i decided to make a website. a static one. this one. with Hugo. the main reason i have for needing a website is as a vanity project, so i have some stuff to host in a Kubernetes cluster i’m running. the k8s cluster is also a vanity project.

because i don’t like software, i wanted a way to deploy my site that doesn’t involve much of it. this post is about that.

Getting Started

i built my site by following the straight-forward Getting Started guide in the Hugo documentation.

i did hugo new site estradiol.cloud. and then cd estradiol.cloud; git init. and then i picked a ridiculous theme “inspired by terminal ricing aesthetics”, installing it like git submodule add https://github.com/joeroe/risotto.git themes/risotto; echo "theme = 'risotto'" >> hugo.toml.¹

at this point, my website is basically finished (i also changed the title in hugo.toml). i probably won’t be putting anything on it, so there’s no point fiddling with other details.

about deployment, the Hugo guide’s Basic Usage page has this to offer:

Most of our users deploy their sites using a CI/CD workflow, where a push¹ to their GitHub or GitLab repository triggers a build and deployment. Popular providers include AWS Amplify, CloudCannon, Cloudflare Pages, GitHub Pages, GitLab Pages, and Netlify.

1. The Git repository contains the entire project directory, typically excluding the public directory because the site is built after the push.

importantly, you can’t make a post about deploying this way. everyone deploys this way. if i deploy this way, this site will have no content.

this approach also involves a build system somewhere that can run Hugo to compile the code and assets and push them onto my host. i definitely already need Hugo installed on my laptop if i’m going to post anything.² so now i’m running Hugo in two places. there’s surely going to be other complex nonsense like webhooks involved.

and hang on. let’s look at this again:

1. The Git repository contains the entire project directory, typically excluding the public directory because the site is built after the push.

you’re telling me i’m going to build a nice static site and not check the actual content into version control? couldn’t be me.

Getting Static

suppose i instead check my content in exactly as i intend to serve it? then i could shell into my server box, pull the site, and nifty-galifty! isn’t this the way it has always been done?

my problem is that i don’t have a server box. i have a container orchestration system. there are several upsides to this³ but it means that somehow my generated content needs to end up in a container. because Pods are ephemeral and i’d like to run my site with horizontal scalability⁴, i don’t want my container to retain runtime state across restarts or replicas.

i could run a little pipeline that builds a container image wrapping my content and pushes it to a registry. when i deploy, the cluster pulls the image, content and all. all ready to go. but now i’ve got software again: build stages and webhooks and, to make matters worse, now i’m hosting and versioning container images.

i don’t want any of this. i just want to put some HTML and static assets behind a web server.

instead, i’d like to deploy a popular container image from a public registry and deliver my content to it continuously.

a minimal setup to achieve this might look like:

• a Pod with:
  • an nginx container to serve the content;
  • a git-pull sidecar that loops, pulling the content;
  • an initContainer to do the initial checkout;
  • an emptyDir volume to share between the containers.
• a ConfigMap to store the nginx config.

when a new Pod comes up, the initContainer mounts the emptyDir at /www and clones the repository into it. i use git sparse-checkout to avoid pulling repository contents i don’t want to serve out:

# git-clone command
git clone https://code.estradiol.cloud/tamsin/estradiol.cloud.git --no-checkout --branch trunk /tmp/www;
cd /tmp/www;
git sparse-checkout init --cone;
git sparse-checkout set public;
git checkout;
shopt -s dotglob
mv /tmp/www/* /www

for the sidecar, i script up a git pull loop:

# git-pull command
while true; do
  cd /www && git -c safe.directory=/www pull origin trunk
  sleep 60
done

and i create a ConfigMap with a server block to configure nginx to use Hugo’s public/ as root:

# ConfigMap; data: default.conf
server {
  listen 80;
  location / {
    root   /www/public;
    index  index.html;
  }
}

the rest of this is pretty much boilerplate:

# estradiol-cloud.yaml
apiVersion: v1
kind: ConfigMap
metadata:
  labels:
    app.kubernetes.io/instance: estradiol-cloud
    app.kubernetes.io/name: nginx
  name: nginx-server-block
data:
  default.conf: |-
    server {
      listen 80;
      location / {
        root   /www/public;
        index  index.html;
      }
    }
---
apiVersion: v1
kind: Pod
metadata:
  name: nginx
  labels:
    app.kubernetes.io/instance: estradiol-cloud
    app.kubernetes.io/name: nginx
spec:
  containers:
  - name: nginx
    image: nginx:1.25.4
    ports:
    - containerPort: 80
    volumeMounts:
    - mountPath: /www
      name: www
    - mountPath: /etc/nginx/conf.d
      name: nginx-server-block
  - name: git-pull
    image: bitnami/git
    command:
    - /bin/bash
    - -ec
    - |
      while true; do
        cd /www && git -c safe.directory=/www pull origin trunk
        sleep 60
      done
    volumeMounts:
    - mountPath: /www
      name: www
  initContainers:
  - name: git-clone
    image: bitnami/git
    command:
    - /bin/bash
    - -c
    - |
      shopt -s dotglob
      git clone https://code.estradiol.cloud/tamsin/estradiol.cloud.git --no-checkout --branch trunk /tmp/www;
      cd /tmp/www;
      git sparse-checkout init --cone;
      git sparse-checkout set public;
      git checkout;
      mv /tmp/www/* /www
    volumeMounts:
    - mountPath: /www
      name: www
  volumes:
  - name: www
    emptyDir: {}
  - name: nginx-server-block
    configMap:
      name: nginx-server-block

my Hugo workflow now looks like:

1. make changes to source;
2. run hugo --gc --minify;⁵
3. git commit & push.

my git pull control loop takes things over from here and i’m on easy street.

Getting Web

this is going great! my Pod is running. it’s serving out my code. i get Continuous Deployment™ for the low price of 11 lines bash. i mean… no one can actually browse to my website⁶ but that will be an easy fix, right? yes. networking is always the easy part.

first, i need a Service. this gives me a proxy to my several replicas⁷ and in-cluster service discovery.

# service.yaml
apiVersion: v1
kind: Service
metadata:
  labels:
    app.kubernetes.io/instance: estradiol-cloud
    app.kubernetes.io/name: nginx
  name: nginx
spec:
  type: ClusterIP
  selector:
    app.kubernetes.io/instance: estradiol-cloud
    app.kubernetes.io/name: nginx
  ports:
  - name: http
    port: 80
    protocol: TCP
    targetPort: http

next, i need an Ingress to handle traffic inbound to the cluster and direct it to the Service:

# ingress.yaml
apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  labels:
    app.kubernetes.io/instance: estradiol-cloud
    app.kubernetes.io/name: nginx
  name: nginx
spec:
  rules:
  - host: estradiol.cloud
    http:
      paths:
      - backend:
          service:
            name: nginx
            port:
              name: http
        path: /
        pathType: Prefix

this part expresses a routing rule: traffic reaching the cluster via estradiol.cloud should go to my Service, and then to one of its backend Pods. to actually apply this rule, i need an ingress controller. mine is ingress-nginx.

when i deployed controller in my cluster, it created some more nginx Pods. these update their configuration dynamically based on the rules in my Ingress resource(s). the controller also creates a Service of type LoadBalancer, which magically creates a load balancer appliance in my cloud provider. off-screen, i can point DNS to that appliance to finish the setup.

you can tell it’s working by looking at your browser bar.

as this has come together, i’ve gotten increasingly anxious about how much YAML i’ve had to write. this is a problem because YAML is software and, as established, i’m hoping not to have much of that. it’s also annoying that most of this YAML really is just boilerplate.

conveniently, Bitnami maintains a Helm Chart that templates out all the boilerplate and does exactly what i’ve just been doing.⁸ i can replace all my YAML with a call out to this chart and a few lines of configuration, assuming i have helm client installed:

# values.yaml
cloneStaticSiteFromGit:
  enabled: true
  repository: "https://code.estradiol.cloud/tamsin/estradiol.cloud.git"
  branch: trunk
  gitClone:
    command:
    - /bin/bash
    - -ec
    - |
      [[ -f "/opt/bitnami/scripts/git/entrypoint.sh" ]] && source "/opt/bitnami/scripts/git/entrypoint.sh"
      git clone {{ .Values.cloneStaticSiteFromGit.repository }} --no-checkout --branch {{ .Values.cloneStaticSiteFromGit.branch }} /tmp/app
      [[ "$?" -eq 0 ]] && cd /tmp/app && git sparse-checkout init --cone && git sparse-checkout set public && git checkout && shopt -s dotglob && rm -rf /app/* && mv /tmp/app/* /app/
ingress:
  enabled: true
  hostname: estradiol.cloud
  ingressClassName: nginx
  tls: true
  annotations: {
    cert-manager.io/cluster-issuer: letsencrypt-prod
  }
serverBlock: |-
  server {
    listen 8080;
    root /app/public;
    index index.html;
  }
service:
  type: ClusterIP

configuration for the git-clone script and our custom server block are added via values.yaml. the git-pull loop configured by the chart works as-is. by using the chart, we get a few other nicities. for instance, my Pods are now managed by a Deployment.⁹ this will make my grand scale-out plans a breeze.

Getting Flux’d

by now, i’m riding high. my whole setup is my static site code and <30 lines of YAML.

i do have a bunch of stuff deployed into my cluster, and none of this is very reproducible without all of that. my workflow has also expanded to:

1. for routine site deploys:
   1. make changes to source;
   2. run hugo --gc --minify;⁵
   3. git commit & push.
2. to update nginx, the chart version, or change config:
   1. make changes to values.yaml
   2. helm upgrade

i could do without the extra helm client dependency on my laptop. i’m also pretty git push-pilled, and i really want the solution to all my problems to take the now familiar shape: put a control loop in my cluster and push to a git repository.

enter flux.

with flux, i decide on a repository (and maybe a path within it) to act as a source for my Kubernetes YAML. i go through a short bootstrap process which installs the flux controllers and add them to repository. to make a change to a resource in my cluster, i edit the YAML and push to the repository. flux listens and applies the changes.

flux supports Helm deploys, so i can get that helm client off my laptop. i can also use it to manage my ingress controller, cert-manager, flux itself and whatever other infrastructural junk i may end up needing.

to move my web stack into flux, i create a HelmRepository resource for the bitnami Helm charts:

# bitnami-helm.yaml
apiVersion: source.toolkit.fluxcd.io/v1beta2
kind: HelmRepository
metadata:
  name: bitnami
  namespace: default
spec:
  url: https://charts.bitnami.com/bitnami

and add a HelmRelease pointing to the repository/chart version and containing my values.yaml:

apiVersion: helm.toolkit.fluxcd.io/v2beta1
kind: HelmRelease
metadata:
  name: web
  namespace: estradiol-cloud
spec:
  interval: 5m
  chart:
    spec:
      chart: nginx
      version: '15.12.2'
      sourceRef:
        kind: HelmRepository
        name: bitnami
        namespace: default
      interval: 1m
  values:
    cloneStaticSiteFromGit:
      enabled: true
      repository: "https://code.estradiol.cloud/tamsin/estradiol.cloud.git"
      branch: trunk
      gitClone:
        command:
          - /bin/bash
          - -ec
          - |
            [[ -f "/opt/bitnami/scripts/git/entrypoint.sh" ]] && source "/opt/bitnami/scripts/git/entrypoint.sh"
            git clone {{ .Values.cloneStaticSiteFromGit.repository }} --no-checkout --branch {{ .Values.cloneStaticSiteFromGit.branch }} /tmp/app
            [[ "$?" -eq 0 ]] && cd /tmp/app && git sparse-checkout init --cone && git sparse-checkout set public && git checkout && shopt -s dotglob && rm -rf /app/* && mv /tmp/app/* /app/
    ingress:
      enabled: true
      hostname: estradiol.cloud
      ingressClassName: nginx
      tls: true
      annotations: {
        cert-manager.io/cluster-issuer: letsencrypt-prod
      }
    serverBlock: |-
      server {
        listen 8080;
        root /app/public;
        index index.html;
      }
    service:
      type: ClusterIP

when i push these to my flux source repository, the Helm release rolls out.

A Note About Software

in the end, i’m forced to admit there’s still a lot of software involved in all of this. setting aside the stuff that provisions and scales my cluster nodes, and the magic LoadBalancer, i have:

• nginx (running from a stock image);
• git & bash (running from a stock image);
• a remote git server (i’m running gitea¹⁰, but github dot com is fine here);
• Kubernetes (oops!);
  • flux, especially kustomize-controller and helm-controller;
  • ingress-nginx controller;
  • cert-manager and Let’s Encrypt;
• the bitnami/nginx Helm chart;

the bulk of this i’ll be able to reuse for the other things i deploy on the cluster¹¹. and it replaces SASS black-boxes like “AWS Amplify, CloudCannon, Cloudflare Pages, GitHub Pages, GitLab Pages, and Netlify” in the recommended Hugo deployment.

to actually deploy my site, i get to maintain a bash scripts for git-clone, my NGINX config, and a couple of blobs of YAML.

at least there are no webhooks.

fin