Monday, 14 December 2009

Checked Exceptions Part 2

So after the big rant over checked exceptions I've been bitten by the problems that come with the them, not for the first time but it seemed all the more acute since I just wrote about them. By modifying the signature of a method to convert ints to enum types I started a change that meant modifying forty files from the the depths of the application to it's highest spire, and overall creating three different exceptions to encapsulate the added exception and IOExceptions that could already occur. Take into account that some of these files were unit tests that would need to change to throw the exception I fit occurred but did not need ant more modification to pass.

The problem with specifying a checked exception here would be the same if Integer.valueOf(String) threw a checked version of NumberFormatException, try-catches and throws statements appeared everywhere. This case is pretty much the worst case for this change as the method is used in more places than I expected, this is an API meant for eventual consumption by people who I may never meet. Who may not have the code there to view the documentation. I left work tonight without committing that work into SVN to think whether this change was really worth all the upheaval it had caused.

To clarify, the method already threw an IllegalArgumentException that was duly documented. The reason I created a custom checked exception was because the unthinkable happened, it actually threw the runtime exception and although it was caught it didn't give me any comfort. Luckily the the method executing the logic throwing the exception handled and logged it. While adding the tries and throws I realised that it was only caught appropriately in about 20% of the cases that it was used. I think this is a good estimate at how many runtime exceptions are caught appropriately in most software, luckily things mostly go our way ;)

After some serious thought I realised that had the method threw the exception when it was created the same amount of error handling code would exist and I would be happy because now the software is more robust. The exception will never be thrown out of core application code i.e. libraries such as Swing with unforeseen results.

The lesson I learned from this isn't that checked exceptions are hard, it's that they are hard to change after the code is in use. This is in exactly the same way that changing method parameters is hard once the code is in use. A couple of things to look out for next time:

Next time the method is going to throw runtime exception perhaps a custom checked exception should be thrown. The runtime exception will most likely not be caught, should the software be allowed to crash if this happens?

Exception hierarchy is your friend. By throwing subclasses of MyCompanyException (which is common but pointless on its own) catchers can easily cope with and log this but many custom versions can be created for more fine grained strategies. Also this may already be caught so no modification required in those cases.

One more conclusion that surprised me is that I actually found some existing errors lurking in code, two stream.close() cases that weren't contained in a finally block. If an exception was thrown then they would not be closed. Not the worst problem in the world but it shows how when I was in error handling mode how I saw code differently.

No comments: