这不仅仅是从异常中恢复的能力问题。在我看来，最重要的是调用者是否对捕获异常感兴趣。

如果您编写的库用于其他地方，或应用程序中的较低级别层，请询问自己调用者是否有兴趣捕获(了解)您的异常。如果他不是，那么使用一个未检查的异常，这样就不会给他带来不必要的负担。

这是许多框架使用的哲学。尤其是Spring和hibernate——它们精确地将已知的受控异常转换为未检查异常，因为Java中过度使用了受控异常。我能想到的一个例子是来自json.org的JSONException，这是一个受检查的异常，最烦人的是——它应该是不受检查的，但开发人员根本没有考虑到这一点。

顺便说一下，大多数情况下，调用者对异常的兴趣与从异常中恢复的能力直接相关，但情况并非总是如此。

在任何一个足够大的系统上，有很多层，检查异常是无用的，因为无论如何，您需要一个架构级策略来处理异常将如何处理(使用故障屏障)。

使用受控异常，您的错误处理策略是微管理的，在任何大型系统上都无法承受。

大多数情况下，您不知道错误是否“可恢复”，因为您不知道API的调用者位于哪一层。

假设我创建了一个StringToInt API，用于将整数的字符串表示形式转换为Int。如果API是用“foo”字符串调用的，我必须抛出检查异常吗?它可以恢复吗?我不知道，因为在他的层中，我的StringToInt API的调用者可能已经验证了输入，如果抛出这个异常，它要么是一个错误，要么是一个数据损坏，它是不可恢复的这一层。

在这种情况下，API的调用者不想捕获异常。他只想让异常“冒出来”。如果我选择了一个受控异常，这个调用者将有大量无用的catch块，只能人为地重新抛出异常。

大多数时候，什么是可恢复的取决于API的调用者，而不是API的编写者。API不应该使用受控异常，因为只有未受控异常才允许选择捕获或忽略异常。

当你想要向调用者提供信息时，受控异常对于可恢复的情况很有用(例如，权限不足，文件未找到等)。

未检查异常很少用于在运行时通知用户或程序员严重错误或意外情况。如果你编写的代码或库将被其他人使用，不要抛出这些异常，因为他们可能不希望你的软件抛出未经检查的异常，因为编译器不会强制捕获或声明这些异常。

来自Java学习者:

When an exception occurs, you have to either catch and handle the exception, or tell compiler that you can't handle it by declaring that your method throws that exception, then the code that uses your method will have to handle that exception (even it also may choose to declare that it throws the exception if it can't handle it). Compiler will check that we have done one of the two things (catch, or declare). So these are called Checked exceptions. But Errors, and Runtime Exceptions are not checked for by compiler (even though you can choose to catch, or declare, it is not required). So, these two are called Unchecked exceptions. Errors are used to represent those conditions which occur outside the application, such as crash of the system. Runtime exceptions are usually occur by fault in the application logic. You can't do anything in these situations. When runtime exception occur, you have to re-write your program code. So, these are not checked by compiler. These runtime exceptions will uncover in development, and testing period. Then we have to refactor our code to remove these errors.

受控异常非常好，只要你知道什么时候应该使用它们。对于SQLException(有时对于IOException)， Java核心API无法遵循这些规则，这就是它们如此糟糕的原因。

受控异常应该用于可预测的、但无法预防的、可以合理恢复的错误。

未检查异常应该用于其他所有事情。

我来解释一下，因为大多数人都误解了这句话的意思。

Predictable but unpreventable: The caller did everything within their power to validate the input parameters, but some condition outside their control has caused the operation to fail. For example, you try reading a file but someone deletes it between the time you check if it exists and the time the read operation begins. By declaring a checked exception, you are telling the caller to anticipate this failure. Reasonable to recover from: There is no point telling callers to anticipate exceptions that they cannot recover from. If a user attempts to read from an non-existing file, the caller can prompt them for a new filename. On the other hand, if the method fails due to a programming bug (invalid method arguments or buggy method implementation) there is nothing the application can do to fix the problem in mid-execution. The best it can do is log the problem and wait for the developer to fix it at a later time.

除非您抛出的异常满足上述所有条件，否则它应该使用未检查异常。

Reevaluate at every level: Sometimes the method catching the checked exception isn't the right place to handle the error. In that case, consider what is reasonable for your own callers. If the exception is predictable, unpreventable and reasonable for them to recover from then you should throw a checked exception yourself. If not, you should wrap the exception in an unchecked exception. If you follow this rule you will find yourself converting checked exceptions to unchecked exceptions and vice versa depending on what layer you are in.

对于已检查和未检查的异常，使用正确的抽象级别。例如，具有两种不同实现(数据库和文件系统)的代码存储库应该通过抛出SQLException或IOException来避免暴露特定于实现的细节。相反，它应该将异常包装在一个跨越所有实现的抽象中(例如RepositoryException)。

You can call it a checked or unchecked exception; however, both types of exception can be caught by the programmer, so the best answer is: write all of your exceptions as unchecked and document them. That way the developer who uses your API can choose whether he or she wants to catch that exception and do something. Checked exceptions are a complete waste of everyone's time and it makes your code a shocking nightmare to look at. Proper unit testing will then bring up any exceptions that you may have to catch and do something with.