I think there's a good chance that upgrading both will fix it. The bug already existed in old versions, it just wasn't normally triggered. Since v2.3.8 this situation is triggered on one dsync side, so the v2.3.9 fix needs to be on the other side.

On 5. Dec 2019, at 8.34, Piper Andreas via dovecot <dovecot@dovecot.org> wrote:

Hello,

upgrading to 2.3.9 unfortunately does *not* solve this issue:

I upgraded one of my replicators from 2.3.7.2 to 2.3.9 and after some seconds replication stopped. The other replicator remained with 2.3.7.2. After downgrading to 2.3.7.2 replication is again working fine.

I did not try to upgrade both replicators up to now, as this is a live production system. Is there a chance, that upgrading both replicators will solve the problem?

The machines are running Ubuntu 18.04

Any help is appreciated.

Thanks,
Andreas

Am 18.10.19 um 13:52 schrieb Carsten Rosenberg via dovecot:
Hi,
some of our customers have discovered a replication issue after
upgraded from 2.3.7.2 to 2.3.8.
Running 2.3.8 several replication connections are hanging until defined
timeout. So after some seconds there are $replication_max_conns hanging
connections.
Other replications are running fast and successful.
Also running a doveadm sync tcp:... is working fine for all users.
I can't see exactly, but I haven't seen mailboxes timeouting again and
again. So I would assume it's not related to the mailbox.
From the logs:
server1:
Oct 16 08:29:25 server1 dovecot[5715]:
dsync-local(username1@domain.com)<FXnVDW22pl0tGAAA1cwDxA>: Error:
dsync(172.16.0.1): I/O has stalled, no activity for 600 seconds (version
not received)
Oct 16 08:29:25 server1 dovecot[5715]:
dsync-local(username1@domain.com)<FXnVDW22pl0tGAAA1cwDxA>: Error:
Timeout during state=master_recv_handshake
server2:
Oct 16 08:29:25 server2 dovecot[8113]: doveadm: Error: read(server1)
failed: EOF (last sent=handshake, last recv=handshake)
There aren't any additional logs regarding the replication.
I have tried increasing vsz_limit or reducing replication_max_conns.
Nothing changed.
--
Both customers have 10k+ users. Currently I couldn't reproduce this on
smaller test systems.
Both installation were downgraded to 2.3.7.2 to fix the issue for now
--
I've attached a tcpdump showing the client showing the client stops
sending any data after the mailbox_guid table headers.
Any idea what could be wrong here or the debug this issue?
Thanks.
Carsten Rosenberg


-- 
________________________________________________________________________
Dr. Andreas Piper, Hochschulrechenzentrum der Philipps-Univ. Marburg
         Hans-Meerwein-Straße 6, 35032 Marburg, Germany
Phone: +49 6421 28-23521  Fax: -26994  E-Mail: piper@HRZ.Uni-Marburg.DE