Running multiple sync_clients
Michael D. Sofka
sofkam at rpi.edu
Fri Nov 19 12:33:44 EST 2010
Bron Gondwana wrote:
> On Thu, Nov 18, 2010 at 12:58:10PM -0500, Michael D. Sofka wrote:
>> Is it safe to run multiple sync_clients? Is there an advantage to doing so?
>> I had to restart sync_client once. After later restarting cyrus two
>> sync_clients were running, and appeared to do well together. Still, I
>> stopped the old process out of caution.
> What version? In 2.3.x, the sync_server will obtain a lock, so you can
> only run one sync_server at a time. This means the sync_clients will
> take it in turns sending data.
2.3.16. This description fits what I was seeing.
> In 2.4, there is no global locking - just locks on each mailbox. So it's
> possible for two sync_clients to run at the same time. There is a
> possibility this will lead to the same data being sent twice. It won't
> cause corruption, but it will cause excess IO!
> Are you sure it's actually two sync_client instances? In every released
> version, sync_client forks immediately upon starting and then forks again
> to actually send data.
Yes. There were a total of 4 sync processes. Two parents, each with a
child worker. The logs showed two sync_client processes taking turns.
> Which is a problem because it opens the BDB environment once and closes it
> many times. So I've now changed it on my development branch to not fork
> the second time, but just keep it all in once process. We'll see if there
> are actually any memory leaks left!
Another question I have is what are /var/lib/imap/db/*? File says they
are Berkeley dbs, but I thought all databases were skiplist in this
release. I can't find the configuration options associated with this
db. I noticed them when looking at files open by sync_client. And not
that all cyrus process have these files open.
I also noticed that when the child process dies, it tends to take out
the parent. Shouldn't the parent catch the error and fork another
child? This has so far only happened twice, and in both cases appears
to be associated with xfering accounts from 2.2.12 to 2.3.16. In at
least one of the cases there was a back-end directory that has been
deleted by the most recent rsync, prior to migration. A reconstruct
Michael D. Sofka sofkam at rpi.edu
C&MT Sr. Systems Programmer, Email, HPC, TeX, Epistemology
Rensselaer Polytechnic Institute, Troy, NY. http://www.rpi.edu/~sofkam/
More information about the Info-cyrus