Configure Spam Filtering#
Improve spam detection by configuring rspamd overrides and DNS-based blocklists.
Problem#
The default rspamd configuration provides basic spam filtering, but does not include:
DNS-based Real-time Blackhole Lists (RBLs) like Spamhaus
Remote fuzzy hash checking against known spam databases
Optimized scoring thresholds
In Kubernetes environments, RBL lookups fail because kube-dns forwards queries to public resolvers (e.g. Cloudflare, Google) which are blocked by RBL providers like Spamhaus.
Solution Overview#
cdk8s-mailu addresses this with:
Unbound DNS sidecar - A recursive DNS resolver running alongside rspamd that sends queries directly to authoritative DNS servers
Rspamd configuration overrides - Stricter thresholds, RBL lists, and remote fuzzy servers via ConfigMap
How It Works#
The RspamdConstruct automatically deploys:
An Unbound container as a sidecar in the rspamd pod
ConfigMaps with rspamd overrides mounted at
/overrides/A DNS override (
options.inc) pointing rspamd to127.0.0.1:53(Unbound)
graph LR
A[Incoming Mail] --> B[Postfix]
B -->|milter| C[rspamd]
C -->|DNS queries| D[Unbound Sidecar]
D -->|recursive| E[Spamhaus]
D -->|recursive| F[Barracuda]
D -->|forward| H[kube-dns]
H -->|.cluster.local| I[K8s Services]
DNS Configuration#
The Unbound sidecar can be configured via the dns config option:
dns:
kubeDnsIp: "10.43.0.10" # kube-dns ClusterIP (K3S default)
unboundImage: "mvance/unbound:1.22.0" # Pinned image version
Default Behavior#
Without explicit dns configuration, Unbound uses:
10.43.0.10as kube-dns for.cluster.localforwardingRecursive resolution for all other queries (including RBL lookups)
QNAME minimization disabled (required by Spamhaus)
Rspamd Overrides#
The following overrides are applied automatically:
Scoring Thresholds (actions.conf)#
reject = 12 # Reject messages with score >= 12 (default: 15)
add_header = 5 # Add X-Spam header at score >= 5 (default: 6)
greylist = 3 # Greylist at score >= 3 (default: 4)
Additional RBL (rbl.conf, groups.conf)#
Adds the Barracuda RBL as RBL_BARRACUDA with a score of 2.0.
Spamhaus ZEN and DBL and the other rspamd default RBLs stay active without an override.
Fuzzy scores (metrics.conf)#
Restores the scores of the rspamd.com fuzzy symbols (FUZZY_DENIED 12, FUZZY_PROB 5, FUZZY_WHITE -2.1), which Mailu’s own metrics.conf drops.
Spam diagnosis header (milter_headers.conf)#
Adds X-Spamd-Result to every scanned message.
Read it in the headers of a missed spam message to see which symbols fired.
Bayes learning#
The Bayes classifier uses Mailu’s defaults. It stores its data in Redis, so give Redis persistent storage. Bayes adds no score until it has learned 200 spam and 200 ham messages. Moving a message into the Junk folder learns it as spam.
Warning
Do not add a classifier-bayes.conf or fuzzy_check.conf override.
Mailu merges these with its own files, turns repeated keys into arrays, and breaks Bayes autolearning.
See Rspamd Architecture for details.
Troubleshooting#
RBL Lookups Not Working#
Verify Unbound is running and resolving correctly:
# Check both containers are running
kubectl get pods -n mailu -l app.kubernetes.io/component=rspamd
# Test RBL resolution via Unbound
kubectl exec -n mailu deploy/rspamd -c unbound -- \
drill @127.0.0.1 2.0.0.127.zen.spamhaus.org
Too Many False Positives#
If legitimate mail is being flagged:
Increase the
rejectthreshold (e.g. to 15)Lower the
greylistthreshold sensitivityCheck rspamd web UI for which symbols triggered
Kubernetes DNS Still Working#
Verify that .cluster.local resolution works through Unbound:
kubectl exec -n mailu deploy/rspamd -c rspamd -- \
nslookup redis.mailu.svc.cluster.local 127.0.0.1
See Also#
Rspamd Architecture - How spam filtering works
Component Specifications - Rspamd technical details
Scale Resources - Adjust resources for spam filtering load