summaryrefslogtreecommitdiffstats
path: root/roles/MX/templates/etc/postfix/virtual/transport.j2
Commit message (Collapse)AuthorAgeFiles
* Explain why we use static transport maps and custom subdomains.Guilhem Moulin2015-06-071
|
* Use $virtual_alias_domains not $virtual_mailbox_domains.Guilhem Moulin2015-06-071
| | | | | | | | | | | | | | | | | | | | | | | | | Quoting postconf(5): smtpd_reject_unlisted_recipient (default: yes) Request that the Postfix SMTP server rejects mail for unknown recipient addresses, even when no explicit reject_unlisted_recipient access restriction is specified. This prevents the Postfix queue from filling up with undeliverable MAILER-DAEMON messages. An address is always considered "known" when it matches a virtual(5) alias or a canonical(5) mapping. […] * The recipient domain matches $virtual_alias_domains but the recipient is not listed in $virtual_alias_maps. * The recipient domain matches $virtual_mailbox_domains but the recipient is not listed in $virtual_mailbox_maps, and $virtual_mailbox_maps is not null. Since we alias everything under special, "invalid", domains (mda.f.o and mailman.f.o), our $virtual_mailbox_maps was null, which led to reject_unlisted_recipient not being triggered for say, "noone@fripost.org". However, replacing $virtual_mailbox_domains with $virtual_alias_domains fits into the second point above.
* Perform the alias resolution and address validation solely on the MX:es.Guilhem Moulin2015-06-071
| | | | | We can therefore spare some lookups on the MDA, and use static:all instead.
* Replace IPSec tunnels by app-level ephemeral TLS sessions.Guilhem Moulin2015-06-071
| | | | | For some reason giraff doesn't like IPSec. App-level TLS sessions are less efficient, but thanks to ansible it still scales well.
* Assume a DNS entry for each role.Guilhem Moulin2015-06-071
| | | | | | E.g., ldap.fripost.org, ntp.fripost.org, etc. (Ideally the DNS zone would be provisioned by ansible, too.) It's a bit unclear how to index the subdomains (mx{1,2,3}, etc), though.
* Decongestion potential bottlenecks on trivial_rewrite(8).Guilhem Moulin2015-06-071
Which might be caused by slow LDAP lookups in transport_maps. Instead, we alias each addresses for which we want a custom transport to a dedicated "dummy" domain, and use a static (CDB) transport_maps to map said domains to their transport; the receiver can then use canonical(8) to restore the original envelope recipient. Since the alias resolution is performed by cleanup(8), which can run in parallel with other instances, it should decongestion bottlenecks under heavy loads. So far only the MX:es have been decongestioned. The list manager and the MDA should be treated as well.