k8s-config/cluster-v2-design.md

312 lines
6.8 KiB
Markdown

# Design for Second Iteration of Cluster/Homelab
## Context
Current cluster was set up just to run CI builds as a
trial.
I'm now sold the k8s is a good approach and would like
to move more of my services to it.
This document will track my design for cluster v2.
## Investigation
### Host OS
Debian:
- on laptop
- already on most of systems
- stable
- not officially tested by k3s
- Will be using apt at work
Stream:
- Tried with k3s and had to disable systemd...
- On second try seemed to work even with error I saw before.
- Cockpit is nice when managing servers.
- Want to like RHEL
- More stable than Fedora
- RPMs are easier to work with
- Using on VM host
Fedora:
- Want to like RHEL
- Tested with k3s
- Latest podman and frieds
- Really fast for something stable...
- Cockpit is nice
- Fedora minimal can't be installed on
cockpit without hitting tab alot.
Decision: Fedora Server
### K3S Distro
RKE2:
- no Debian support
- 4GB Minimum
- 2 CPU
- cilium and nginx not default
k3s:
- k3d is a thing
- documentation online is good
- 512 MB of RAM
- 1 CPU
- easy installation
Decision: k3s
### How many clusters?
Decision: Exactly two (one for "need to work" services one for CI and messing around).
The mess with longhorn scared me... it wouldn't be that big a deal if it only effected
CI, but it also effect Kanboard and git.
### Files
Decision: Host local.
Files are not something I want to have to think about.
The longhorn mess scared me.
NFS not working with postgres is annoying.
### How many nodes per cluster?
The current cluster has lots of small VMs, with VMs added with more
CPUs/RAM as the requirements grew.
I'd rather limit myself to fewer more powerful VMs, and let the VM OS manage
CPU and memory.
More nodes would be useful if they were on different base hardware.
Realistically I'm never going to pay for more than the Ingress VM...
Decision:
1 big VM per cluster.
Both VMs hosted on current hardware.
If we add hardware, can add an additional node at that time.
### Networking
Status quo is flannel with vxlan with Traefik and Klipper LB and CoreDNS.
#### DNS
CoreDNS is great.
#### Load Balancer
Klipper works fine now.
MetalLB is the other option, is more complicated and doesn't
seem to give much particularly with a single node cluster.
Decision: Klipper
#### Ingress
Traefik:
- Status Quo.
- Works fine.
- Outside of k8s I don't like.
nginx-ingress:
- Google
- Used by a lot of people.
- Nothing sexy or risky.
- auth exposed in annotations
ingress-nginx:
- nginx upstream.
- extra features like stream support that I'm using
on lightsail now.
- full blown virtual server support.
- maybe too complicated?
- exposes same features as I have on lightsail through
annotations, which could be a thing to get keycloack to
work.
- auth in annotations is behind paywall, but available through
a virtual server
Decision: nginx-ingress
Use LB for stuff I would use the virtual server for.
#### CNI
flannel vxlan
- status quo
- works fine
cilium
- label based network policies
- leaning toward this plus multus though I doubt
I'll ever write a policy
- I want the ability to write a policy...
- if set up different pod cidr can do multi-cluster later
- cluster name and cluster id at install time
- can do transparent encryption (not worth it...)
cilium multi-cluster networking:
- not worth the complexity
- will manage connections with ingress/egress methods.
flannel wg
- encrypt traffic and set up overlay if I want to interact with
cloud machines
- can do the same with a manual wireguard network...
istio
- I dislike side car containers
- Traffic I'm interested in is mainly not L7.
- blessed by Air Force
Decision: flannel vxlan
not worth the extra complexity of cilium.
## What goes on each cluster/VM?
Lightsail:
1. Wireguard
2. Apt/RPM repos
3. Main NGINX Proxy
Infra Cluster:
- On Host:
1. CoreDNS
2. Wireguard
- On Cluster:
1. Keycloak
2. Kanboard
3. OneDev
4. Harbor
Main Cluster:
- On Host:
1. Wireguard
- On Cluster:
1. Tekton
2. MQTT Broker
3. Squid
4. j7s-os-deployment
## Deployments
Manually kubectl apply:
- Easy to reason about
- running apply is fun
- using flux has chicken and egg problem if git is also
deployed from flux
Flux:
- More git ops-y
- chicken and egg problem is conquerable, in a maybe
confusing way
Decision:
1. Infra:
1. kubectl apply/helm everything
2. Drop keycloak image manually in k3s either using cri or
placing in magic place after k3s install.
3. Use helm with values for onedev.
4. Get rid of Kanboard custom image.
Use kubectl apply.
2. Test:
1. Mostly kubectl apply for tekton.
2. Use flux for:
1. MQTT
2. j7s-os-deploy
3. squid
## VM Resources
Lightsail:
- Leave alone
Infra Cluster:
- RAM 4 GiB total
- 3 CPUs
- 200Gib Hardrive
Main Cluster:
- RAM 4 GiB total
- 3 CPUs
- 200Gib Hardrive
## Experiments
### k3s with cilium and nginx on Centos Stream 9
```
systemctl disable firewalld --now
export INSTALL_K3S_EXEC="server --disable traefik --flannel-backend=none --disable-network-policy --selinux"
curl -sfL https://get.k3s.io | sh -s -
```
I see an error about selinux policies conflicting, but I'm not sure if it matters?
Install cilium following instructions here:
https://docs.cilium.io/en/v1.12/gettingstarted/k3s/
Install nginx with:
```
helm upgrade --install ingress-nginx ingress-nginx \
--repo https://kubernetes.github.io/ingress-nginx \
--namespace ingress-nginx --create-namespace
```
### k3s with nginx on fedora server
```
sudo systemctl disable firewalld --now
export INSTALL_K3S_EXEC="server --disable traefik --selinux"
curl -sfL https://get.k3s.io | sh -s -
sudo chown jimmy:jimmy /etc/rancher/k3s/k3s.yaml
sudo dnf install helm
export KUBECONFIG=/etc/rancher/k3s/k3s.yaml
helm upgrade --install ingress-nginx ingress-nginx --repo https://kubernetes.github.io/ingress-nginx --namespace ingress-nginx --create-namespace
```
Import simple-ros2.
Laptop:
```
podman save -o simple-ros2.tar simple-ros2:latest
scp simple-ros2.tar 192.168.1.106:~/.
```
On server:
```
sudo ctr images import ./simple-ros2.tar
# wait forever....
```
Test yaml:
```
---
apiVersion: v1
kind: Pod
metadata:
name: test-pod
spec:
containers:
- name: simple-ros2
image: localhost/simple-ros2:latest
imagePullPolicy: Never
args: [ros2, launch, j7s-simple, j7s_publisher_launch.py]
```
### VM Host set up
I **think** I ran something like this when I set up the VM host.
I don't remember exactly, and I didn't document it...
This should be carefully looked at before running.
```
nmcli connection add ifname br0 type bridge con-name br0 connection.zone trusted
nmcli connection add type bridge-slave ifname enp4s0 master br0
nmcli connection modify br0 bridge.stp no
nmcli connection modify enp4s0 autoconnect no
nmcli connection down enp4s0
nmcli connection up id br0
```