Using Dovecot for a Geo-replicated World-wide Email Service

[Lennart Oldenburg]

Hello everyone,

my colleagues and I from Technische Universität Berlin are currently
looking into the best way to setup a Dovecot-powered email service that
is simultaneously accessible world-wide and still provides short
response times by redirecting users to the geographically-closest
cluster. We imagine the setup to make use of multiple data center
regions in public clouds around the world, e.g. in Amazon Web Services
(AWS) or Google Compute Platform (GCP). At the same time, state is
considered to be shared, i.e. changes made on file system in any local
cluster will have to be synchronized with all other clusters around the
globe (mailbox replication) and mailboxes might be manipulated from any
data center in the world, potentially at the same time. Naturally,
problems arise. We have multiple questions and would be very interested
in the best way to approach them based on your experiences and
recommendations:

1) Which path should we take replication-wise? Please keep in mind that
we are targeting replication between more than 2 clusters.
Traditionally, in a rather small local setup with mostly reliable
network, a distributed file system such as NFS or GlusterFS might be the
best option. Unfortunately, though, when scaling to more than 2 data
centers geographically-distributed around the globe, latency becomes
severe and distributed file systems might not be the way to go anymore.
What is your recommendation for replicating state?

2) We are mostly interested in providing short response times to client
requests while maintaining global state consistency (that is, avoiding
conflicts in state and not losing any user data). What are your
recommendations in that direction? Any particular performance
optimizations we could apply to Dovecot?

3) Is there a recommended way to measure and monitor Dovecot's service
quality? For example, is there a way to export "live" metrics of a
running Dovecot deployment to something such as Prometheus
(https://prometheus.io)?

4) Maildir as the way to represent mailboxes and deliver emails on file
system is kind of a given. I wonder, though, how well is dbox
replicatable? What happens when dbox files are synchronized (e.g. via
dsync) and a conflict arises? Does dbox offer a performance improvement
over Maildir?

5) We perceive Dovecot to come with many optimizations to speed up
"read" IMAP commands, i.e. non-state-changing ones. Is this perception
correct? Or is Dovecot able to use its various index and log files for
fast "write" IMAP commands (e.g. CREATE, DELETE, RENAME, APPEND, STORE,
...) as well?

6) Would DovecotPRO by Dovecot Oy offer any advantage to state
replication? For example, could the Object Storage Plugin be used to
replicate mailboxes through some object store service offered world-wide?

7) Is the setup we have in mind based on Dovecot a feasible solution for
world-wide email comparable to services like Gmail? If not, what would
you change or not do at all?

Please excuse if some questions might be of rather basic nature, we are
not at all Dovecot experts. I tried to gather as much information on
that topic as possible from the wiki and sources found online but I
might have missed an important page. In that case, please point me
towards existing resources on the questions raised above.


Thanks in advance and kind regards,
Lennart Oldenburg
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0x4A70E1C7.asc
Type: application/pgp-keys
Size: 6068 bytes
Desc: not available
URL: <http://dovecot.org/pipermail/dovecot/attachments/20170522/f1a5be8c/attachment.bin>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: OpenPGP digital signature
URL: <http://dovecot.org/pipermail/dovecot/attachments/20170522/f1a5be8c/attachment.sig>