Lockers keeps going higher...

[Bron Gondwana]

On Wed, Jul 15, 2009 at 04:23:12PM -0500, Derek Chen-Becker wrote:
> Simon Matter wrote:
> > I'm quite sure in your case the problem is with duplicate_db (deliver.db).
> > Otherwise you wouldn't see those locker errors because they are not coming
> > from skiplist. So I suggest to convert deliver.db the same way to
> > skiplist.
> 
> This thread seems to indicate that using skiplist for deliver.db isn't
> optimal:
> 
> http://lists.andrew.cmu.edu/pipermail/info-cyrus/2005-September/019694.html

It may be a little bit faster, but we haven't noticed any difference 
when we switched to using skiplist about a month ago.

In the past when the skiplist backend was as buggy as whatever, it
probably would have caused some pain - but it's had a LOT of polishing
done on it.  There might still be bugs of course, but it's had stacks
of testing.
 
> My question would be, if I can only ever have 10 concurrent lmtpd
> processes running (based on my postfix config), how can I possibly have
> 1600+ lockers on the database? Is there some code path that isn't
> releasing the lock on deliver.db?

I really don't know, I haven't dug deeply into the Berkeley backend.
I suspect Cyrus must be doing _something_ wrong, becase it seems to 
have more BDB issues than any other software that uses it (note:
"wrong" could just mean exercising bits of the API that other
software doesn't, and that aren't tested as well.)

Bron.