[Zope] ZEO clients going zombie

[Dieter Maurer]

Bengt Giger wrote at 2007-4-10 15:36 +0200:
>with our new ZEO client systems, we frequently observe this problem:
>
>a ZEO process starts to use 100% CPU time (user) without a significant
>increase of requests. Sometimes (but not always) the process stops answering
>requests, still using 100% CPU.
>
>When we kill such a process, it changes to zombie state (shown in top as 'Z'
>and '<defunct>'), still using 100% CPU, but now its system time, not user. The
>HTTP port is still in use, so we have to reboot this node to restart the ZEO
>client. This usually fails because some filesystems cannot be unmouted, there
>are still files locked.
>
>I tried both start modes, runzope and zopectl, but no difference.
>
>All that is in opposition of what I know about zombie processes, they should
>use no CPU time.

I may have seen a similar problem (though I am not sure about the details):

  The problem was a buggy Python signal handling together with
  a doubtful Linux thread implementation (2.4 kernel).

  When a fatal signal occured, the signal killed the main thread
  but all other threads were in a strange state. Only a "SIGKILL"
  could get them out of this state. Of course, the ports
  remained open although nobody serviced them.

  I am not sure whether the threads were in zombie state or used
  CPU. Probably not.



-- 
Dieter