It has only really become viable a few releases ago. One of my colleagues here at SAP did a diff plugin for visualizing what's going to happen in the next upgrade: https://github.com/databus23/helm-diff
For me, the value of helm is that the templates are something I can commit and track through source control.
This essentially bypasses that.
I can see how it might be handy in a pinch, or during development, but I'd rather invest in tooling to ensure I can foster a commit to release rather than have operators one-offing my cluster.
I agree, but for me, Helm releases give the ability to track changes in configuration of my apps (you can see the history and roll back if something is wrong)
That's funny, because Helm is actually already keeping a revision history with Kubernetes primitives (the pods belonging to the old release are scaled down and Helm marks the ReplicaSet as "superseded", from the prior release, that's how rollback works.) To begin with I have no doubt there are great reasons to keep all of that in a formal source control.
I have always wondered about the decision to put all of the charts into a single "Kubernetes/charts" repo. It seems like it would make a lot of sense to have a repo for each chart, but then I guess you lose the capability of being able to pull them all down at once, or quickly update your local copies of everything in one shot. The charts are extremely small by design, maybe the overhead of a single git repo per chart was judged to be too much, and the value of having all the stable charts in one place was considered to be of greater magnitude.
One common pattern I've seen is to put your charts directly into the application's source tree (which is certainly under version control.) This is how Draft sets up a chart. That seems OK to me, it seems very much like the creators of Helm would like for you to be able to get to work and having rollback capabilities without creating any (user-facing) dependency on a Git repository, or necessarily knowing anything about how to use Git or other SCM tools.
(Draft takes this a step further; if you're using it to the fullest extent, then it generates a new release in Helm for every FS_NOTIFY as you save your work.)
>I have always wondered about the decision to put all of the charts into a single "Kubernetes/charts" repo.
There's discussion happening right now among the Charts maintainers about changing this. If you're interested you can join the conversation in #charts on the Kubernetes slack.
Full Disclosure: I'm an author of App Registry which is a project that attempts to integrate projects like Helm with registries like Quay on a multiple-repository basis.
This is basically what we're doing: With each app, we put the chart in the app's repo itself, and we built a wrapper around Helm that performs deploys for us, using "helm upgrade --install ./k8s" from the file system. Keeping the chart with the app means it's versioned along with the app.
(It's not a general-purpose wrapper; it's the CLI we use to manage anything related to our clusters and apps, basically: DNS, secrets, ad-hoc shells etc.)
One of the things this wrapper does is ensure that every deploy is annotated with metadata — git revision, git branch, git repo, and also who deployed it. This allows a developer to run a status command to determine what versions of each app are currently deployed, and also show a log of new commits that haven't been deployed yet. It also means you can easily see that an app was deployed from a specific branch, for example.
I think Helm's design is a bit flawed here. It's catering to two different purposes: One is deploying packaged versions of third-party software (e.g. PostgreSQL); the other is deploying your own software. There's an inherent conflict there in that the release process is different. Why can't Helm pull directly from git and use its revisions as version numbers? It doesn't make any sense for us to "publish" an app to a "repository" when we already have git.
Right now, we're basically using Helm as a glorified templating system, and it's a somewhat buggy one, so we may decide to ditch Helm and use a pure templating solution instead. I haven't looked at Draft, but seeing as it's based on Helm, I'm not sure that it can improve on the basic design?
Draft is a very simple tool and it's not just based on Helm, it is using Helm under the hood.
You will probably not find much use for Draft given the more advanced state of your tooling. It's been described as "a gateway drug to k8s" because, for the uninitiated, it simplifies the process of "OK, I want to put my app on K8s, what now?" – it sets you up with a reasonable default chart, and a dummy Dockerfile based on what kind of app the Draft logic decides your app is.
(Eg. it's going to recommend ruby:onbuild if you have a Ruby app.)
That might work out of the box, more likely you need to make some changes in the Dockerfile, but that could very well be all.
It's really more like an advertisement for Deis. It's the "zeroconf" version of Deis. If you've never heard of Deis, I'll be happy to tell you more about that; Deis has a builder service/pod and acts as a git repository, so you set it up as a git remote with the deis commandline tool and 'git push deis' to prompt a release upgrade. It is more likely what you are probably expecting as a piece of your CI/CD solution.
For people that aren't already fluent users of git, k8s, helm, charts... Draft is easier to get started with, and it gives a chance to expose you to all of those things, without making any of it prerequisite knowledge to spin up your app on k8s at once.
For literally anyone else, you're going to get a much better experience from Deis, which is not only out of Beta but long past the 2.0 milestone (when it moved to K8s from the prior CoreOS/Fleet-based stack.) Deis is installed via helm, but after that you do not need to use the Helm client. Your apps can use Procfile-style or Dockerfile for deployments.
The builder is very modern and developed, and as it evolves, it picks up a lot of efficiencies that you might have seen before on Heroku if you're familiar with that platform, making it very convenient to work with. (That also means it brings the baggage of 12-Factor, if you consider that as baggage...)
To be clear, I think that Deis is using helm under the hood too, but you never have to interact with helm or your chart.
Draft is written so you could probably avoid ever running the helm client, but it is a much thinner layer of sugar on top. You are very near the helm charts, but in Deis you might never even realize that there are helm releases associated with your app. The Deis client interface is pretty fully featured, almost drop-in as a comparison to heroku client. Using Draft, you are encouraged to edit your charts, but with Deis it is well-hidden and there is probably never any need to do that.
Service brokers (for supporting dynamic allocation of things like shared instances of postgres and redis, like you get with heroku) are a thing, but they are newer and I have not seen them being used in practice yet.
By Deis, are you referring to Workflow? [1] There's no product called "Deis" anymore.
Workflow seems too high-level and large for my taste. It's an entire PaaS on top of Kubernetes, which also brings way too much to the table (edge routing, log/metrics management, alerting, etc.).
The advantage of Helm over static K8s configs or simple templates is that charts can use other charts for complex dependencies, and can have multiple releases with different values. The latter is useful for different environments (dev, staging, production) or for changing dynamic values separate from template code (replicas, resources, versions).
Helm has separation between charts and templates, and releases and values. The former should always be version controlled and deployed like code. The latter are local to cluster. They can be version controlled but then need tool to sync with the cluster.
It looks like this is a tool for editing data on cluster. Which is something that helm currently makes hard cause need to dump current values, edit them locally, and update with changes.
If you (and others on this sub-thread) like managing kubernetes through version control, I humbly suggest trying out the ksonnet/kubecfg toolchain (https://github.com/ksonnet/kubecfg and the closely related ksonnet/ksonnet-lib repo - Disclaimer: I'm one of the contributors).
Helm is nice, but (imo, and as some of the comments here are also pointing out) oriented more towards consuming 3rd-party pre-built charts from some sort of centralised marketplace/repo. (This is an important use case, and yes I know it is certainly possible to use helm to manage your own internal software too.)
18 comments
[ 1.9 ms ] story [ 48.2 ms ] threadThis essentially bypasses that.
I can see how it might be handy in a pinch, or during development, but I'd rather invest in tooling to ensure I can foster a commit to release rather than have operators one-offing my cluster.
I have always wondered about the decision to put all of the charts into a single "Kubernetes/charts" repo. It seems like it would make a lot of sense to have a repo for each chart, but then I guess you lose the capability of being able to pull them all down at once, or quickly update your local copies of everything in one shot. The charts are extremely small by design, maybe the overhead of a single git repo per chart was judged to be too much, and the value of having all the stable charts in one place was considered to be of greater magnitude.
One common pattern I've seen is to put your charts directly into the application's source tree (which is certainly under version control.) This is how Draft sets up a chart. That seems OK to me, it seems very much like the creators of Helm would like for you to be able to get to work and having rollback capabilities without creating any (user-facing) dependency on a Git repository, or necessarily knowing anything about how to use Git or other SCM tools.
(Draft takes this a step further; if you're using it to the fullest extent, then it generates a new release in Helm for every FS_NOTIFY as you save your work.)
There's discussion happening right now among the Charts maintainers about changing this. If you're interested you can join the conversation in #charts on the Kubernetes slack.
Full Disclosure: I'm an author of App Registry which is a project that attempts to integrate projects like Helm with registries like Quay on a multiple-repository basis.
(It's not a general-purpose wrapper; it's the CLI we use to manage anything related to our clusters and apps, basically: DNS, secrets, ad-hoc shells etc.)
One of the things this wrapper does is ensure that every deploy is annotated with metadata — git revision, git branch, git repo, and also who deployed it. This allows a developer to run a status command to determine what versions of each app are currently deployed, and also show a log of new commits that haven't been deployed yet. It also means you can easily see that an app was deployed from a specific branch, for example.
I think Helm's design is a bit flawed here. It's catering to two different purposes: One is deploying packaged versions of third-party software (e.g. PostgreSQL); the other is deploying your own software. There's an inherent conflict there in that the release process is different. Why can't Helm pull directly from git and use its revisions as version numbers? It doesn't make any sense for us to "publish" an app to a "repository" when we already have git.
Right now, we're basically using Helm as a glorified templating system, and it's a somewhat buggy one, so we may decide to ditch Helm and use a pure templating solution instead. I haven't looked at Draft, but seeing as it's based on Helm, I'm not sure that it can improve on the basic design?
You will probably not find much use for Draft given the more advanced state of your tooling. It's been described as "a gateway drug to k8s" because, for the uninitiated, it simplifies the process of "OK, I want to put my app on K8s, what now?" – it sets you up with a reasonable default chart, and a dummy Dockerfile based on what kind of app the Draft logic decides your app is. (Eg. it's going to recommend ruby:onbuild if you have a Ruby app.)
That might work out of the box, more likely you need to make some changes in the Dockerfile, but that could very well be all.
It's really more like an advertisement for Deis. It's the "zeroconf" version of Deis. If you've never heard of Deis, I'll be happy to tell you more about that; Deis has a builder service/pod and acts as a git repository, so you set it up as a git remote with the deis commandline tool and 'git push deis' to prompt a release upgrade. It is more likely what you are probably expecting as a piece of your CI/CD solution.
For people that aren't already fluent users of git, k8s, helm, charts... Draft is easier to get started with, and it gives a chance to expose you to all of those things, without making any of it prerequisite knowledge to spin up your app on k8s at once.
For literally anyone else, you're going to get a much better experience from Deis, which is not only out of Beta but long past the 2.0 milestone (when it moved to K8s from the prior CoreOS/Fleet-based stack.) Deis is installed via helm, but after that you do not need to use the Helm client. Your apps can use Procfile-style or Dockerfile for deployments.
The builder is very modern and developed, and as it evolves, it picks up a lot of efficiencies that you might have seen before on Heroku if you're familiar with that platform, making it very convenient to work with. (That also means it brings the baggage of 12-Factor, if you consider that as baggage...)
Draft is written so you could probably avoid ever running the helm client, but it is a much thinner layer of sugar on top. You are very near the helm charts, but in Deis you might never even realize that there are helm releases associated with your app. The Deis client interface is pretty fully featured, almost drop-in as a comparison to heroku client. Using Draft, you are encouraged to edit your charts, but with Deis it is well-hidden and there is probably never any need to do that.
Service brokers (for supporting dynamic allocation of things like shared instances of postgres and redis, like you get with heroku) are a thing, but they are newer and I have not seen them being used in practice yet.
Workflow seems too high-level and large for my taste. It's an entire PaaS on top of Kubernetes, which also brings way too much to the table (edge routing, log/metrics management, alerting, etc.).
[1] https://deis.com/workflow/
It looks like this is a tool for editing data on cluster. Which is something that helm currently makes hard cause need to dump current values, edit them locally, and update with changes.
Helm is nice, but (imo, and as some of the comments here are also pointing out) oriented more towards consuming 3rd-party pre-built charts from some sort of centralised marketplace/repo. (This is an important use case, and yes I know it is certainly possible to use helm to manage your own internal software too.)