_id"/>
        <one-to-many class="Eg.Child, Eg"/>
    </set>
</class>

<class name="Eg.Child, Eg">
    <id name="Id" column="id"/>
    ....
    <many-to-one name="Parent" class="Eg.Parent, Eg" column="parent_id"/>
</class>

Mapping one end of an association with inverse="true" doesn't affect the operation of cascades, both are different concepts!

There are some additional considerations for bidirectional mappings with indexed collections (where one end is represented as a <list> or <map>) when using NHibernate mapping files. If there is a property of the child class that maps to the index column you can use inverse="true" on the collection mapping:

<class name="Parent">
    <id name="Id" column="parent_id"/>
    ....
    <map name="Children" inverse="true">
        <key column="parent_id"/>
        <map-key column="name"
            type="string"/>
        <one-to-many class="Child"/>
    </map>
</class>

<class name="Child">
    <id name="Id" column="child_id"/>
    ....
    <property name="Name" column="name"
        not-null="true"/>
    <many-to-one name="Parent"
        class="Parent"
        column="parent_id"
        not-null="true"/>
</class>

If there is no such property on the child class, the association cannot be considered truly bidirectional. That is, there is information available at one end of the association that is not available at the other end. In this case, you cannot map the collection inverse="true". Instead, you could use the following mapping:

<class name="Parent">
    <id name="Id" column="parent_id"/>
    ....
    <map name="Children">
        <key column="parent_id"
            not-null="true"/>
        <map-key column="name"
            type="string"/>
        <one-to-many class="Child"/>
    </map>
</class>

<class name="Child">
    <id name="Id" column="child_id"/>
    ....
    <many-to-one name="Parent"
        class="Parent"
        column="parent_id"
        insert="false"
        update="false"
        not-null="true"/>
</class>

Note that in this mapping, the collection-valued end of the association is responsible for updates to the foreign key.

There are two possible approaches to mapping a ternary association. One approach is to use composite elements (discussed below). Another is to use an IDictionary with an association as its index:

<map name="Contracts" lazy="true">
    <key column="employer_id"/>
    <index-many-to-many column="employee_id" class="Employee"/>
    <one-to-many class="Contract"/>
</map>

<map name="Connections" lazy="true">
    <key column="node1_id"/>
    <index-many-to-many column="node2_id" class="Node"/>
    <many-to-many column="connection_id" class="Connection"/>
</map>

The <many-to-any> and <index-many-to-any> elements provide for true heterogeneous associations. These mapping elements work in the same way as the <any> element - and should also be used rarely, if ever.

The previous sections are pretty confusing. So lets look at an example. This class:

using System;
using System.Collections.Generic;

namespace Eg

    public class Parent
    {
        public long Id { get; set; }

        private ISet<Child> Children { get; set; }

        ....
        ....
    }
}

has a collection of Eg.Child instances. If each child has at most one parent, the most natural mapping is a one-to-many association:

<hibernate-mapping xmlns="urn:nhibernate-mapping-2.2"
    assembly="Eg" namespace="Eg">

    <class name="Parent">
        <id name="Id">
            <generator class="sequence"/>
        </id>
        <set name="Children" lazy="true">
            <key column="parent_id"/>
            <one-to-many class="Child"/>
        </set>
    </class>

    <class name="Child">
        <id name="Id">
            <generator class="sequence"/>
        </id>
        <property name="Name"/>
    </class>

</hibernate-mapping>

This maps to the following table definitions:

create table parent (Id bigint not null primary key)
create table child (Id bigint not null primary key, Name varchar(255), parent_id bigint)
alter table child add constraint childfk0 (parent_id) references parent

If the parent is required, use a bidirectional one-to-many association:

<hibernate-mapping xmlns="urn:nhibernate-mapping-2.2"
    assembly="Eg" namespace="Eg">

    <class name="Parent">
        <id name="Id">
            <generator class="sequence"/>
        </id>
        <set name="Children" inverse="true" lazy="true">
            <key column="parent_id"/>
            <one-to-many class="Child"/>
        </set>
    </class>

    <class name="Child">
        <id name="Id">
            <generator class="sequence"/>
        </id>
        <property name="Name"/>
        <many-to-one name="parent" class="Parent" column="parent_id" not-null="true"/>
    </class>

</hibernate-mapping>

Notice the NOT NULL constraint:

create table parent ( Id bigint not null primary key )
create table child ( Id bigint not null
                     primary key,
                     Name varchar(255),
                     parent_id bigint not null )
alter table child add constraint childfk0 (parent_id) references parent

On the other hand, if a child might have multiple parents, a many-to-many association is appropriate:

<hibernate-mapping xmlns="urn:nhibernate-mapping-2.2"
    assembly="Eg" namespace="Eg">

    <class name="Parent">
        <id name="Id">
            <generator class="sequence"/>
        </id>
        <set name="Children" lazy="true" table="childset">
            <key column="parent_id"/>
            <many-to-many class="Child" column="child_id"/>
        </set>
    </class>

    <class name="eg.Child">
        <id name="Id">
            <generator class="sequence"/>
        </id>
        <property name="Name"/>
    </class>

</hibernate-mapping>

Table definitions:

create table parent ( Id bigint not null primary key )
create table child ( Id bigint not null primary key, name varchar(255) )
create table childset ( parent_id bigint not null,
                        child_id bigint not null,
                        primary key ( parent_id, child_id ) )
alter table childset add constraint childsetfk0 (parent_id) references parent
alter table childset add constraint childsetfk1 (child_id) references child

See also Chapter 22, Example: Parent/Child.