Friday, February 5, 2010

Consider composition over inheritance

This week I want to discuss another bullet from my code review checklist.

Composition is a method we use to combine simple objects into more complex objects.  With composition we have an owner class that changes behavior by delegating the implementation of certain behaviors to smaller and simpler objects.  When we destroy the owner object, all of these smaller objects the owner uses are also destroyed.  If we were to use inheritance to achieve this same goal (changing behavior) we would use a subclass to change the behavior of the base class.  As I mentioned in my last post,  the base class and subclass are tightly coupled and this can tend to be a bit brittle.  

Consider the very real case where we are writing an application that needs to work with both Oracle and SQL Server.  We could use inheritance to define a base class and have one subclass for SQL Server and another for Oracle.  Our application can figure out which object to instantiate at runtime and that will work pretty well. Here is what our class diagram might look like:

image 

If we take a little different approach and use composition instead we might have a class diagram similar to the following.

image

The CompositionDatabaseWriter does not need to implement IDatabaseWriter but I did that because I think it makes the example easier to understand.  The application will use the CompositionDatabaseWriter to do database work and CompositionDatabaseWriter will determine whether to use SqlServerDatabaseWriter or OracleDatabaseWriter at runtime perhaps by using a configuration file entry.  When one of the CompositionDatabaseWriter methods is called, CompositionDatabaseWriter simply calls the corresponding method on the Sql Server or Oracle object. 

Both designs allow us to interact with an oracle or Sql Server database which was our goal.  Now here is where the flexibility of composition comes in;  Suppose we now need to support a MySql database.  If we use inheritance to solve this problem, at a minimum we must create a MySqlDatabaseWriter subclass, recompile, test, and deploy our data layer.  If we use composition instead we can create a MySqlDatabaseWriter in a new assembly and all we have to do to use it is deploy this single assembly and change a config file.  Very powerful, because we used composition we can change the behavior at runtime via configuration.

The fact of the mater is that our sample is so simple we didn’t even use composition at all so imagine the considerably more complex scenario where what we are doing is creating a persistence object that must transactionally work with the file system and a database.  We define the persistence object and within the persistence object we delegate the responsibility of dealing with the file system to one class, and dealing with the database to another.  The persistence object is composed of a FileSytemWriter and a DatabaseWriter.  Our Persistence object has only one concern – coordinating database and file persistence.  Our FileSystemWriter has only one concern – managing File System interactions, and finally our DatabaseWriter is only concerned about interacting with a database of one type.  We have decomposed a fairly complex problem into a few smaller more managable problems, we came up with a nice robust design thanks to composition and we have managed to do it in such a way that there are no SRP violations either!

No comments:

Post a Comment