We (Twitter) looked at our running Rails instances (under mongrel) and found that much of our latency was caused by extremely large amounts of time spent generating backtraces because of exceptions being raised.
The vast majority of the raised exceptions originated from the locations fixed in these two patches.
Normally, using raise / rescue wouldn't be a problem, but in many of these instances, the stack trace is hundreds of frames deep, and takes a really long time to generate. Since the generation of the stack trace happens automatically when an exception is raised, the best approach is to avoid raising exceptions altogether where avoidable.
I ran a number of benchmarks against the changes, but as mentioned above, these issues really only surface when the stack traces are large and time-consuming to generate; within the context of tests, it's not a significant difference (although relevant tests ran reliably approximately 2 milliseconds faster).
I've attached some brief output that includes the total time spent in CPU and the corresponding total time spent in CPU in backtrace generation, over one second intervals, running against our live mongrel processes.
Please contact me at blaine@obvious.com if any more information is required. Thanks!
