It’s all too easy to kill a k3s cluster. I’ve been using k3s for years now and I’ve had plenty of adventures tweaking various aspects of running it. Before it’d take just a small change to an Argo Application to trigger a cascading failure. Hopefully now it’s a bit more resilient. Just a bit.
Helm was meant to be the package manager for Kubernetes. One common problem for package managers is “how do I find my packages?” Many package systems opt for having a default central repository for stuff. Distros have their central repos for apt. Programming languages too: for Node it’s npm, for Ruby it’s RubyGems, for Java it’s Maven central, for Clojure it’s Clojars. Of course most if not all systems have a way to add other package repositories or at least some other means to pull in dependencies (referencing git commits for example).
For Helm the central repository of charts/stable used to be the obvious default. You can of course add other repositories too, but defaults are powerful and many people will just give up if something is not available in the default source. On the other hand, having everything in one place puts a huge burden on the maintainers of that one place, as was the case of charts/stable. So they deprecated it.
My cluster is now running on k3s 1.20.6 and Argo CD 2.0.0 with its Helm chart at 3.2.2. Actually, upgrading Argo itself wasn’t much of a problem. I just changed the targetRevision of the Application and it was up and running in a few minutes. Then a few days later things got interesting.
There were no downtimes, but I noticed that Argo started failing to sync itself. Apparently a new minor version of the Helm chart came out (though it was still the same application version) that added support for the networking.k8s.io/v1 version of Ingress. However, it also accidentally broke clusters running Kubernetes before 1.19. And mine was one such.
While the Argo people are figuring out how to fix this (if), I decided to go and take this opportunity to upgrade my cluster. This wasn’t as painless as it should’ve been though.
My 4-node k3s cluster (where this blog is hosted too) kept dying every now and then. Looking at kubectl describe nodes it quickly became evident that this was caused by the nodes running out of disk space. Once a node gets tainted with HasDiskPressure, pods might get evicted and the kubelet will be using (quite a lot of) CPU trying to free disk space by garbage collecting container images and freeing ephemeral storage.
My setup by default uses local storage (the local-path provider) where volumes are actually local folders on the node. This means that pods that use persistence are stuck with the same node forever and can’t just move around. This makes eviction a problem, since they have nowhere else to go. It also means that disk usage is actually disk usage on the node, and not on some block volume over the network.
It’s been a year since I wrote about bootstrapping a cluster with Argo and using Argo Rollouts for canary deploys based on Prometheus metrics. Since then many things have changed. I moved from Digital Ocean to Linode (mostly because Linode has a Tokyo region) and from a single-node k3s “cluster” to a 4-node one. But most of how I use Argo CD for GitOps hasn’t changed.
kubectl has the feature to select objects by filtering on labels using the -l flag. Labels are key-value pairs attached to objects as metadata and they don’t have to be unique. I’ve most often seen them used to identify what project or app an individual resource belongs to. Helm uses labels to mark resources with the app, chart and revision they belong to.
But wait, if they’re not unique and there is a way to select multiple values with set operators, how does that work? The database backing Kubernetes by default, etcd is a key-value store. While it can natively select multiple records by prefix matching, it’d be hard to imagine labels working like that. There are many of them and the selectors are complex.
So I dove into Kubernetes’s source code to figure out how it works.
A while back I wrote about bootstrapping a Kubernetes cluster. I’ve been refining the setup so that it requires as little manual kubectl‘ing as possible. I still use ArgoCD to get everything rolling, and there is one bit that kept going red: persistent volumes.
I wish it could be completely automated… But for now I’ve just automated as much as possible (and convenient). The ingredients:
I have a repository for the purposes of playing around with Kubernetes tooling like this – and hopefully turn it into an actual application eventually. I have big plans and lots of stuff I want to try out, but time is limited. All the code examples in this post use the namespaces and naming choices in the repository. The folder structure (relevant to this bit) is like…
system
├┬ apps
│└─ (bootstrapped Argo CD app manifests)
├┬ argo
│└─ the local "umbrella chart" for Argo CD and Argo Rollouts
├┬ bootstrap
│└─ boilerplate project and application manifests
└┬ manifests
└─ manifests I didn't bother turning into a Helm chart
referenced by the raw-manifests.yaml application
アプリコードを変えました。後は機械がなんとかするはずのところ、そこから何十分もかかる手動のデプロイ作業が待っていた。アプリが動く台が増えるとその時間も台数の分だけ倍増する。
理想は、コードの変更がgit上で主ブランチにコミットとして現れたら、CI/CDパイプラインがコンテナイメージを生成して（マージ前にすでにテストが通っている前提）本番環境にカナリヤの方式でデプロイされる。
Kubernetes使おうとすると早い段階で永続化の話が自然と出てくる。データベースをクラスタ内で動かすとか、ユーザー界の理由もありうるし、クラスタの土台となるツールが永続ボリューム（Persistent Volume）を要求することもある。
