Getting tired of Helm – any better way to handle deployments in Kubernetes?
I’ve been deep in Helm templates lately and it’s starting to feel like YAML hell.
It was fine when we had a few services, but now it’s just hard to manage.
Anyone found a workflow that avoids Helm altogether? Or made Helm manageable at scale?
Kustomize is easier to manage at scale, but some upfront effort is required. Many charts are distributed as Helm packages, so you’ll often need to export them as raw YAML manifests. In an ideal setup, ArgoCD combined with Kustomize should cover most deployment needs. However, depending on your workflow, you may eventually need a way to dynamically replace variables. If the built-in tools in recent Kustomize versions aren’t sufficient, consider using envsubst as a fallback.
> I handle deploy time dynamic variables with `sed`
I brought up envsubst because it’s a simpler, cleaner, and often overlooked option for variable substitution.
> Not that I haven't had the need, I've use jsonnet with libk8s at scale. But if you're asking the question this simply, you probably don't need it
In my view, Jsonnet isn’t an improvement - it’s complicated to learn, cumbersome to use, and prone to mistakes.
That said, if an organization decides to adopt any specific tool, I believe consistency in tooling, design, and practices is more important than the tool itself.
For complex stuff I write Python or Go programs to build manifests, then shell out to kubectl apply. An old example - deploying a multi-instance modded Arma 3 server on k3s: https://github.com/dharmab/homelab-k3s/tree/main/lab
T.b.h. if I were to write a manifest generator, I would still probably commit the thing into a repo and let argo do the rest. Maybe even fiddled around to make the generator into a config-management-plugin ... but that feels like over-doing it.
I've been working on improving a tool called kr8+, which uses jsonnet to combine cluster config and apply it to components: https://github.com/ice-bergtech/kr8
What exactly are you doing with Helm that's making it so painful to use, and what does your development workflow look like? I've certainly had my fair share of issues with Helm, especially when trying to get a bit too fancy with creating Helm libraries, and standardised charts. I've also found that trying to aggregate multiple charts into a single chart for deploying an environment can also become a nightmare to manage.
I'm currently looking at Helmfile so that I don't need to aggregate charts into a 'parent chart', and i'd also like to move towards a single standardised chart that all microservices can use, rather than spin up a new chart for each service.
there's also https://timoni.sh/ if you want a helm-like experience, but with cue instead of templating.
If you're working with upstream projects, unfortunately many of them will only provide helm charts, so you got to decide between rewriting them to suit your env/tool, or just live with the crappiness of helm.
Recently I have been writing more stings in jsonnet. If I were with more haskell-friendly team, might even try dhall. In general, I feel like writing the yaml in something else than yaml is the way to go, and as long as you get imports and way to do templating that is not just string interpolation, you are good.
IMHO at scale (both in terms of complexity and org level) having something consistent helps more than trying to fight the de-facto standard. Since most upstream projects are mainly distributed as Helm charts, going with anything else will require more effort eventually.
Holos.run for my homelab cluster. Cuelang has a learning curve but works well with argo unlike Timoni and let's you import existing charts, bare manifests, and use kustomize. Let's me abstract config with custom types and unification/(inheritance if that is easier to think about but a bit wrong)
Yeah the whole 'git repo = helm chart' just does not feel great at all. As we all know, the only thing worse is not using helm and having to deal with writing all those service, pv, pvc, ingress yaml files individually :)
(Not that I haven't had the need, I've use jsonnet with libk8s at scale. But if you're asking the question this simply, you probably don't need it)
I brought up envsubst because it’s a simpler, cleaner, and often overlooked option for variable substitution.
> Not that I haven't had the need, I've use jsonnet with libk8s at scale. But if you're asking the question this simply, you probably don't need it
In my view, Jsonnet isn’t an improvement - it’s complicated to learn, cumbersome to use, and prone to mistakes.
That said, if an organization decides to adopt any specific tool, I believe consistency in tooling, design, and practices is more important than the tool itself.
For complex stuff I write Python or Go programs to build manifests, then shell out to kubectl apply. An old example - deploying a multi-instance modded Arma 3 server on k3s: https://github.com/dharmab/homelab-k3s/tree/main/lab
T.b.h. if I were to write a manifest generator, I would still probably commit the thing into a repo and let argo do the rest. Maybe even fiddled around to make the generator into a config-management-plugin ... but that feels like over-doing it.
I'm currently looking at Helmfile so that I don't need to aggregate charts into a 'parent chart', and i'd also like to move towards a single standardised chart that all microservices can use, rather than spin up a new chart for each service.
i like generating k8s yaml with cue, example: https://github.com/cue-labs/cue-by-example/tree/main/003_kub...
there's also https://timoni.sh/ if you want a helm-like experience, but with cue instead of templating.
If you're working with upstream projects, unfortunately many of them will only provide helm charts, so you got to decide between rewriting them to suit your env/tool, or just live with the crappiness of helm.
https://github.com/carvel-dev/carvel
works really well
https://github.com/mkubaczyk/helmsman