Re: Quality Assurance of mailserver?
On Sun, 15 May 2011, Andrew McGlashan wrote:
> Yes, but most won't be using ksplice, that would be the exception --
> I did say "generally" for a reason. Besides, it is true that even
> with ksplice that you can get by for over 12 months without a reboot
> and also have the latest kernel and all other updates working fine?
ksplice is half-a-reboot. And no, it is not advised to have extremely long
uptimes, with or without ksplice. Anyone with a clue uses ksplice to keep
things limping along until the next downtime window, and _has_ such regular
downtime windows scheduled throughout the year. Often they're so short
(covers just the window where HA failover could couse some service
disruption) you don't even know about them.
Reboots fix any latent kernel state corruption, most hardware state
corruption, forces all the hardware buses to be retrained (QPI, HT, PCIe,
SAS/SATA...), scrubs all the ECC memory, gives the baseboard management
system a spin, and does a lot of interesting housekeeping. At least on
servers.
Besides, he who doesn't reboot, doesn't know if the service will come up
when he finally *has* to reboot outside of a maintenance window. Best
practice is to have fast HA failover and use that to get invisible
maintenance windows for boxes, with no or very little service disruption.
Also, if you use 2.6.32 (confirmed in Debian squeeze), you currently are
WELL ADVISED TO REBOOT WITHIN 190 DAYS. There is a nasty bug somewhere that
triggers *sometimes* and causes kernel state corruption. So far, it is
proving to be very difficult to track down. This is NOT the first time
we've had bugs like this, that show up only on very long running systems.
Maybe ksplice is enough to avoid it. Maybe it isn't.
--
"One disk to rule them all, One disk to find them. One disk to bring
them all and in the darkness grind them. In the Land of Redmond
where the shadows lie." -- The Silicon Valley Tarot
Henrique Holschuh
Reply to: