Upstream fixes
Adopted upstream. Poste.io 2.5.20 ships all three fixes (the learning scripts call
rspamc --rcpt, remove their temporary file and are no longer world-writable), so theupstream-fixesadd-on is gone from the Dockerfiles. On a base older than 2.5.20, build from release 1.2.0. The end-to-end tests still verify the learning behaviour on every build. The rest of this page documents the three problems as they were.
Three problems of the Poste.io image around spam learning: moving a message
to Junk (or out of it) is supposed to teach rspamd, and up to 2.5.19 it
did not (fixes 1 and 2); and every message learned stayed behind in /tmp
(fix 3). The first two were reported to Poste.io; releases up to 1.2.0 carried
a workaround, marked UPSTREAM-FIX, removed now that the base image fixes
them.
1. The learning scripts are world-writable
In the base image /etc/dovecot/sieve and the spam / ham pipe scripts are
0777. Dovecot refuses to execute a world-writable program
(executable ... is world-writable in /data/log/mail.log), so the learning
never happens.
Fix: chmod 0755 on the folder and the two scripts, 0644 on the .sieve
files.
2. Junk learning goes into a bucket nobody reads
Poste.io configures the rspamd Bayes classifier per domain (per_user
returns the recipient domain), so scanning reads the key RS<domain>. The Junk
scripts, however, call rspamc learn_spam / learn_ham without a
recipient: rspamd then learns into RS<address>, which no scan ever reads.
Fix: imapsieve passes the IMAP user to the scripts
(pipe :copy "spam" ["${IMAPUSER}"]) and the scripts call
rspamc --rcpt "$1" learn_spam, so the learning lands in the domain bucket.
The block ends with grep -q checks: if Poste.io changes the scripts, the
build fails instead of shipping a fix that no longer applies.
3. Every learned message stays in /tmp
The spam / ham scripts copy the message into a temporary file
(input_tmpfile=$(mktemp)) to give it to rspamc twice, and never remove it.
Each message moved into or out of Junk leaves a copy of itself in the
container's /tmp, readable by the mail user, until the container is
recreated.
Fix: right after the mktemp line the scripts get
so the file is removed when the script ends, also when it is stopped. As in
fix 2, a grep check fails the build if the scripts change.
Are they still needed?
No, from Poste.io 2.5.20 (checked on poste.io/mailserver:2.5.20 and
analogic/poste.io:2.5.20, scripts of 2026-10-21). On an older base, look at
the base image:
docker run --rm --entrypoint sh poste.io/mailserver:<VERSION> -c \
'ls -l /etc/dovecot/sieve; grep -n "pipe :copy" /etc/dovecot/sieve/*.sieve; grep -n "rspamc.*learn_\|tmpfile" /etc/dovecot/sieve/spam /etc/dovecot/sieve/ham'
(analogic/poste.io:<VERSION> for the open source edition.)
- scripts
-rwxr-xr-x→ remove fix 1; pipewith an argument andrspamcwith--rcpt(or an equivalent solution) → remove fix 2;- the scripts remove their temporary file (
rm, atrap, or no temporary file at all) → remove fix 3.
If the build fails on the grep checks instead, the scripts changed: read them
before adapting the block.
Checking that learning works
Move a spam message to Junk, then in the container:
The counter of your domain must increase. (rspamc stat sums all keys and
can be misleading.) Bayes starts classifying after 200 learned messages of each
kind.