It took us six weeks to break the curse of the singleton. Six weeks! By the end of it, we’d rewritten most of our data access layer.
We began the process of removing singletons innocently enough. I thought I was well prepared for the task. I’d just finished reading The Pragmatic Programmer (my review of The Pragmatic Programmer) and Working Effectively with Legacy Code (my review of Legacy Code). I remember telling Gordon I’d tackle the problem over the weekend…
The Singleton Design Pattern is one of the first patterns introduced in many software design books. But don’t let this fool you like it did me. Its prominent position has nothing to do with its importance. The Singleton is usually listed first because it’s the easiest pattern to explain and implement. It made a convenient place for the author to start, but the Singleton’s real uses are very limited.
Which is appropriate, actually, since the real use of the singleton is to limit usage. A class that implements the Singleton pattern allows only one object to be instantiated at a time. There are a few cases where this is desirable. For example, classes that control access to a single hardware device or that set up global variables. But the danger of the Singleton is that there are many cases where you’ll want to misuse it.
Scott Densmore lists the four key characteristics of the Singleton and how each can get you into trouble in his Why Singletons are Evil blog post.
For another cautionary tale of the cycle attraction, infatuation, disappointment, and rejection, read Singleton, I love you, but you’re bringing me down.
In our case, we’d gleefully implemented Singletons for database access, content indexing, security and access control, and in a few other places where we thought we needed just one instance. If Steve Yegge were here, he’d call what we’d done an instance of the Simpleton pattern — a failure to grasp basic principles of object-oriented programming. You can read more about Yegge’s thoughts on the singleton and design patterns for dummies.
Our automated tests were running slowly because we had to set up and tear down the database for every test. Making a change to one component would frequently cause several tests to fail. Everything was tied together at the hip — at the Singleton classes — and it was impossible to disentangle our code to test particular items in isolation. We had tests, but not unit tests. They were integration tests, and the points of integration were the handful of singleton classes we’d built.
Worse, our database performance was lousy. Since we had a global variable for our database object, we could sprinkle database access code throughout the rest of our object model. We discovered that we were opening and closing database connections all the time. And we’d had to implement tricky locking code to guarantee that our SQL statements would get executed in the right order.
The Singleton let us be lazy about our programming habits. It allowed us to make assumptions we shouldn’t have. You can call it premature optimization or a retreat into procedural programming techniques from an earlier era. Ultimately, we’d found that it allowed us to cut too many corners.
So we slowly rooted out each Singleton class from our API and reimplemented the functionality in other ways. Fortunately, we had a large battery of integration tests to help guide us. And luckily, we’d decided to tackle the problem during our first Alpha test, when we could still afford to make sweeping changes. But correcting bad design takes much longer than avoiding it in the first place — even if you’ve read all the right books.
Six weeks later, we finally sorted out the mess we’d made for ourselves. There’s a handful of odds and ends left to do, but the design feels better. My gut tells me it’s an improvement, and our tests — now we have both unit and integration tests — show that we’ve almost tripled the speed of the data access layer.
It was worth our time to break the Curse of the Singleton. Beware lest ye, too, fall under its spell!