On 11 Mar 2003 18:21:02 +0100 Henrik Nordstrom <[EMAIL PROTECTED]> wrote:
> The above continues until all server instances on the backend is busy > processing requests and the TCP backlog has been filled, causing > further connection requests to the server to be rejected. > > If you are using some health monitoring then this monitoring should > detect the situation and mark the backend server as offline. The problem is we have multiple NFS mounts on our origin and have instances were only one fails, therefore, only a subset of objects are affected. The origin, for a while, continues to serve normally for all other objects. Eventually, the origin backs up and the checker notices and, in our hacked version, puts squid into offline mode. This is usually temporary (less than a minute) and then the origin is able to server, so squid goes back into normal operation. This cycle repeats. The affected objects, even after the origin servers them correctly, are never served from cache until a reconfgiguration or restart. --Brian
