collect_children is only called in process_request, so at least the last
process forked is not collected. This could be handled by calling
collect_children every X seconds or by handling SIGCHLD.

Can you provide a patch?

Next time please fill in type, components and versions, too. :)

Using signals is asking for all sorts of trouble.

What is the point of collecting the last process? Having one zombie
child max as a steady state strikes me as totally acceptable.

Well, one child isn't the maximum. All children which do not finish
before the last client connects, will live on as zombies (and will be
collected when the next client connects). If that is acceptable, then
just close this report.

Using signals would be easier if http://bugs.python.org/issue1089358
goes in. But even then a single accept call might not suffice and the
code might need to select on the listening socket and another
filedescriptor (on which the signal handler writes).

The simplest solution might be to use select with a timeout and check
for dead children periodically.

(and sorry for not filling out the complete form.)