Should Language Changes Break Existing Code?

comments edit

I received a lot of comments (a lot for me) on my post entitled "The Difficulties of Language Design".

I wanted to follow up on one interesting comment by a reader named Jocelyn:

"Language changes shouldn't break existing code..."Well, yes or: \ \ - obsolete features (like the lock keyword) could be flagged\ - tools could be developped to update existing source code\ - the language could be versionned:\ #version 1.1\ #version 2.0\

That's why I qualified my statement with "...Too Much.". There are certainly cases where you have to take the plunge and risk breaking existing code. The things she mentioned are certainly great ways to mitigate the impact of changes, but they aren't enough.

I think the real difficulty is when you slightly change the behavior of a language feature such as a keyword. This change won't show up when you recompile your code because you aren't marking the feature as obsolete. Likewise, it can be quite hard for code analysis tools to check to see if the semantics of your code relies on the old behavior (though in some cases this might be possible) and would have problems with the new behavior. The best it could do is flag the keyword and say "Hey! The behavior of this keyword has changed." This might be helpful in some cases, but imagine if the behavior of the lock statement changed slightly. That's a lot of places you're going to have to check by hand.

The end result is that you recompile your code using the newer language and everything looks hunky dory. But days, maybe weeks, later you find a subtle problem with your code that is difficult to track down. In the end, it may be the end result of a chain of events that started at the point where your code relied on a certain behavior of the language and that behavior changed. The point here is that the error might not occur at the point where you rely on the faulty behavior, but somewhere down the line.

I'm not advocating that the behavior of language features should never change, especially if the behavior is wrong to begin with. I'm merely pointing out the risks and hazards of doing so. It's a heavy cost and the benefit sure as hell better be worth it. I think this is why you see so few breaking changes.

Comments