Hey all, We've recently been encountering an issue where memcached will keep connections open to our back-end servers even after those servers have closed their connection to memcached. Our hunch is that the backends, when they time out, don't cleanly close the connection and memcached keeps holding onto them for hours afterwards.
The curr_connections will report tens of thousands of connections, even after we've bounced all application servers to point to a different memcached. netstat -na also shows these connections, as well as lsof. We use pylibmc 1.2.3 and libmemcached to connect to memcached. Any ideas for tweaks to kernel settings or libmemcached settings to keep this from happening? We use ketama/md5 for our hashing, and have tcp_nodelay and noblock enabled. Thanks, Mike
