So about three years ago, around the time of my last blog post, I promised a friend I would write this post. Keeping promises is a good thing, and this is my first step towards easing back into regular blogging. I fully expect him to return from Pennsylvania to buy me a beer over this. However, it’s been an… ahem… eventful three years or so, and blogging, unfortunately, got pushed to the back burner on my priority list, along with a few other career minded activities. Now that the personal drama of the past three years is more or less resolved, it’s time to put a few things back on the front burner.

What I consider to be proper exception handling practices is relatively well known these days. There are plenty of blog posts out there already on this topic which more or less echo my opinions on this topic. I’ll try to include a few links at the bottom of the post. Several years ago I had an argument with a co-worker who posited that exceptions should be caught at every level and logged. This might seem like sanity on the surface, but the resulting error log looked something like this:

Error: System.SomeException
Followed by small stack trace.

Error: System.SomeException
Followed by slightly bigger stack trace.

Error: System.SomeException
Followed by slightly bigger stack trace.

Error: System.SomeException
Followed by slightly bigger stack trace.

Error: System.SomeException
Followed by slightly bigger stack trace.

Error: System.SomeException
Followed by slightly bigger stack trace.

Error: System.SomeException
Followed by slightly bigger stack trace.

Error: System.SomeException
Followed by slightly bigger stack trace.

These were all the same exception. The problem with this approach is that the error log, if you run any kind of analytics on in, becomes skewed depending on how far up the stack trace your exception was thrown. To mitigate this problem, we came up with the concept of the “PreLoggedException”. Basically, we would log the exception at the very top level and subsequently throw the exception back up the stack encapsulated in this pre-logged type, which our logging system knew to ignore. Now the error log looked like this:

Error: System.SomeException
Followed by small stack trace.

Much cleaner, right? Well, there’s still a problem. When your exception happens in production and you go about trying to figure out what happened, you’ve lost more or less all context for where and how this exception was thrown, because all you really know is what method it was thrown in, but really nothing about who was calling the method or why. What gives you this clue is the entire stack trace, which we’re losing here. I believe that was further mitigated by having the logging system pull a system stack trace and add it to the log entry, but what you’re actually getting is the stack for how you got to the logging code. You’re still losing context about the actual error. Not to mention you’re executing a whole slew of catch blocks which are sloooooooowwwww………

In other words, we started with a bad idea and kept band-aiding it until it didn’t suck quite so bad.

When I argued for not catching exceptions at every level but rather catching them following a certain set of rules, my co-worker warned me “do yourself a favor, never express that view in any future interviews.” I suppose this is my ultimate dismissal of that advice, but I’m not too worried.

My approach for exception handling follows three basic rules:

Only catch an exception if

1. You can do something about it.
2. You can add useful information to it.
3. You’re at an application boundary.

Here’s what that means:

1. Only catch an exception if you can do something about it.

We’ll start with a trivial example of a login system that uses a file. Please, never actually do this in production code, it’s just concocted example. So if our code goes to open a file and the file isn’t there, we get a FileNotFound exception. If the calling code doesn’t know what to do with this, it should bubble up. However, if we know how to create the file from scratch we can create the file and continue on our merry way. When you run into situations like this though, What should really run through your head is “How can I avoid handling an exception at all?” In this case, it’s a trivial matter to simply check for the existence of the file before trying to open it. If we detect that the file isn’t there, we can accomplish the same thing without having to handle in in a catch block.

2. Only catch an exception if you can do something about it.

Continuing with the poorly thought out file based login system we contrived in part 1, if the code calls a Login(…) method and the FileNotFound exception is thrown higher up the stack, the code that calls Login must account for a FileNotFound exception. This is kind of counterintuitive because the calling code should not need to know the internals of the Login method, and the data file is an implementation detail. What makes more sense, assuming that we didn’t implement any of the good advice from step 1, is for Login to catch the FileNotFound exception and wrap it in a new exception. For argument’s sake we’ll say LoginSystemFailureException. (Sorry, couldn’t think of anything better at the moment.) This gives us two stack traces, preserving the original stack trace in the inner exception, and also is much more informative to the calling code.

3. Only catch an exception if you’re at an application boundary.

At some point we have to catch all the exceptions, even the ones we don’t know what to do with. WinForms, ASP.Net, and most other UI technologies have some kind of built in mechanism for catching unhandled exceptions without fatally terminating the application. It’s still a good idea to somehow gracefully exit the application in this case if possible though, because you can no longer be sure what state your application is in, but nothing annoys a user more than an application just exploding. These unhandled exceptions need to be logged, and this is a good place to catch them. Ideally you never want this option to be exercised, but code as though it will be. When you log these exceptions, give them a “Fatal” status (e.g. Log4Net) and make sure these bugs get handled in your next release.

That’s it in a nutshell. If you do it right each exception will only get logged once and with the largest stack trace possible which will make those 2am emergency severity 1 debugging sessions much shorter and less frustrating.

Here’s a few people who also have interesting things to say on this topic: 

http://blogs.msdn.com/b/ericlippert/archive/2008/09/10/vexing-exceptions.aspx

http://www.codeproject.com/Articles/9538/Exception-Handling-Best-Practices-in-NET

I know there’s more but I can’t find them at the moment.