java singleton instantiation

Question

I've found three ways of instantiating a Singleton, but I have doubts as to whether any of them is the best there is. I'm using them in a multi-threaded environment and prefer lazy instantiation.
Sample 1:

private static final ClassName INSTANCE = new ClassName();

public static ClassName getInstance() {
    return INSTANCE;
}

Sample 2:

private static class SingletonHolder { 
    public static final ClassName INSTANCE = new ClassName();
}

public static ClassName getInstance() {
    return SingletonHolder.INSTANCE;
}

Sample 3:

private static ClassName INSTANCE;

public static synchronized ClassName getInstance()
{
    if (INSTANCE == null)
        INSTANCE = new ClassName();

    return INSTANCE;
}

The project I'm using ATM uses Sample 2 everywhere, but I kind of like Sample 3 more. There is also the Enum version, but I just don't get it.

The question here is - in which cases I should/shouldn't use any of these variations? I'm not looking for lengthy explanations though (there's plenty of other topics about that, but they all eventually turn into arguing IMO), I'd like it to be understandable with few words.

Answer 1

Sample 1 does not use lazy initialisation.

Sample 2 and 3 are both lazy. Sample 2 uses the Initialization on demand holder idiom (IODH) which has no synchronisation overhead. Therefore it is faster than Sample 3.

In Effective Java (Item 3), Joshua Bloch recommends that a single-element enum type is the best way to implement a singleton.

However, if you are unsure about the enum type, stick with IODH.

Answer 2

First of all, make absolutely sure that you need a singleton, and that you want to provide "global-level" access to the singleton. I've found that in many cases clients of the singleton have no need to know that it is a singleton. Rather, they just need to get a service when they are instantiated.

Thus, regardless of how you obtain the singleton (if at all), consider changing the way your classes gain access to this object. While this means modifying constructors and changing "distribution changes", I've found that dependency injection frameworks reduce the cost. The DI framework can also then take care of singleton instantiation (e.g., Guice does that).

Aside from that. option 3 is the typical and most common (and thread safe) version that I'm familiar with. Option 1 is mostly for non-lazy initialization (not always acceptable). I've never seen 2 used.