[2.3.8] possible replication issue

[KOCIK Fabien (Acoss)]

Hi,

I'm dealing for few days with replication issues between two dovecot containerized instances in Kubernetes environment.
After some search, I found this thread that reports exactly the same symptoms I encountered : https://dovecot.org/pipermail/dovecot/2019-October/117353.html
I can confirm that downgrade of dovecot to 2.3.7.2 fixes the issue.

Testing usage of -T parameter with a very low value (10s) on 2.3.8 version is reducing perceptible impact of the issue (mails are replicated within the following five minutes).

I'm not expert in dovecot architecture but I tried to put an eye in the code and I'm wondering if this commit is not creating a potential phantom stalled I/O timeout ?
https://github.com/dovecot/core/commit/ec817bb2185bb21b34ba6bdd83b32af16dd0b4ad#diff-f0ef7f961c147b56c28b246b06eb5eb6

In fact, I was unable to see when the counter incremented by io_set_pending is decremented ?

Hoping that this information can help...

Regards.
Fabien.
+--------+
For technical details :
Container images are based on Arch Linux distro image.
Here is the configuration of Dovecot server :
dovecot -n
# 2.3.7.2 (3c910f64b): /etc/dovecot/dovecot.conf
# OS: Linux 4.19.66-coreos x86_64  ext4
# Hostname: mailstore-0.mailstore.piafe.svc.cluster.local
disable_plaintext_auth = no
doveadm_password = # hidden, use -P to show it
doveadm_port = 12345
login_greeting = PIAFE Mail Server Ready
mail_location = mdbox:/var/mail/%Ld/%Ln
mail_plugins = " notify replication quota"
namespace inbox {
  hidden = no
  inbox = yes
  list = yes
  location =
  mailbox Brouillons {
    auto = subscribe
    special_use = \Drafts
  }
  mailbox Corbeille {
    auto = subscribe
    special_use = \Trash
  }
  mailbox "Courrier indésirable" {
    auto = subscribe
    special_use = \Junk
  }
  mailbox Envoyés {
    auto = subscribe
    special_use = \Sent
  }
  prefix =
  separator = /
  subscriptions = yes
  type = private
}
passdb {
  args = /etc/dovecot/ldap.conf.ext
  driver = ldap
}
plugin {
  mail_replica = tcp:mailstore-1.mailstore.piafe.svc.cluster.local
  quota = dirsize:User quota
  quota_grace = 10%%
  quota_max_mail_size = 20M
  quota_rule = *:storage=1G
  quota_rule2 = Trash:storage=+200M
  quota_status_access = DUNNO
  quota_status_nouser = DUNNO
  quota_status_overquota = 552 5.5.2 Mailbox is full
}
postmaster_address = postmaster at piafe.recouv.fr
replication_dsync_parameters = -d -N -l 30 -U -T 10
service aggregator {
  fifo_listener replication-notify-fifo {
    user = mailer
  }
  unix_listener replication-notify {
    user = mailer
  }
}
service doveadm {
  inet_listener {
    port = 12345
  }
  inet_listener http {
    port = 8080
  }
}
service lmtp {
  inet_listener lmtp {
    address = *
    port = 24
  }
}
service replicator {
  process_min_avail = 1
  unix_listener replicator-doveadm {
    group = mailer
    mode = 0660
  }
}
ssl = no
userdb {
  args = /etc/dovecot/ldap.conf.ext
  default_fields = uid=500 gid=500 home=/var/mail/%Ld/%Ln
  driver = ldap
}
protocol imap {
  mail_plugins = " notify replication quota quota"
}

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://dovecot.org/pipermail/dovecot/attachments/20191108/1b06576f/attachment.html>