The Singleton pattern solves many of your problems. You know that you only need a single instance. You have a guarantee that this instance is initialized before it's used. It keeps your design simple by having a global access point. It's all good. What's not to like about this classic design pattern?
Quite a lot, it turns out. Tempting they may be, but experience shows that most singletons really do more harm than good. They hinder testability and harm maintainability. Unfortunately, this additional wisdom is not as widespread as it should be and singletons continue to be irresistible to many programmers. But it is worth resisting:
-
The single-instance requirement is often imagined. In many cases it's pure speculation that no additional instances will be needed in the future. Broadcasting such speculative properties across an application's design is bound to cause pain at some point. Requirements will change. Good design embraces this. Singletons don't.
-
Singletons cause implicit dependencies between conceptually independent units of code. This is problematic both because they are hidden and because they introduce unnecessary coupling between units. This code smell becomes pungent when you try to write unit tests, which depend on loose coupling and the ability to selectively substitute a mock implementation for a real one. Singletons prevent such straightforward mocking.
-
Singletons also carry implicit persistent state, which again hinders unit testing. Unit testing depends on tests being independent of one another, so the tests can be run in any order and the program can be set to a known state before the execution of every unit test. Once you have introduced singletons with mutable state, this may be hard to achieve. In addition, such globally accessible persistent state makes it harder to reason about the code, especially in a multi-threaded environment.
-
Multi-threading introduces further pitfalls to the singleton pattern. As straightforward locking on access is not very efficient, the so-called double-checked locking pattern (DCLP) has gained in popularity. Unfortunately, this may be a further form of fatal attraction. It turns out that in many languages DCLP is not thread-safe and, even where it is, there are still opportunities to get it subtly wrong.
The cleanup of singletons may present a final challenge:
-
There is no support for explicitly killing singletons, which can be a serious issue in some contexts. For example, in a plug-in architecture where a plug-in can only be safely unloaded after all its objects have been cleaned up.
-
There is no order to the implicit cleanup of singletons at program exit. This can be troublesome for applications that contain singletons with interdependencies. When shutting down such applications, one singleton may access another that has already been destroyed.
-
Some of these shortcomings can be overcome by introducing additional mechanisms. However, this comes at the cost of additional complexity in code that could have been avoided by choosing an alternative design.
Therefore, restrict your use of the Singleton pattern to the classes that truly must never be instantiated more than once. Don't use a singleton's global access point from arbitrary code. Instead, direct access to the singleton should be from only a few well-defined places, from where it can be passed around via its interface to other code. This other code is unaware, and so does not depend on whether a singleton or any other kind of class implements the interface. This breaks the dependencies that prevented unit testing and improves the maintainability. So, next time you are thinking about implementing or accessing a singleton, hopefully you'll pause, and think again.
by Sam Saariste