On 18.9.2012, at 2.02, Kelsey Cummings wrote:
We are seeing a few (0-15) proxy failures like the following out of ~3m successful proxied connections a day. Average session creation load over our peak hour is about 47/sec. The backend servers aren't logging anything that would suggest any internal problem like insufficient processes to handle the load. It doesn't seem to happen when utilization is lowest at night.
dovecot: imap-login: Error: proxy(foo): connect(1.1.1.1, 143) failed: Connection timed out (after 63 secs)
I'm curious if anyone else has seen any similar problems or has any suggestions.
I once had similar problems when the proxy backend was Courier. The problems went away after migration to Dovecot was complete. The possibilities are either:
a) The backend server is busy and doesn't have a chance to accept() the connection.
b) Packets get dropped in the network and the retry packet is slow in coming (or also gets lost).
Changing some kernel settings might help with a). There are also kernel settings that specify how SYN resend is attempted, you could try reducing it to a few seconds.