Question about MUPDATE master mailboxes list

[Rob Siemborski]

On Tue, 17 Feb 2004, Etienne Goyer wrote:

> Why do you recommend *against* relying on this behavior ?

Because it takes much longer to write all the entires to the database than
just writing anything that the database has incorrect (as, I guess, you
discovered).

> For about 500K mailboxes (mailboxes, not accounts), how long should
> "ctl_mboxlist -m" from two backends run ?  I know it's impossible to
> give exact as it rely on many factors (hardware, network speed, etc),
> but I would be satisfied with a rough estimation.

It mostly should depend on how much of the database is already populated
in the mupdate master.  If there's nothing there, and the database is
syncing on every write, then I would expect 500,000 entries to take a VERY
long time to run.

> I am asking because, starting with a clean (just did mkimap) configdir
> on the mupdate master, it took 8 hours for us with fairly powerful and
> otherwise idle machines.

That seems to be a not-unsurprising timeframe given the number of updates.
(If you are using skiplist, there are two fsync() calls for every update).

> > It shouldn't be, provided you are running ctl_mboxlist -m at startup of
> > the backends.
>
> Well, we do.
>
> On the mupdate master, is it advisable to :
>
> 1. have master run "ctl_cyrusdb -r" at START ?

Yes.  This is an outright requirement of *any* cyrus system that uses
atleast one database.

> 2. have master checkpoint the database with "ctl_cyrusdb -c" as a
> regular EVENTS ? (I know I am going to hate myself when you answer this
> one ...)

Yes.  *Especially* if it was berkeley DB.

-Rob


-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
Rob Siemborski * Andrew Systems Group * Cyert Hall 207 * 412-268-7456
Research Systems Programmer * /usr/contributed Gatekeeper

---
Home Page: http://asg.web.cmu.edu/cyrus
Wiki/FAQ: http://cyruswiki.andrew.cmu.edu
List Archives/Info: http://asg.web.cmu.edu/cyrus/mailing-list.html