The C++ standard library’s exception hierarchy is IMHO pretty arbitrary and meaningless. For example, it would probably just create problems if anyone started actually using e.g. std::logic_error instead of terminating when it’s clear that the program has a Very Nasty Bug™. For as the standard puts it,
“The distinguishing characteristic of logic errors is that they are due to errors in the internal logic of the program.”
Thus, at the point where it might otherwise seem reasonable to throw a std::logic_error the program state might be unpredictably fouled up, and continued execution might put the user’s data in harm’s way.
Still, like std::string the standard exception class hierarchy has a really really practically important and useful feature, namely that it’s formally standard.
So any custom exception class should be derived indirectly or (although I would not recommend it) directly from std::exception.
Generally, when the debates about custom exception classes raged ten years ago, I recommended deriving only from std::runtime_error, and I still recommend that. It is the standard exception class that supports custom messages (the others generally have hardcoded messages that one preferably should not change, since they have value in being recognizable). And one might argue that std::runtime_error is the standard exception class that represents recoverable failures (as opposed to unrecoverable logic errors, which can’t be fixed at run time), or as the standard puts it,
“runtime errors are due to events beyond the scope of the program. They cannot be easily predicted in advance”.
Sometimes the C++ exception mechanism is used for other things, treated as just a low-level dynamic destination jump mechanism. For example, clever code can use exceptions for propagating a successful result out of a chain of recursive calls. But exception-as-failure is the most common usage, and that’s what C++ exceptions are typically optimized for, so mostly it makes sense to use std::runtime_error as root for any custom exception class hierarchy – even if that forces someone who wants to be clever, to throw a “failure”-indicating exception to indicate success…
Worth noting: there are three standard subclasses of std::runtime_error, namely std::range_error, std::overflow_error and std::underflow_error, and that contrary to what their names indicate the latter two are not required to be be generated by floating point operations and are not in practice generated by floating point operations, but are AFAIK only generated by some – surprise! – std::bitset operations. Simply put, the standard library’s exception class hierarchy seems to me to have been thrown in there just for apperance’s sake, without any real good reasons or existing practice, and even without a does-it-make-sense check. But maybe I missed out on that and if so, then I still have something new to learn about this. :-)
So, std::runtime_error it is, then.
At the top of a hierarchy of custom exception classes, with C++03 it was useful to add in the important stuff missing from C++03 standard exceptions:
Virtual clone method (especially important for passing exceptions through C code).
Virtual throwSelf method (same main reason as for cloning).
Support for chained exception messages (standardizing a format).
Support for carrying a failure cause code (like e.g. Windows or Posix error code).
Support for getting a standard message from a carried failure cause code.
C++11 added support for much of this, but except for trying out the new support for failure cause codes and messages, and noting that unfortunately it’s pretty Unix-specific and not very suitable for Windows, I haven’t yet used it. Anyway, for completeness: instead of adding cloning and virtual rethrowing (which is the best that an ordinary application programmer can do in a custom exception class hierarchy, because as an application programmer you cannot hoist a current exception object out of the storage that the implementation’s exception propagation uses), the C++11 standard adds free functions std::current_exception() and std::rethrow_exception(), and instead of support for chained exception messages it adds a mixin class std::nested_exception and free functions std::rethrow_nested and std::rethrow_if_nested.
Given the partial C++11 support for the above bullet points, a new and modern custom exception class hierarchy should better integrate with the C++11 support instead of addressing the C++03 shortcomings. Well, except for the C++11 failure code thing, which seems to be very unsuitable for Windows programming. So, at the top of the custom hierarchy, right under std::runtime_error, there will ideally be at least one general exception class, and derived from that, one exception class that supports propagation of failure codes.
Now, finally, to the gist of the question: should one now best derive a unique exception class for every possible failure cause, or at least for major failure causes?
I say no: DON’T ADD NEEDLESS COMPLEXITY.
If or where it is can be useful for a caller to distinguish a certain failure cause, a distinct exception class for that is very useful. But in most cases the only information of interest to a caller is the single fact that an exception has occurred. It is very rare that different failure causes lead to different attempted fixes.
But what about failure cause codes?
Well, when that's what an underlying API gives you, it is just added work to create corresponding exception classes. But on the other hand, when you are communicating failure up in a call chain, and the caller might need to know the exact cause, then using a code for that means the caller will have to use some nested checking and dispatch inside the catch. So these are different situations: (A) your code is the original source of a failure indication, versus (B) your code uses e.g. a Windows or Posix API function that fails and that that indicates failure cause via a failure cause code.
