Handling failovers (rewrite)

[Dustin Sallings]

	[I responded to this off-list because I somehow received a copy in  
my inbox that didn't include the list]

On Jul 4, 2007, at 2:15 , a. wrote:

> Isn't it true that a memcached server more likely "goes down"  
> becaouse its machine has been restarted, or the process crashed?
>
> So usually you'll end up with an empty server which does not have  
> stale data.

   The scenario we were discussing here was a server going down  
(causing all requests to go to another server), coming back (writes  
causing updates to the server that isn't primary), and then going  
down again (causing stale reads).

   It's not the server that went down, but the next server in the  
list whose freshness is a concern.


	Losing connectivity to a node does cause a failure mode where the  
secondary was updated, and then the primary comes back with stale  
data.  Sending a flush on reconnect would prevent that from being a  
problem, but could wreak havoc on your primary data store when a  
client has a network instability.

-- 
Dustin Sallings


-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.danga.com/pipermail/memcached/attachments/20070704/3b0b9bf3/attachment.html