Kubernetes (also known by its abbreviation K8s) is a production-grade orchestrating tool for containers. This article describes how to deploy `docker-mailserver` to K8s. K8s differs from Docker especially when it comes to separation of concerns: Whereas with Docker Compose, you can fit everything in one file, with K8s, the information is split. This may seem (too) verbose, but actually provides a clear structure with more features and scalability. We are going to have a look at how to deploy one instance of `docker-mailserver` to your cluster.
We assume basic knowledge about K8s from the reader. If you're not familiar with K8s, we highly recommend starting with something less complex, like Docker Compose.
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.
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.
We can also make use of user-provided configuration files, e.g. `user-patches.sh`, `postfix-accounts.cf` and more, to adjust `docker-mailserver` to our likings. We encourage you to have a look at [Kustomize][kustomize] for creating `ConfigMap`s 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.
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](#exposing-your-mail-server-to-the-outside-world).
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 K8s users will only want to have one instance anyway, and users that need high availability know how to do it anyways.
Last but not least, the `Deployment` becomes the most complex component. It instructs Kubernetes how to run the docker-mailserver container and how to apply your ConfigMaps and persisted storage. Additionally, we can set options to enforce runtime security here.
By now, the mailserver starts, but does not really work for long (or at all), because we're lacking certificates. You will need to choose yourself, which approach you'd want to go with. The [TLS][docs-tls] section provides you with an overview.
!!! attention "Sensitive Data"
For storing OpenDKIM keys, TLS certificates or any sort of sensitive data, you should be using `Secret`s. You can mount secrets like `ConfigMap`s and use them the same way.
## Exposing your Mailserver to the Outside World
The more difficult part with K8s is to expose a deployed mailserver to the outside world. K8s provides multiple ways for doing that; each has downsides and complexity. The major problem with exposing the mailserver to outside world in K8s is to [preserve the real client IP][k8s-service-source-ip]. The real client IP is required by the mailserver 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][docs-postfix] by dropping the line that states `check_policy_service unix:private/policyd-spf`.
The easiest approach was covered above, using `#!yaml 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 the mailserver as a [Service][k8s-network-service] with [external IPs][k8s-network-external-ip]. 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.
The [proxy pod][k8s-proxy-service] 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][k8s-nodes] you want to expose mailserver on.
One way to preserve the real client IP is to use `hostPort` and `hostNetwork: true`. This comes at the cost of availability; you can talk to the mailserver from outside world only via IPs of [Node][k8s-nodes] where mailserver is deployed.
This way is ideologically the same as [using a proxy pod](#proxy-port-to-service), but instead of a separate proxy pod, you configure your ingress to proxy TCP traffic to the mailserver pod using the PROXY protocol, which preserves the real client IP.
With an [NGINX ingress controller][k8s-nginx], set `externalTrafficPolicy: Local` for its service, and add the following to the TCP services config map (as described [here][k8s-nginx-expose]):
With [HAProxy][dockerhub-haproxy], the configuration should look similar to the above. If you know what it actually looks like, add an example here. :smiley:
