So, the requirements change and you now need to record a new piece of information in one of your tables. That’s a pretty easy one, right? Just add a column and you’re laughing.
Well this is probably the way I see this problem solved 100% of the time (Rails migrations and Grails GORM for example) and I actually think it is the worst solution.
When you add a column to a table you should always consider whether you can add a sensible value. Adding a nullable column to a table is a really bad idea. Firstly, it isn’t relational (and it is almost guaranteed to reduce your normal form). Secondly, it is likely to introduce performance problems if the column needs to be searchable. Finally, nullable columns have a higher cost of maintenance as they imply logic (the null state) but don’t define it. The logic as to when the column may be null will always exist outside the schema.
If you can define a sensible default then you might be able to add the column to the table. For example boolean flag fields could be handled this way as long as the historic values can be accurately mapped to, say, false.
This method however falls down if your historical data cannot be truthfully mapped to a default value. If for example the new value represents something you have started to track rather than some brand new piece of information then you cannot truthfully map the historic data to the default value as you don’t know whether the default really applies or not.
For example if you send goods to customers using a regular or overnight delivery method and you need to start tracking the delivery method it would be wrong to map all historic data to the regular delivery. The truth is you don’t know whether a historical order was an overnight delivery or not. If you do map it incorrectly then you will end up skewing all your data and ultimately poison the very data you want to be tracking.
In this case it is far better to simply introduce a relational child table to hold the new information. Using child tables is a more accurate record of the data as a child row will only exist where the data is known. For unknown historic records there will be no child entry and your queries will not need any special cases.
When using child table data like this you can easily separate your data via the EXISTS predicate and in most databases EXISTS is very performant.
I think using child tables to record new information is relational way to solve the problem of adding new columns but it is usually turned down for one of two reasons (there naturally may be others but I’ve only ever heard these used in anger).
Firstly there is the argument that this technique leads to a proliferation of child tables. This is an accurate criticism but unfortunately if you want your database to be accurate and the new information is introduced piecemeal then you do not have a lot of choice. Pretending that you have historic data you don’t actually have doesn’t solve the problem.
One thing that can help this situation is to separate your database into its transaction processing element and its warehousing or archive aspect. In the warehouse the structure of the database might be quite involved but since the accuracy of the data is more important than the simplicity of the data model and the queries tend to be of a reporting or aggregating nature there tends to be less issues with having many tables. Often views can allow you to present a high-level view of the data and introducing null values into the heading of query, while ugly, is more acceptable than breaking normal form in the raw data itself (though god help you if you start joining aggregate views on the gappy data rather than the underlying true state of the data).
The transactional version of the database is then free to reflect the state of the current data model alone rather than the historical permutations. Here the model can be clean and reflect more of the Domain Model although again you want to retain the Relational aspect of the data.
Having separate datastores with different schemas often makes app developers lives easier as well. Usually an application is only concerned with current interactions and not historical transactions. As long as historical data can be obtained as a Service then you have no need to reflect historical schema concerns in your application.
However even if you cannot separate the data there is still little reason to introduce null value columns into the database. ORM often makes the simple assumption that one data entity is one object, but that is simply to make ORM easier. The application persistence layer shouldn’t be allowed to distort the underlying relational data model. If your data isn’t relational or you can’t define a fixed schema, don’t use an RDBMS. Look into things like CouchDB.
In fact some persistence schemes, like JPA, actually allow you to reflect the reality of the relational data by allowing embedded entities in objects. This is a leaky abstraction because from an object point of view there often is very little reason why, say, an Order object contains a Delivery Dispatch Type object.
If you want to avoid the leak then use the Repository pattern to create a true Domain object and be prepared to handle the complexity of transforming the relational data to the object-orientated data within the Repository. Remember that the Repository represents a true boundary in the system: you are hoping to derive real benefits from both the relational representation and the object representation, not screw one in favour of the other.
So please; don’t just wack in another nullable column and then wonder why queries on your database are slow.