I am having trouble scaling some regular dovecot cleanup operations on our servers. On a daily basis, I'm wanting to do this on each server, which contains its own isolated set of user storage:

/usr/bin/doveadm expunge -A mailbox Trash* savedbefore 21d /usr/bin/doveadm expunge -A mailbox Spam savedbefore 7d /usr/bin/doveadm expunge -A mailbox Sent savedbefore 120d

but these are a very expensive operations. For example, just doing the Spam expunge takes 30 minutes (or more, depending on the load) of heavy disk operations, on each machine it is run on. There are approximately 20k users on each machine.

It seems like it does not use the iterate query, but rather it looks into the database at the expires table and rather it iterates over every user mentioned there. This is a problem because I've got multiple dovecot machines with different sets of users on them, using the same table, so that means that its doing stat() calls on each system for every user mentioned there and failing to find the user on the filesystem (because the user is on another system). In these cases it spits out an error: doveadm(user): Info: User no longer exists, skipping

For users that do exist on this system, it seems to do something on the order of 15 stat() calls, at minimum.

What are some things I can do to make this less of an expensive process? If I had a shared storage system that each machine used, this would cut down on the resource waste because I'd only need to run the query once, but unfortuntely, that isn't how these systems were designed.

thanks for any ideas, tips etc. micah