During my development of toppra, a thing bugs me quite often: how to handle failure? The problem is that toppra is a fairly complicated algorithm with several hotspots for numerical instabilities. In another words, it fails often and in different places. For example, consider the below pseudo-code:

def solve_toppra(inputs):
    ret = func1()
    ret = func2(ret)
    ret = func3(ret)
	return ret

Now, the algorithm can fail at different points. As a result, there are a lots of error-checking stuffs in my implementation:

def solve_toppra(inputs):
    ret = func1()
    if ret is None:
        return None
    ret = func2(ret)
    if ret is None:
        return None
    ret = func3(ret)
    return ret

Now, the code is rather convoluted. The original version without error checking looks much cleaner and resemble the pseudo-code better. Furthermore, it is also very difficult to know why the algorithm fails, as the result of solve_toppra is None, conveying next to nothing information. Indeed, everytime toppra fails I have to look into the logs to figure which part has gone wrong, which is really tedious.

The solution http://www.cauldwell.net/patrick/blog/ThisIBelieveTheDeveloperEdition.aspx