这是一个“圣战”式的问题。换句话说，两种选择都是好的，但人们会有自己的偏好，他们会誓死捍卫这些偏好。

公认的做法是使用IllegalArgumentException(字符串消息)来声明一个参数是无效的，并给出尽可能多的细节…所以说，一个参数被发现是空的，而异常是非空的，你会这样做:

if( variable == null )
    throw new IllegalArgumentException("The object 'variable' cannot be null");

你几乎没有理由隐式地使用“NullPointerException”。NullPointerException是当您试图在空引用(如toString())上执行代码时由Java虚拟机抛出的异常。

我完全同意你说的话。早失败，快失败。非常好的异常咒语。

抛出哪个Exception主要是个人喜好的问题。在我看来，IllegalArgumentException似乎比使用NPE更具体，因为它告诉我问题是我传递给方法的参数，而不是执行方法时可能生成的值。

我的2美分

抛出一个排除空参数的异常(无论是NullPointerException还是自定义类型)使得自动化空测试更加可靠。这种自动化测试可以通过反射和一组默认值来完成，就像在Guava的NullPointerTester中一样。例如，NullPointerTester将尝试调用以下方法…

Foo(String string, List<?> list) {
  checkArgument(string.length() > 0);
  // missing null check for list!
  this.string = string;
  this.list = list;
}

...with two lists of arguments: "", null and null, ImmutableList.of(). It would test that each of these calls throws the expected NullPointerException. For this implementation, passing a null list does not produce NullPointerException. It does, however, happen to produce an IllegalArgumentException because NullPointerTester happens to use a default string of "". If NullPointerTester expects only NullPointerException for null values, it catches the bug. If it expects IllegalArgumentException, it misses it.

在这种情况下，IllegalArgumentException使用API向用户传达了“不应该为空”的明确信息。正如其他论坛用户指出的那样，只要你使用API向用户传达正确的信息，你就可以使用NPE。

GaryF和tweak放弃了“有效Java”(我发誓)的参考，建议使用NPE。看看其他好的API是如何构造的，这是了解如何构造你的API的最好方法。

Another good example is to look at the Spring APIs. For example, org.springframework.beans.BeanUtils.instantiateClass(Constructor ctor, Object[] args) has a Assert.notNull(ctor, "Constructor must not be null") line. org.springframework.util.Assert.notNull(Object object, String message) method checks to see if the argument (object) passed in is null and if it is it throws a new IllegalArgumentException(message) which is then caught in the org.springframework.beans.BeanUtils.instantiateClass(...) method.

实际上，在我看来，抛出IllegalArgumentException或NullPointerException的问题只是对Java中不完全理解异常处理的少数人的“圣战”。一般来说，规则很简单，如下:

argument constraint violations must be indicated as fast as possible (-> fast fail), in order to avoid illegal states which are much harder to debug in case of an invalid null pointer for whatever reason, throw NullPointerException in case of an illegal array/collection index, throw ArrayIndexOutOfBounds in case of a negative array/collection size, throw NegativeArraySizeException in case of an illegal argument that is not covered by the above, and for which you don't have another more specific exception type, throw IllegalArgumentException as a wastebasket on the other hand, in case of a constraint violation WITHIN A FIELD that could not be avoided by fast fail for some valid reason, catch and rethrow as IllegalStateException or a more specific checked exception. Never let pass the original NullPointerException, ArrayIndexOutOfBounds, etc in this case!

至少有三个非常好的理由反对将所有类型的参数约束违反映射到IllegalArgumentException，第三个理由可能非常严重，以至于标志着这种做法的糟糕风格:

(1) A programmer cannot a safely assume that all cases of argument constraint violations result in IllegalArgumentException, because the large majority of standard classes use this exception rather as a wastebasket if there is no more specific kind of exception available. Trying to map all cases of argument constraint violations to IllegalArgumentException in your API only leads to programmer frustration using your classes, as the standard libraries mostly follow different rules that violate yours, and most of your API users will use them as well!

(2) Mapping the exceptions actually results in a different kind of anomaly, caused by single inheritance: All Java exceptions are classes, and therefore support single inheritance only. Therefore, there is no way to create an exception that is truly say both a NullPointerException and an IllegalArgumentException, as subclasses can only inherit from one or the other. Throwing an IllegalArgumentException in case of a null argument therefore makes it harder for API users to distinguish between problems whenever a program tries to programmatically correct the problem, for example by feeding default values into a call repeat!

(3) Mapping actually creates the danger of bug masking: In order to map argument constraint violations into IllegalArgumentException, you'll need to code an outer try-catch within every method that has any constrained arguments. However, simply catching RuntimeException in this catch block is out of the question, because that risks mapping documented RuntimeExceptions thrown by libery methods used within yours into IllegalArgumentException, even if they are no caused by argument constraint violations. So you need to be very specific, but even that effort doesn't protect you from the case that you accidentally map an undocumented runtime exception of another API (i.e. a bug) into an IllegalArgumentException of your API. Even the most careful mapping therefore risks masking programming errors of other library makers as argument constraint violations of your method's users, which is simply hillareous behavior!

With the standard practice on the other hand, the rules stay simple, and exception causes stay unmasked and specific. For the method caller, the rules are easy as well: - if you encounter a documented runtime exception of any kind because you passed an illegal value, either repeat the call with a default (for this specific exceptions are neccessary), or correct your code - if on the other hand you enccounter a runtime exception that is not documented to happen for a given set of arguments, file a bug report to the method's makers to ensure that either their code or their documentation is fixed.