Good or bad practice? Initializing objects in getter

Question

I have a strange habit it seems... according to my co-worker at least. We've been working on a small project together. The way I wrote the classes is (simplified example):

[Serializable()]
public class Foo
{
    public Foo()
    { }

    private Bar _bar;

    public Bar Bar
    {
        get
        {
            if (_bar == null)
                _bar = new Bar();

            return _bar;
        }
        set { _bar = value; }
    }
}

So, basically, I only initialize any field when a getter is called and the field is still null. I figured this would reduce overload by not initializing any properties that aren't used anywhere.

ETA: The reason I did this is that my class has several properties that return an instance of another class, which in turn also have properties with yet more classes, and so on. Calling the constructor for the top class would subsequently call all constructors for all these classes, when they are not always all needed.

Are there any objections against this practice, other than personal preference?

UPDATE: I have considered the many differing opinions in regards to this question and I will stand by my accepted answer. However, I have now come to a much better understanding of the concept and I'm able to decide when to use it and when not.

Cons:

• Thread safety issues
• Not obeying a "setter" request when the value passed is null
• Micro-optimizations
• Exception handling should take place in a constructor
• Need to check for null in class' code

Pros:

• Micro-optimizations
• Properties never return null
• Delay or avoid loading "heavy" objects

Most of the cons are not applicable to my current library, however I would have to test to see if the "micro-optimizations" are actually optimizing anything at all.

LAST UPDATE:

Okay, I changed my answer. My original question was whether or not this is a good habit. And I'm now convinced that it's not. Maybe I will still use it in some parts of my current code, but not unconditionally and definitely not all the time. So I'm going to lose my habit and think about it before using it. Thanks everyone!

Answer 1

What you have here is a - naive - implementation of "lazy initialization".

Short answer:

Using lazy initialization unconditionally is not a good idea. It has its places but one has to take into consideration the impacts this solution has.

Background and explanation:

Concrete implementation:
Let's first look at your concrete sample and why I consider its implementation naive:

1. It violates the Principle of Least Surprise (POLS). When a value is assigned to a property, it is expected that this value is returned. In your implementation this is not the case for null:

   foo.Bar = null;
   Assert.Null(foo.Bar); // This will fail
   

2. It introduces quite some threading issues: Two callers of foo.Bar on different threads can potentially get two different instances of Bar and one of them will be without a connection to the Foo instance. Any changes made to that Bar instance are silently lost.
   This is another case of a violation of POLS. When only the stored value of a property is accessed it is expected to be thread-safe. While you could argue that the class simply isn't thread-safe - including the getter of your property - you would have to document this properly as that's not the normal case. Furthermore the introduction of this issue is unnecessary as we will see shortly.

In general:
It's now time to look at lazy initialization in general:
Lazy initialization is usually used to delay the construction of objects that take a long time to be constructed or that take a lot of memory once fully constructed.
That is a very valid reason for using lazy initialization.

However, such properties normally don't have setters, which gets rid of the first issue pointed out above.
Furthermore, a thread-safe implementation would be used - like Lazy<T> - to avoid the second issue.

Even when considering these two points in the implementation of a lazy property, the following points are general problems of this pattern:

1. Construction of the object could be unsuccessful, resulting in an exception from a property getter. This is yet another violation of POLS and therefore should be avoided. Even the section on properties in the "Design Guidelines for Developing Class Libraries" explicitly states that property getters shouldn't throw exceptions:

   Avoid throwing exceptions from property getters.

   Property getters should be simple operations without any preconditions. If a getter might throw an exception, consider redesigning the property to be a method.

2. Automatic optimizations by the compiler are hurt, namely inlining and branch prediction. Please see Bill K's answer for a detailed explanation.

The conclusion of these points is the following:
For each single property that is implemented lazily, you should have considered these points.
That means, that it is a per-case decision and can't be taken as a general best practice.

This pattern has its place, but it is not a general best practice when implementing classes. It should not be used unconditionally, because of the reasons stated above.

---

In this section I want to discuss some of the points others have brought forward as arguments for using lazy initialization unconditionally:

1. Serialization:
   EricJ states in one comment:

   An object that may be serialized will not have it's contructor invoked when it is deserialized (depends on the serializer, but many common ones behave like this). Putting initialization code in the constructor means that you have to provide additional support for deserialization. This pattern avoids that special coding.

   There are several problems with this argument:

   1. Most objects never will be serialized. Adding some sort of support for it when it is not needed violates YAGNI.
   2. When a class needs to support serialization there exist ways to enable it without a workaround that doesn't have anything to do with serialization at first glance.

2. Micro-optimization: Your main argument is that you want to construct the objects only when someone actually accesses them. So you are actually talking about optimizing the memory usage.
   I don't agree with this argument for the following reasons:

   1. In most cases, a few more objects in memory have no impact whatsoever on anything. Modern computers have way enough memory. Without a case of actual problems confirmed by a profiler, this is pre-mature optimization and there are good reasons against it.

   2. I acknowledge the fact that sometimes this kind of optimization is justified. But even in these cases lazy initialization doesn't seem to be the correct solution. There are two reasons speaking against it:

      1. Lazy initialization potentially hurts performance. Maybe only marginally, but as Bill's answer showed, the impact is greater than one might think at first glance. So this approach basically trades performance versus memory.
      2. If you have a design where it is a common use case to use only parts of the class, this hints at a problem with the design itself: The class in question most likely has more than one responsibility. The solution would be to split the class into several more focused classes.