As an update, we're not 100% stable yet. We've had the main Apache process crash out a couple of times while running mod_fcgid, and we're still looking into the cause. Any thoughts from folks here would be much appreciated. For a little over a weeks worth of traffic, we get:

• Over 800 instances of "An established connection was aborted by the software in your host machine. : core_output_filter: writing data to the network"
• Over 670 instances of "An existing connection was forcibly closed by the remote host. : core_output_filter: writing data to the network"
• Over 430 instances of "mod_fcgid: read timeout from pipe", the same number of "exit(communication error)" messages with a variety of return codes (-1073741819, 9, 255), and the same number of "Premature end of script headers" (these errors are therefore likely related)
• Over 300 warnings of "mod_fcgid: process X graceful kill fail, sending SIGKILL"
• A small number of other warnings ("Connected party did not respond properly", "mod_fcgid can't apply process slot", "request failed: error reading the headers")

For info, we are running with the following options:

KeepAlive On
MaxKeepAliveRequests 200
KeepAliveTimeout 5

ThreadsPerChild 125
MaxRequestsPerChild 10000

Win32DisableAcceptEx
EnableMMAP off
EnableSendfile off

I wonder whether bringing the ThreadsPerChild down might help? AFAIK the threads are created and are unused, since all the traffic goes through FCGI. I am tempted to turn off KeepAlive too. If we completely run out of ideas, I will try replacing FCGI with FastCGI. Any thoughts how I might trace the core issue?