In a recent post I talked about how good design attempts to minimize the impact of changes to a system, often through Design Patterns.
When used appropriately, Design Patterns are a great tool for building a great design, but there is an important caveat to keep in mind anytime you apply a pattern. A Design Pattern might minimize the impact of one kind of change at the expense of amplifying another type of change.
What do I mean by this? One common pattern is the Abstract Factory pattern which is often manifested in .NET code via the Provider Model pattern. The Provider Model abstracts access to an underlying resource by providing a fixed API to the resource. This does a bang up job of insulating the consumer of the provider when changing the underlying resource.
MembershipProvider is one such implementation of the provider model pattern. The consumer of the
MembershipProvider API doesn’t need to change if the
SqlMembershipProvider is swapped in favor of the
ActiveDirectoryMembershipProvider. This is one way that the provider pattern attempts to minimize the impact of changes. It insulates against changes to the underlying data store.
However there is a hidden tradeoff with this pattern. Suppose the API itself changes often. Then, the impact of a single API change is multiplied across every concrete implementation of the provider. In the case of the
MembershipProvider, this is pretty much a non-issue because the likelihood of changing the API is very small.
But the same cannot be said of the data access layer for software such as a blog (or similar software). A common approach is to implement a
BlogDataProvider to encapsulate all data access so that the blog software can make use of multiple databases. The basic line of thought is that we can simply implement a concrete provider for each database we wish to support. So we might implement a
FireBirdBlogDataProvider, and so on.
This sounds great in theory, but it breaks down in practice because unlike the API to the
MembershipProvider, the API for a
BlogDatabaseProvider is going to change quite often. Pretty much every new feature one can think of often needs a backing data store.
Everytime we add a new column to a table to support a feature, the impact of that change is multiplied by the number of providers we support. I discussed this in the past in my post entitled Where the Provider Model Falls Short.
Every Design Pattern comes with inherent tradeoffs that we must be aware of. There is no silver bullet.
The key here when looking to apply patterns is to not follow a script for applying patterns blindly. Look at what changes often (in this case the database schema) and figure out how to minimize the impact of that change. In the above scenario, one option is to simply punt the work of supporting multiple databases to someone else in a more generic fashion.
For example, using something like NHibernate or Subsonic in this situation might mean that a schema change only requires changing one bit of code. Then NHibernate or Subsonic is responsible for making sure that the code works against its list of supported databases.
One might object to these approaches because they feel these approaches cannot possibly query every database they support as efficiently as database specific SQL as one would do in a database specific provider. But I think the 80/20 rule applies here. Let the dynamic query engine get you 80% of the way, and use a provider just for the areas that need it.
So again, this is not an indictment of the provider model. The provider model is extremely useful when used appropriately. As I mentioned, the Membership Provider is a great example. But if you really need to support multiple databases AND your database schema is succeptible to a lot of churn, then another pattern may be in order.