Talk:Blogs:AndreiBadea:EnumsInAPIs

Normally in a release version one wouldn't want any asserts enabled so maybe not a very big problem? The general principle is if you are designing a library you don't make any assumptions on how your client calls your API. Perhaps I'm wrong though...

--Alex 20:21, 16 July 2008 (CEST)

In this example, if the user wanted to ensure that he covers all the fields of the enumeration, then he wrote something wrong, he should not have the "default:" case.

If the switch covers all the fields of the enumeration, the "default:" is not needed and not warnings/errors are generated at compile time.

We can also assume that if the user use the default close here, his meaning is not to cover all the fields of the enumeration, but to only cover a subset and throw an assert exception when encounter a field of the complementary subset or null.

--Vincent Cantin 03:26, 17 July 2008 (CEST)

To Alex: I have also seen IllegalStateException's thrown instead of an assert statement. Regarding the assumptions, you are right. But this case looks like one where you could want to break the rule. Not being able to add fields to the enum could be a big limitation.

To Vincent: you are right, many people have a default branch just to be sure, even when they know they have covered all fields of the enum (of the current version of the enum, that is).

--AndreiBadea 15:41, 17 July 2008 (CEST)

Could want to break a rule? This rule having been expressed in various ways by hthe MONSTERS of computer science would be then too strong for me to want to break it. There must be a subtlety I'm missing when following your example. I shall work on that further. That is after I've studied the design patterns. Otherwise no chance for me to pass the exams. Thanks!!

--Alex 23:28, 17 July 2008 (CEST)

Doesn't this depend on how the enum is used?

If a value of the enumerated type is in the result of an API method, and the switch statement above is in the client, and the new value can be returned in a result produced by previous-version client behavior, then adding a value to the enum can cause breakage.

But if the only uses of the enumerated type are in arguments to API methods, or the new value can only appear in a result for a request that used other new–version aspects, then this specific opportunity for error doesn't arise.

It is still possible for new semantic requirements (e.g. a rule that the new enum value is required in certain cases) to break the client.

I suggest that the real issue is this: if the new API still accepts all prior-version arguments and (for those arguments) returns the same results, then there's no breakage. Breakage occurs when a new version fails to accept previously-valid arguments, or returns results that previously could not occur.

--Joel Neely 13:37, 25 August 2008 (CEST)

If you compile your release with _NDEBUG the switch will fall through just fine since the assert macros are removed from the release code. This should be decent behaviour even with new binary versions of the code if the new enums are always added at the end.

Developers running debug code *might* catch the new enum value if it is set by the library, and the failed assertion will alert them to the new value that is available.

Basically it comes down to programming practices and disciplines for it to work, which are sometimes difficult to maintain in multi-user distributed development.