I've recently set up a director proxy environment on my test servers, with the intention of deploying on our cluster soon.
One thing I found confusing in the proxying documentation [1] was the first bit about their being two ways to do the authentication...either you have the proxy forward the auth to the real server for authentication, or you have the proxy authenticate it and then login to the real server with a master password.
Well, we use /bin/checkpassword authentication which hooks into a variety of subsytems for various specific customer needs, and sometimes we need to know the username AND password of the user in order to determine their home directory information. So, using a master password (which requires the back-end server not getting the user password) is out.
However, when we have the front-end server do a static director proxy, the problem is that authentication failures are logged on the back-end server with a source IP of the proxy, and no authentication failure with the client IP address is logged on the proxy. So, fail2ban (which is a MUST these days, at least for us) will not be able to properly filter out the brute force attackers.
My solution was an alternative: I authenticate with our /bin/checkpassword on the proxy, which authenticates the user and only at that point returns the proxy=y nopassword=y switch to proxy the connection and forward the authentication.
As a result, we get logs on the proxy for failed attempts, and the full username and password is supplied to the back-end server for proper processing.
Food for thought in case anybody else is implementing this.
Thanks, Andy
[1] http://wiki2.dovecot.org/PasswordDatabase/ExtraFields/Proxy