[scjody]

As a relatively new Kubernetes user, the big thing on my wishlist is templates (for example so I can share YAML between my production and staging environments). Unfortunately the k8s team have already said "no" to that, and point to a list of alternatives. But without a simple solution that lots of people use, it's hard to know which one to pick and it doesn't look like any are particularly good. I took a quick look at a bunch of them and they seem to be either half-baked solutions that someone put up on GitHub, or tools that introduce way more complexity than I need right now (like Helm).

[linkmotif]

Helm is a lot simpler than it looks, and it’s awesome.

I also had this exact perspective, not wanting to introduce more complexity and therefore held off on Helm for a long time. But then I got comfortable enough with Kubertnes and also my project got complex enough that I needed templating and then Helm was great to pick up.

I just use one chart that I wrote for my application and I don’t mess around with repositories. I render the chart locally with helm template[0] to build my containers on Google Cloud container builder[1]. Then I deploy with helm upgrade. It’s a great flow!

[0]https://github.com/technosophos/helm-template

[1]https://github.com/dminkovsky/kube-cloud-build

[bogomipz]

I've really confused by Helm. From their github repo Helm bills itself as:

"The Kubernetes package manager."

Then it goes on to state:

"Helm is a tool for managing Kubernetes charts."

Then reading further I see:

"Charts are packages of pre-configured Kubernetes resources."

I find this extremely confusing. Also its my understanding that unit of deploy in K8 is a pod which is built on Docker containers. Isn't installing packages into running containers considered an anti-pattern?

Could someone give a better explanation of Helm/Charts and when someone might use it?

[linkmotif]

Yeah it’s not the best docs. But the tool is great.

Helm just renders a bunch of templates of kubernetes manifests against a values file. It then takes the resulting manifests, compares them to what’s running on your cluster and makes sure nothing less or more is running. That’s it. What’s nice about it is that you can then change the manifest templates or values files, and when you install the updated result, you have a single concrete new state that your entire application is in. Which makes it easy to version your entire application, whether you changed just one env var in one manifest or made many changes thoughout.

There’s a package manager component but you can skip that completely and imo it’s just confusing that they lead front and center with that component.

The essential commands are helm ls and helm upgrade. And helm rollback is nice too.

[bogomipz]

Thank you for this. This is helpful.

So its kind of like configuration management for your manifests then? Would that also be an accurate analogy?

[linkmotif]

Great to hear.

> So its kind of like configuration management for your manifests then?

Basically, yep.

Which is why, for example, it makes it easy to deploy staging environments. Instead of writing your manifests are the final finished product, you set up your manifests to be interpolatable templates. Then you can deploy your application one way to your production environment, but then staging another way (different env vars, different replicas, different anything, but still the same app).

The whole package management/sharing of charts thing falls out as a result of this, but I've yet to find this useful. I don't pay attention to my chart's version number (it's always 1.0), but instead aggregate all the tags for each image I need throughout my manifest templates in the values file. Then when I upgrade my helm release to deploy those latest tags, helm provides me with a monotonically increasing integer for the resultant release. I can always point to this integer and ask, "what manifests were running at this release?"

[ecnahc515]

Similar. The reason they say "package manager" rather than "configuration management" for your manifests is because it's actually much more about packaging than it is config management.

The chart packaging format and template language don't allow you to easily encapsulate or abstract away your dependent charts, it's really just a simple package which you can use as-is, but not really possible "tweak" or "extend" charts without forking them. Contrasted this to a CM system which typically make it possible to build new modules/cookbooks/recipes/whatever on top of others, with the possibility to hide away the details of the underlying module/recipe/etc.

It's really just tarballs with yaml templates and a way to specify values for the templates.

[linkmotif]

I have yet to appreciate with the whole tarball/package thing.

My application is 5-6 deployments, a 3 statefulset and a cron job. I have it all in one chart that I keep unpackaged on in a directory. The workflow is (i) update the templates/values, (ii) ensure necessary containers live on GCR, `helm upgrade` my application. And that's it.

I've yet to find a need for breaking this single chart into multiple charts with dependencies. Maybe once it gets much, much bigger? I also don't like the idea of pulling in third party charts as dependencies, even if they're public from github. This last point is probably irrational and me-specific, but something about it is less appealing than just copy/pasting the manifests I want into my project. So much more direct.

[bogomipz]

Thanks, this really helps to intuit the project's documentation for me. Cheers.

[xur17]

I do something similar, but we only use helm to render the templates - we do not utilize any of its deployment processes. Instead we use gitlab to render and deploy a template as a normal k8s object.

[web007]

You can use any templating system you like, then "| kubectl create -f -"

I would like YAML includes, or some way to re-use templates directly. It's mildly inconvenient to have all of my env vars templated, but to require N copies for N variants on runcommand / base name / port / etc. I could Xzibit templates in my templates I suppose, by having a simpler format than YAML with #include would be awesome.

[donaldguy]

But if you don't love YAML itself you might just wanna check out ksonnet: https://ksonnet.io/

[nameless912]

I just hacked together a solution using Jinja2. I literally throw template strings in whatever fields I need templated and render the whole thing to /tmp/ using a stupid simple python script. It's not pretty, and admittedly it's a little fragile and home spun, but it works.

[scjody]

I'll probably end up doing the same. For now we're just duplicating things in two places.

[donaldguy]

helm recently(ish) added a helm template command to do client side rendering of a chart

So if you wanted to just create a folder with a Chart.yaml, and a templates/file.yml.tpl with some `{{.Values.foo}}` template strings

It is sufficient to just run `helm template --set foo=bar | kubectl apply -f -` at that point.

Idk if that is little enough complexity for you. Similarly I use https://github.com/gliderlabs/sigil a few places where I want to include specific fields from external json files

[linkmotif]

Same here, helm template is great. It’s nice to pipe output from helm template to https://github.com/dminkovsky/kube-cloud-build if you’re on GKE.

If you’re using helm, is your json interpolation for things that can’t be in values.yaml?

[lobster_johnson]

You can use Helm with local files, instead of fiddling with the awkward "repository" concept:

  helm install .

We use this to wrap Helm with a nice deployment tool that does a lot more (e.g. automatically annotate our resources with git commit information).