This is cool. I implemented an extremely similar system for my last job. The data was dumped into mongo and then we visualized it with stacked bar graphs.

One thing to consider since you're using a count+total model is that the most interesting timings will often be the 90th or 99th percentile, so by calculating averages you might be missing useful information.

I ran into some issues with the implementation after switching to using an async framework since the code was no longer a series of nested function calls. Since the current best practice is coroutines where this will still work I think it's okay, but you should consider how someone using callbacks might time their code. In my case I was in a hurry so I manually called the equivalent of your __enter__ and __exit__, but it was pretty ugly and left a lot of room for bugs.

See also: http://pycounters.readthedocs.org/en/latest/

Trivial but useful. Next time, instead of kicking in my own custom timing decorator (5-liner, but...) I'll probably use this. [Edit: pycounter looks even nicer, didn't heard of them before, thanks for sharing!]

The only downside I see, it does record function name, but doesn't record module name (and, for class members, classname). For example, it wouldn't be too useful to see "authorize" instead of "ppp.common.authorize" in RADIUS server profiling logs. :)

We built a similar decorator/context manager for profiling, but it is really much more useful if you can export the data to statsite + graphite so that you can graph and view it on an on going basis. The insight into the runtime is much more valuable when you have historic data to compare it to.