Re: Replication Problems

[kevin mcgowan <clunis@xxxxxxxxx>]

replication errors don't surprise me at all.  I'd be very surprised if 
an application were getting errors, though.  Replication is 
intentionally lossy.  We designed it to drop replication events under 
load rather than block on user requests.  The filters are designed to 
query all available cosign servers so a replication failure is not a 
problem.

From the application user's perspective even a 100% loss rate for 
replication should not cause a problem (whatever is causing the 100% 
loss rate might cause a problem, but replication is never absolutely 
necessary, it just makes everything more efficient).

We have about 39,533 students (graduate, undergraduate, professional), 
5,330 faculty, 10,287 hospital staff, 12,757 non-hospital staff, and 
about 26,000 Friend guest accounts.  We had 172,395 login events across 
our three weblogin servers yesterday.  We saw roughly 259,035 distinct 
REGISTER requests yesterday divided among 95 services (interestingly, 
the distribution of users per service roughly follows Zipf's law).  We 
don't log CHECK at all anymore unless you're running with debugging 
turned on so I'm not sure how to get you that information.  Our 
server's cosign databases are normally 90 - 95% in sync with each 
other.  I hope this is helpful.

Our heaviest usage period is in the late afternoon.  Between 3 and 4 
o'clock yesterday we saw about 10,551 LOGIN events and 17,397 distinct 
REGISTER events.

Kevin

On Feb 2, 2005, at 4:28 PM, Brett Lomas wrote:

> That's fine, thanks heaps for that Kevin
>
> I am not surpirsed about the replication error though, because if the
> machine is under high load, and has like 1000 processes most of which
> are cosign daemons servicing requests, then the replication process 
> will
> only get a small amount of CPU time to actually process the events. At
> least that is my understanding of things. That is why I am worndering 
> if
> I am putting an unreasonable load on it.
>
>
> Brett

                ... "you can't give yourself a nickname." ...