Skip to main content

Migrating from SpamAssassin to Rspamd

This guide explains how to migrate an existing SpamAssassin (SA) setup to Rspamd. It focuses on practical steps, key differences, and safe rollout.

At a glance​

  • Performance: Rspamd is typically much faster and more resource‑efficient than SA.
  • Rules: Many SA checks exist natively in Rspamd. Optional SA rules import is available via the spamassassin module.
  • Actions vs. score: Rspamd recommends actions (no action → greylist → add header → rewrite subject → reject) that your MTA applies.
  • Stats: SA Bayes data cannot be reused; you must retrain Rspamd statistics.

Migration checklist​

  1. Install Rspamd using official packages (/downloads).
  2. Integrate with your MTA: see MTA integration for Postfix, Exim, Sendmail and HTTP options.
  3. Configure Redis and statistics: see Statistics settings. Keep defaults unless you know what to change.
  4. Decide on SA rules migration:
    • Prefer native Rspamd modules and rules out‑of‑the‑box.
    • If you rely on custom SA rules, use the SpamAssassin rules module.
  5. Train statistics with your corpora (ham/spam) and/or enable autolearn.
  6. Stage rollout: mirror traffic, test thresholds, then switch production.

Key differences from SpamAssassin​

  • Engine: Rspamd is modular (Lua/C) with asynchronous I/O and Redis; SA is Perl‑based.
  • Verdict: Rspamd provides actions for the MTA; SA typically marks as spam/ham.
  • Policy: Rspamd combines multiple modules (DKIM, DMARC, SURBLs, fuzzy, Bayes, etc.) by default.
  • Config: HCL‑style files and Lua; native WebUI for monitoring and control.

Integrating Rspamd​

Use the guide at MTA integration. Recommended patterns:

  • Postfix/Sendmail: milter mode via Rspamd proxy worker.
  • Exim and HTTP‑capable MTAs: HTTP JSON integration.
  • LDA mode: deliver via rspamc --mime --exec to add headers and operate per‑user statistics.

Migrating rules (optional)​

If you need SA rules compatibility, enable the spamassassin module. Prefer native Rspamd features where possible for performance and quality.

Example minimal configuration (see full options at /modules/spamassassin):

spamassassin {
	# One or more SA rules files or globs
	ruleset = "/path/to/rules/*.cf";
	# Optionally include baseline SA rules
	base_ruleset = "/var/db/spamassassin/*/*.cf";
	# Optional tuning
	match_limit = 100k; # limit regex search window
}

Notes:

  • Rspamd optimizes SA rules; some plugins/functions are not supported. See the module doc for limits and performance notes.
  • Prefer keeping only the custom rules you truly need.

Statistics and training​

Rspamd uses a different statistical algorithm and storage; SA Bayes cannot be imported. Configure and train as follows:

  1. Ensure Redis is available and reachable by Rspamd.
  2. Use the default Bayes classifier in $CONFDIR/statistic.conf (or classifier-bayes.conf). See Statistics settings.
  3. Initial training from corpora (Maildir‑like directories):
rspamc learn_ham /path/to/ham
rspamc learn_spam /path/to/spam
  1. Optionally enable autolearn in classifier-bayes.conf:
autolearn {
	spam_threshold = 6.0;
	ham_threshold = -0.5;
	check_balance = true;
}
  1. Maintain statistics expiry with the bayes_expiry module.

Learning by email (optional)​

You can expose aliases to feed samples to rspamc:

learn-spam-<secret>: "| rspamc learn_spam"
learn-ham-<secret>:  "| rspamc learn_ham"

Use unpredictable aliases to prevent abuse and pollution.

Mapping SA concepts to Rspamd​

  • SA “spam” ≈ Rspamd “add header” by default. Consider your users’ workflows when choosing action thresholds.
  • SA network plugins like Razor/Pyzor/DCC are not used by Rspamd; similar functionality exists via native RBLs, fuzzy checks, reputation and URL checks.
  • Meta rules are supported; some SA plugins are unsupported—see module docs.

Rollout strategy​

  1. Shadow run: run Rspamd alongside SA and compare headers/reports.
  2. Tune actions: adjust thresholds to meet your FP/FN targets.
  3. Cutover: switch the MTA content filter to Rspamd, keep SA disabled but available for fallback.
  4. Monitor: use WebUI, logs and metrics; review BAYES_* rates and autolearn balance.

Troubleshooting and pitfalls​

  • Incompatible Bayes: do not attempt to import SA Bayes; retrain.
  • Excess SA rule load: importing the entire SA ruleset can slow Rspamd—migrate only necessary custom rules.
  • Missing Redis: statistics and some features rely on Redis; ensure connectivity and authentication.
  • Multi‑recipient mail with per‑user stats: prefer LDA mode to avoid attributing to the wrong user.

See also​