On 5/26/2010 8:06 AM, Blake Hudson wrote: > I wish it were that straightforward. After performing several > switchovers where DNS A records were repointed, many clients (days > later) continue trying to access the old servers. TTL on the DNS records > are set appropriately short, this is simply a case of mail clients > inappropriately caching DNS results. The fix for this is to restart the > client, sometimes requiring a reboot of the machine to accomplish. > I cannot state STRONGLY ENOUGH that using DNS A records for service load balancing is irresponsible & antiquated in an HA datacenter environment as the current clunky state of the "larger world" stands. There are too many ill-behaved caching devices out there. There are far too many Windows machines that do not pay any attention at all to TTL. Let me give you an example. About 2 years ago we moved our Perdition proxy pool to a new IP. I moved 2 of the 4 servers to the new IP after hours before the DNS update. 2 days later we were still seeing traffic at the old IP, but traffic had subsided sufficiently to move the 3rd machine. A MONTH LATER the traffic at the old IP had finally dropped to a very low (but still not zero) level to move the 4th machine and just live with the possibility of a few phone calls. We take great pains now to not rely on DNS robin for *anything*, or DNS itself to split load or buy us any HA, and we account for the pitiful REALITY of how poorly DNS propagates now in planning service moves. From more recent moves I can tell you the situation has not improved, the decay of usage of old IP takes a long time to decay below 5% if you are dealing with many users outside your site. ---- Cyrus Home Page: http://cyrusimap.web.cmu.edu/ Cyrus Wiki/FAQ: http://cyrusimap.web.cmu.edu/twiki List Archives/Info: http://asg.web.cmu.edu/cyrus/mailing-list.html
