When I first worked with Sybase/SQL Server, we thought our databases were impressively large but they were, by today’s standards, pathetically small. We had one script to build the whole database. Every script I ever read was richly annotated; it was more like reading a document. Every table had a comment block, and every line would be commented too. At the end of each routine (e.g. procedure) was a quick integration test, or series of test assertions, to check that nothing in the build was broken. We simply ran the build script, stored in the Version Control System, and it pulled everything together in a logical sequence that not only created the database objects but pulled in the static data.
This worked fine at the scale we had. The advantage was that one could, by reading the source code, reach a rapid understanding of how the database worked and how one could interface with it. The problem was that it was a system that meant that only one developer at the time could work on the database. It was very easy for a developer to execute accidentally the entire build script rather than the selected section on which he or she was working, thereby cleansing the database of everyone else’s work-in-progress and data.
It soon became the fashion to work at the object level, so that programmers could check out individual views, tables, functions, constraints and rules and work on them independently. It was then that I noticed the trend to generate the source for the VCS retrospectively from the development server. Tables were worst affected. You can, of course, add or delete a table’s columns and constraints retrospectively, which means that the existing source no longer represents the current object. If, after your development work, you generate the source from the live table, then you get no block or line comments, and the source script is sprinkled with silly square-brackets and other confetti, thereby rendering it visually indigestible. Routines, too, were affected. In our system, every routine had a directly attached string of unit-tests. A retro-generated routine has no unit-tests or test assertions. Yes, one can still commit our test code to the VCS but it’s a separate module and teams end up running the whole suite of tests for every individual change, rather than just the tests for that routine, which doesn’t scale for database testing.
With Extended properties, one can get the best of both worlds, and even use them to put blame, praise or annotations into your VCS. It requires a lot of work, though, particularly the script to generate the table. The problem is that there are no conventional names beyond ‘MS_Description’ for the special use of extended properties. This makes it difficult to do splendid things such ensuring the integrity of the build by running a suite of tests that are actually stored in extended properties within the database and therefore the VCS.
We have lost the readability of database source code over the years, and largely jettisoned the use of test assertions as part of the database build. This is not unexpected in view of the increasing complexity of the structure of databases and number of programmers working on them. There must, surely, be a way of getting them back, but I sometimes wonder if I’m one of very few who miss them.