[Catalog-sig] psycoph errors from pypi

["Martin v. Löwis"]

> Ah, OK, that explains it.  This is a reasonable thing to do from a
> performance point of view.  Thanks for plugging away at this. :)
> 
> (Of course it's too bad we don't have a better way of testing changes.
> Oh well.)

If there were volunteer testers, it would be possible to test changes
for some period of time. Such testers would have to build themselves
a PyPI installation, and then checkout all changes that have been
committed (or install them from a tracker where they float around).

Alternatively, if somebody contributed a unit test suite, certain
problems might get caught.

In the specific case, I tested whether PyPI "works" on my local
installation, and I apparently didn't not manage to trigger the
problem. My guess is that it was originally triggered by some
failing concurrent access, which is really hard to test for.

Regards,
Martin