Comment by debarshri
9 hours ago
You can pretty install it without ssh under 60s. The fun starts after it has been installed.
We have been running into lot of issues at production with k3s. There I embarked on journey to writing a kubernetes compliant and equivalent platform in rust with the help of claude [1]. It is a fun little project for now, still figuring out stuff, idea is to keep it minimal and single binary every embedded including CNI, and support various runtimes like docker, containerd etc but also wasm, vms and also jvm.
Very interesting!
Architecturally - where do you run Postgres ? I assume it would be external to the cluster ? (doing it internally would create a circular dependency ?)
Yes, it is external to the cluster.
If you want to do a quick setup, it creates a SQLite DB for the metadata.
You have to be careful trying to do this kind of thing. The problems you describe having below are problems with peripheral components, not k3s itself. The runtime handles garbage collection and image pinning. Your embedded runtime is using libcontainer, the same thing containerd uses, so the behavior should be identical. Since you support other runtimes, how they handle image pinning, if they support it at all, will vary. Whether or not you embed the CNI plugins and networking controllers, you're seemingly still using CNI since that's how container runtimes attach containers to a network, so whatever problems you had with CNI before would still happen. The DR VM not wanting to join sounds like it was probably due to etcd storing node IPs in the cluster member metadata. If you transfer that to a new host and it doesn't have the same IP, you need to first correct that metadata out of band, which no Kubernetes distro I'm aware of handles automatically but it's a simple etcdctl one-liner. You also need to make sure the client certificate you're using to authenticate with etcd is reissued with the new host IP in its IP SANs, which k3s does do automatically. If you're not using etcd, well, good in a way because it has a lot of cruft and I'm not a fan, but that will be difficult to support because the entire Kubernetes API and many third-party controllers are all designed around how etcd works. k3s doesn't actually require etcd and can use any SQL-based RDBMS thanks to its kine compatibility shim.
With all respect, "building it because I want to" and "working toward making (it) production grade" doesn't inspire a ton of confidence. k3s has been part of the CNCF for many years and its developer Darren Shepherd was the founding CTO for both cloud.com and Rancher Labs, which were acquired by Citrix and SUSE. It looks like you're running your own B2B company and hoping to swap out k3s as the underlying engine for multitenancy. That's very risky. Surely Claude can help you understand and use k3s just as readily as help you write a replacement, and I'm sure SUSE sells professional services. I have no clue what they charge but typically you're talking like $300 an hour and you'd probably only need 40 hours.
Sure, with full disclosure i dont expect anyone to run in production until i have. Absolutely understand your trust deficit.
Once i have embarked on the journey building this from scratch, there are new innovative ideas i can implement not bound to any foundation nor org.
Ps. We do not sell as product it is 100% free and opensource with MIT license.
Do you have a writeup what problems you ran into?
We do, let me check with my team and post it here.
There were many issues. On top of my mind was, after a DR drill where in a VM was booted, node did not join the cluster. Apart from that bunch of issues due to etcd, longhorn.
Another major one was the CNI stopped work for a particular node. Garbage collection for images was another, we labelled the images, it would still remove then from the node.
Bunch of these kind of issues when our requirement is fairly straightforward. Therefore we are working towards a strip down version.
There is lot of operation complexity in general and most of us can do without.
I've found a lot of issues come through somewhat naive networking setup - which is encouraged by the "just yolo it" installation instruction in the documentation. If you want to start understanding what's going on you'll end up in very weird corners very quickly. Also, if you don't want the API endpoints available to the world the documentation is not much help.
I've found things more stable if you can give a dedicated interface just for internal k3s communication. It can be a bridge interface on top of a vlan interface - but not the vlan interface itself, or some things will break in very interesting ways. Also, even when using IPv6, just stick with internal IPs and nat everything - touching internal IP ranges is no fun. Plus, if there's a chance you'd ever want to use dual stack, set it up with internal v6 addresses, and just don't use the v6 addresses for now. There's also a lot of unintuitive behaviour around dual stack networking - and lots of areas where documentation is just plain wrong.
I'm scripting our stuff with ansible - one of the more useful things was the realisation that in some areas changes which shouldn't break anything can lead to cluster communication being interrupted, which is a very interesting thing to deal with, especially when you can't pin it to that change that didn't touch anything close to that, and therefore should not be responsible. I've learned, and sprinkled checks to make sure all members can still reach each other in there now, so that at least when I break it on changes I directly know why.
Meanwhile our architecture team that surely supported 0 real life k8s went with no vendor, on premises deployments, claiming it was as easy as booting a VM, after 2y, there is 2 apps running and supposedly all future apps will be deployed on that cluster.
I cannot wait for the end of this month to leave that place.
1 reply →