Use an Address class to encapsulate street, suburb, state, postcode. This encourages code reuse and simplifies refactoring.
NHibernate makes identifier properties optional. There are all sorts of reasons why you should use them. We recommend that identifiers be 'synthetic' (generated, with no business meaning) and of a non-primitive type. For maximum flexibility, use Int64 or String.
Don't use a single monolithic mapping document. Map Eg.Foo in the file Eg/Foo.hbm.xml. This makes particularly good sense in a team environment.
Place mapping files along with the classes they map and declare them as Embedded Resources in Visual Studio.
This is a good practice if your queries call non-ANSI-standard SQL functions. Externalising the query strings to mapping files will make the application more portable.
As in ADO.NET, always replace non-constant values by "?". Never use string manipulation to bind a non-constant value in a query! Even better, consider using named parameters in queries.
NHibernate lets the application manage ADO.NET connections. This approach should be considered a last-resort. If you can't use the built-in connections providers, consider providing your own implementation of NHibernate.Connection.IConnectionProvider.
Suppose you have a type, say from some library, that needs to be persisted but doesn't provide the accessors needed to map it as a component. You should consider implementing NHibernate.UserTypes.IUserType. This approach frees the application code from implementing transformations to / from an NHibernate type.
In performance-critical areas of the system, some kinds of operations (eg. mass update / delete) might benefit from direct ADO.NET. But please, wait until you know something is a bottleneck. And don't assume that direct ADO.NET is necessarily faster. If need to use direct ADO.NET, it might be worth opening a NHibernate ISession and using that SQL connection. That way you can still use the same transaction strategy and underlying connection provider.
From time to time the ISession synchronizes its persistent state with the database. Performance will be affected if this process occurs too often. You may sometimes minimize unnecessary flushing by disabling automatic flushing or even by changing the order of queries and other operations within a particular transaction.
When using a distributed architecture, you could pass persistent objects loaded in the middle tier to and from the user interface tier. Use a new session to service each request. Use ISession.Update() or ISession.SaveOrUpdate() to update the persistent state of an object.
Database Transactions have to be as short as possible for best scalability. However, it is often neccessary to implement long running Application Transactions, a single unit-of-work from the point of view of a user. This Application Transaction might span several client requests and response cycles. Either use Detached Objects or, in two tiered architectures, simply disconnect the NHibernate Session from the ADO.NET connection and reconnect it for each subsequent request. Never use a single Session for more than one Application Transaction usecase, otherwise, you will run into stale data.
This is more of a necessary practice than a "best" practice. When an exception occurs, roll back the ITransaction and close the ISession. If you don't, NHibernate can't guarantee that in-memory state accurately represents persistent state. As a special case of this, do not use ISession.Load() to determine if an instance with the given identifier exists on the database; use Get() or a query instead.
Use eager (outer-join) fetching sparingly. Use proxies and/or lazy collections for most associations to classes that are not cached in the second-level cache. For associations to cached classes, where there is a high probability of a cache hit, explicitly disable eager fetching using fetch="select". When an outer-join fetch is appropriate to a particular use case, use a query with a left join fetch.
Hide (NHibernate) data-access code behind an interface. Combine the DAO and Thread Local Session patterns. You can even have some classes persisted by handcoded ADO.NET, associated to NHibernate via an IUserType. (This advice is intended for "sufficiently large" applications; it is not appropriate for an application with five tables!)
If you compare objects outside of the ISession scope, you have to implement Equals() and GetHashCode(). Inside the ISession scope, object identity is guaranteed. If you implement these methods, never ever use the database identifier! A transient object doesn't have an identifier value and NHibernate would assign a value when the object is saved. If the object is in an ISet<T> while being saved, the hash code changes, breaking the contract. To implement Equals() and GetHashCode(), use a unique business key, that is, compare a unique combination of class properties. Remember that this key has to be stable and unique only while the object is in an ISet<T>, not for the whole lifetime (not as stable as a database primary key). Never use collections in the Equals() comparison (lazy loading) and be careful with other associated classes that might be proxied.
Good usecases for a real many-to-many associations are rare. Most of the time you need additional information stored in the "link table". In this case, it is much better to use two one-to-many associations to an intermediate link class. In fact, we think that most associations are one-to-many and many-to-one, you should be careful when using any other association style and ask yourself if it is really neccessary.