[OpenAFS] DB servers seperate from fileservers

Jeffrey Hutzelman jhutz@cmu.edu
Tue, 08 Aug 2006 15:13:35 -0400

On Tuesday, August 08, 2006 02:49:41 PM -0400 Derek Atkins 
<warlord@MIT.EDU> wrote:

> Ken Hornstein <kenh@cmf.nrl.navy.mil> writes:
>>> I was specifically talking about DB servers.  Having one of them go
>>> down, provided there are no volumes on that server, should not cause a
>>> problem, right?
>> In theory, no.  In practice ... it's annoying.
>> The problem is that AFS clients pick a DB server to talk to at random,
>> and for most cases stick with that server.  For a long-running bit of
>> software like the cache manager, a single server going down isn't a huge
>> problem; the client notices that server X is down, and then won't use
>> that server again.  But for userland programs, they have no idea what
>> the current state is for ubik servers, so they pick one at random ...
>> and sometimes they pick the down one, and you get delays when using
>> "vos" or "pts".  Those commands will eventually work ... they just
>> sometimes take a while to find the right DB servers.
> I wonder if we could add a cache manager query so that apps could
> ask the cache manager for server stats?  Obviously this would be
> an optional query, and the app should continue to use the current
> random choice if there is no cache manager.  But I think it would
> help this case if there were some way to store this kind of state
> on a client.

You mean like VIOCCKSERV, which tells you which servers the cache manager 
currently thinks are down?  Or more like VIOCGETSPREFS, which tells you 
what preference order the cache manager has assigned to servers?

-- Jeff