Kubernetes
Introduction
This article describes how to deploy DMS to Kubernetes. Please note that there is also a Helm chart available.
Requirements
We assume basic knowledge about Kubernetes from the reader. Moreover, we assume the reader to have a basic understanding of mail servers. Ideally, the reader has deployed DMS before in an easier setup with Docker (Compose).
About Support for Kubernetes
Please note that Kubernetes is not officially supported and we do not build images specifically designed for it. When opening an issue, please remember that only Docker & Docker Compose are officially supported.
This content is entirely community-supported. If you find errors, please open an issue and provide a PR.
Manifests
Configuration
We want to provide the basic configuration in the form of environment variables with a ConfigMap. Note that this is just an example configuration; tune the ConfigMap to your needs.
---
apiVersion: v1
kind: ConfigMap
metadata:
name: mailserver.environment
immutable: false
data:
TLS_LEVEL: modern
POSTSCREEN_ACTION: drop
OVERRIDE_HOSTNAME: mail.example.com
FAIL2BAN_BLOCKTYPE: drop
POSTMASTER_ADDRESS: postmaster@example.com
UPDATE_CHECK_INTERVAL: 10d
POSTFIX_INET_PROTOCOLS: ipv4
ONE_DIR: '1'
ENABLE_CLAMAV: '1'
ENABLE_POSTGREY: '0'
ENABLE_FAIL2BAN: '1'
AMAVIS_LOGLEVEL: '-1'
SPOOF_PROTECTION: '1'
MOVE_SPAM_TO_JUNK: '1'
ENABLE_UPDATE_CHECK: '1'
ENABLE_SPAMASSASSIN: '1'
SUPERVISOR_LOGLEVEL: warn
SPAMASSASSIN_SPAM_TO_INBOX: '1'
# here, we provide an example for the SSL configuration
SSL_TYPE: manual
SSL_CERT_PATH: /secrets/ssl/rsa/tls.crt
SSL_KEY_PATH: /secrets/ssl/rsa/tls.key
We can also make use of user-provided configuration files, e.g. user-patches.sh, postfix-accounts.cf and more, to adjust DMS to our likings. We encourage you to have a look at Kustomize for creating ConfigMaps from multiple files, but for now, we will provide a simple, hand-written example. This example is absolutely minimal and only goes to show what can be done.
---
apiVersion: v1
kind: ConfigMap
metadata:
name: mailserver.files
data:
postfix-accounts.cf: |
test@example.com|{SHA512-CRYPT}$6$someHashValueHere
other@example.com|{SHA512-CRYPT}$6$someOtherHashValueHere
Static Configuration
With the configuration shown above, you can not dynamically add accounts as the configuration file mounted into the mail server can not be written to.
Use persistent volumes for production deployments.
Persistence
Thereafter, we need persistence for our data. Make sure you have a storage provisioner and that you choose the correct storageClassName.
---
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
name: data
spec:
storageClassName: local-path
accessModes:
- ReadWriteOnce
resources:
requests:
storage: 25Gi
Service
A Service is required for getting the traffic to the pod itself. The service is somewhat crucial. Its configuration determines whether the original IP from the sender will be kept. More about this further down below.
The configuration you're seeing does keep the original IP, but you will not be able to scale this way. We have chosen to go this route in this case because we think most Kubernetes users will only want to have one instance.
---
apiVersion: v1
kind: Service
metadata:
name: mailserver
labels:
app: mailserver
spec:
type: LoadBalancer
selector:
app: mailserver
ports:
# Transfer
- name: transfer
port: 25
targetPort: transfer
protocol: TCP
# ESMTP with implicit TLS
- name: esmtp-implicit
port: 465
targetPort: esmtp-implicit
protocol: TCP
# ESMTP with explicit TLS (STARTTLS)
- name: esmtp-explicit
port: 587
targetPort: esmtp-explicit
protocol: TCP
# IMAPS with implicit TLS
- name: imap-implicit
port: 993
targetPort: imap-implicit
protocol: TCP
Deployments
Last but not least, the Deployment becomes the most complex component. It instructs Kubernetes how to run the DMS container and how to apply your ConfigMaps, persisted storage, etc. Additionally, we can set options to enforce runtime security here.
---
apiVersion: apps/v1
kind: Deployment
metadata:
name: mailserver
annotations:
ignore-check.kube-linter.io/run-as-non-root: >-
'mailserver' needs to run as root
ignore-check.kube-linter.io/privileged-ports: >-
'mailserver' needs privilegdes ports
ignore-check.kube-linter.io/no-read-only-root-fs: >-
There are too many files written to make The
root FS read-only
spec:
replicas: 1
selector:
matchLabels:
app: mailserver
template:
metadata:
labels:
app: mailserver
annotations:
container.apparmor.security.beta.kubernetes.io/mailserver: runtime/default
spec:
hostname: mail
containers:
- name: mailserver
image: ghcr.io/docker-mailserver/docker-mailserver:latest
imagePullPolicy: IfNotPresent
securityContext:
allowPrivilegeEscalation: false
readOnlyRootFilesystem: false
runAsUser: 0
runAsGroup: 0
runAsNonRoot: false
privileged: false
capabilities:
add:
# file permission capabilities
- CHOWN
- FOWNER
- MKNOD
- SETGID
- SETUID
- DAC_OVERRIDE
# network capabilities
- NET_ADMIN # needed for F2B
- NET_RAW # needed for F2B
- NET_BIND_SERVICE
# miscellaneous capabilities
- SYS_CHROOT
- KILL
drop: [ALL]
seccompProfile:
type: RuntimeDefault
# You want to tune this to your needs. If you disable ClamAV,
# you can use less RAM and CPU. This becomes important in
# case you're low on resources and Kubernetes refuses to
# schedule new pods.
resources:
limits:
memory: 4Gi
cpu: 1500m
requests:
memory: 2Gi
cpu: 600m
volumeMounts:
- name: files
subPath: postfix-accounts.cf
mountPath: /tmp/docker-mailserver/postfix-accounts.cf
readOnly: true
# PVCs
- name: data
mountPath: /var/mail
subPath: data
readOnly: false
- name: data
mountPath: /var/mail-state
subPath: state
readOnly: false
- name: data
mountPath: /var/log/mail
subPath: log
readOnly: false
# certificates
- name: certificates-rsa
mountPath: /secrets/ssl/rsa/
readOnly: true
# other
- name: tmp-files
mountPath: /tmp
readOnly: false
ports:
- name: transfer
containerPort: 25
protocol: TCP
- name: esmtp-implicit
containerPort: 465
protocol: TCP
- name: esmtp-explicit
containerPort: 587
- name: imap-implicit
containerPort: 993
protocol: TCP
envFrom:
- configMapRef:
name: mailserver.environment
restartPolicy: Always
volumes:
# configuration files
- name: files
configMap:
name: mailserver.files
# PVCs
- name: data
persistentVolumeClaim:
claimName: data
# certificates
- name: certificates-rsa
secret:
secretName: mail-tls-certificate-rsa
items:
- key: tls.key
path: tls.key
- key: tls.crt
path: tls.crt
# other
- name: tmp-files
emptyDir: {}
Certificates - An Example
In this example, we use cert-manager to supply RSA certificates. You can also supply RSA certificates as fallback certificates, which DMS supports out of the box with SSL_ALT_CERT_PATH and SSL_ALT_KEY_PATH, and provide ECDSA as the proper certificates.
---
apiVersion: cert-manager.io/v1
kind: Certificate
metadata:
name: mail-tls-certificate-rsa
spec:
secretName: mail-tls-certificate-rsa
isCA: false
privateKey:
algorithm: RSA
encoding: PKCS1
size: 2048
dnsNames: [mail.example.com]
issuerRef:
name: mail-issuer
kind: Issuer
Attention
You will need to have cert-manager configured. Especially the issue will need to be configured. Since we do not know how you want or need your certificates to be supplied, we do not provide more configuration here. The documentation for cert-manager is excellent.
Sensitive Data
Sensitive Data
For storing OpenDKIM keys, TLS certificates or any sort of sensitive data, you should be using Secrets. You can mount secrets like ConfigMaps and use them the same way.
The TLS docs page provides guidance when it comes to certificates and transport layer security. Always provide sensitive information vai Secrets.
Exposing your Mail Server to the Outside World
The more difficult part with Kubernetes is to expose a deployed DMS to the outside world. Kubernetes provides multiple ways for doing that; each has downsides and complexity. The major problem with exposing DMS to outside world in Kubernetes is to preserve the real client IP. The real client IP is required by DMS for performing IP-based SPF checks and spam checks. If you do not require SPF checks for incoming mails, you may disable them in your Postfix configuration by dropping the line that states: check_policy_service unix:private/policyd-spf.
The easiest approach was covered above, using externalTrafficPolicy: Local, which disables the service proxy, but makes the service local as well (which does not scale). This approach only works when you are given the correct (that is, a public and routable) IP address by a load balancer (like MetalLB). In this sense, the approach above is similar to the next example below. We want to provide you with a few alternatives too. But we also want to communicate the idea of another simple method: you could use a load-balancer without an external IP and DNAT the network traffic to the mail server. After all, this does not interfere with SPF checks because it keeps the origin IP address. If no dedicated external IP address is available, you could try the latter approach, if one is available, use the former.
External IPs Service
The simplest way is to expose DMS as a Service with external IPs. This is very similar to the approach taken above. Here, an external IP is given to the service directly by you. With the approach above, you tell your load-balancer to do this.
---
apiVersion: v1
kind: Service
metadata:
name: mailserver
labels:
app: mailserver
spec:
selector:
app: mailserver
ports:
- name: smtp
port: 25
targetPort: smtp
# ...
externalIPs:
- 80.11.12.10
This approach
- does not preserve the real client IP, so SPF check of incoming mail will fail.
- requires you to specify the exposed IPs explicitly.
Proxy port to Service
The proxy pod helps to avoid the necessity of specifying external IPs explicitly. This comes at the cost of complexity; you must deploy a proxy pod on each Node you want to expose DMS on.
This approach
- does not preserve the real client IP, so SPF check of incoming mail will fail.
Bind to concrete Node and use host network
One way to preserve the real client IP is to use hostPort and hostNetwork: true. This comes at the cost of availability; you can reach DMS from the outside world only via IPs of Node where DMS is deployed.
---
apiVersion: extensions/v1beta1
kind: Deployment
metadata:
name: mailserver
# ...
spec:
hostNetwork: true
# ...
containers:
# ...
ports:
- name: smtp
containerPort: 25
hostPort: 25
- name: smtp-auth
containerPort: 587
hostPort: 587
- name: imap-secure
containerPort: 993
hostPort: 993
# ...
With this approach,
- it is not possible to access DMS via other cluster Nodes, only via the Node DMS was deployed at.
- every Port within the Container is exposed on the Host side.
Proxy Port to Service via PROXY Protocol
This way is ideologically the same as using a proxy pod, but instead of a separate proxy pod, you configure your ingress to proxy TCP traffic to the DMS pod using the PROXY protocol, which preserves the real client IP.
Configure your Ingress
With an NGINX ingress controller, set externalTrafficPolicy: Local for its service, and add the following to the TCP services config map (as described here):
25: "mailserver/mailserver:25::PROXY"
465: "mailserver/mailserver:465::PROXY"
587: "mailserver/mailserver:587::PROXY"
993: "mailserver/mailserver:993::PROXY"
HAProxy
With HAProxy, the configuration should look similar to the above. If you know what it actually looks like, add an example here. 
Configure the Mailserver
Then, configure both Postfix and Dovecot to expect the PROXY protocol:
HAProxy Example
kind: ConfigMap
apiVersion: v1
metadata:
name: mailserver.config
labels:
app: mailserver
data:
postfix-main.cf: |
postscreen_upstream_proxy_protocol = haproxy
postfix-master.cf: |
smtp/inet/postscreen_upstream_proxy_protocol=haproxy
submission/inet/smtpd_upstream_proxy_protocol=haproxy
submissions/inet/smtpd_upstream_proxy_protocol=haproxy
dovecot.cf: |
# Assuming your ingress controller is bound to 10.0.0.0/8
haproxy_trusted_networks = 10.0.0.0/8, 127.0.0.0/8
service imap-login {
inet_listener imap {
haproxy = yes
}
inet_listener imaps {
haproxy = yes
}
}
# ...
---
kind: Deployment
apiVersion: extensions/v1beta1
metadata:
name: mailserver
spec:
template:
spec:
containers:
- name: docker-mailserver
volumeMounts:
- name: config
subPath: postfix-main.cf
mountPath: /tmp/docker-mailserver/postfix-main.cf
readOnly: true
- name: config
subPath: postfix-master.cf
mountPath: /tmp/docker-mailserver/postfix-master.cf
readOnly: true
- name: config
subPath: dovecot.cf
mountPath: /tmp/docker-mailserver/dovecot.cf
readOnly: true
With this approach,
- it is not possible to access DMS via cluster-DNS, as the PROXY protocol is required for incoming connections.