Closed viglesiasce closed 6 years ago
Things that need to be discussed:
Thanks for capturing this @viglesiasce. @sameersbn put together a great set of kubernetes manifest and helm chart authoring guides (https://github.com/bitnami/charts/tree/master/_docs) that could provide a base for something.
The one thing that has been bugging me is docker standards. @viglesiasce we probably need to get together with a couple of people to discuss, or we can hijack a sig apps.
I know I am beating this dead horse just a bit, but we need to go through the existing charts and get http://kubernetes.io/docs/user-guide/compute-resources/ laid out. Probably should update any standards docs / guidelines that we have.
A doc for us to collaborate on: https://docs.google.com/document/d/120IobQc4Q1__wSL5XZ7cSJEm2LDePoamUbX6N9SVuBA/edit?usp=sharing
@michelleN should we set 15 min in the helm sig or sig-apps to knock this out? I am knee deep in writing the HA upgrade guide, so I am not sure how we can knock this out.
@chrislovecnm have a look at the technical requirements that I bolstered today. I think it will be best to just build out that section in the docs: https://github.com/kubernetes/charts/blob/master/CONTRIBUTING.md#technical-requirements
Nice! Added a bunch of notes to the doc.
We need a linter to check for that stuff btw
For ease of reading
Also upgrades are not supported by Petsets... Just a note to the section
My pet list of things to consistify, after reading through a good number of these charts:
My best practice recs:
{{ template name . }}
{{ template name . }}-{{ .Values.container1.name }}
, {{ template name . }}-{{ .Values.container2.name }}
, etc.app
labels on resources should also use {{ template name . }}
instead of current {{ template fullname . }}
patterncomponent
labels to the respective resources
component
as a selectorI explained my reasoning for the first two items in #168 (specifically here, here, and here).
The last item is a pattern I use at work, explained here.
I would like to add a best practice item request... I would love to see the recommendation and practice of using a global property
persistence:
enabled: true/false
in all official stable charts that completely disables PVs/PVCs in the main Chart and Dependent charts.
Why? On k8s clusters with no PVs you have to go figure out what non standard parameters disable the PVC for each snowflake chart values and then also for the its dependencies. This is true e.g. for existing charts like WordPress in K8s Stable (you need to turn off PVs for MariaDB specifically).
In my view, this is a UX problem for newbies using the official Stable charts because on a cluster like "kube-solo" or "minikube", that many start on, there are no PVs and in those cases there will never be a mixed situation where part of the chart should have PVs. If this became a standard (you should also be able to individually turn chart PVCs off or on) it will undoubtedly mean less grief for users and chart maintainers...
Also, if this became a best practice standard then perhaps this could be added as a Helm install option (further improving UX).
I would like to add a best practice item request... I would love to see the recommendation and practice of using a global property
I'm -1 on this, as outlined in the matching discussion on the gitlab-ce PR:
https://github.com/kubernetes/charts/pull/202#issuecomment-260711737
tl;dr: Some charts like factorio, gitlab, and even Cassandra store various bits of data in different places. There are cases when some of your users might want persistence for one, but not the other. For example, we'd be actively wasting money/quota/resources for many of our factorio users, since the default install doesn't come with any mods to persist.
I think our policy of sane defaults covers the split, in that users with special needs will read the chart README/values and adjust as needed.
On k8s clusters with no PVs you have to go figure out what non standard parameters disable the PVC for each snowflake chart values and then also for the its dependencies
What percentage of our userbase do we think will be put in a situation like this? Is it enough to warrant the loss of flexibility?
Part of this is requiring 1.4+, which it looks like we may be leaning towards (based on @prydonius last comment and the thumbs ups): https://github.com/kubernetes/helm/issues/1528
@gtaylor did you read my answer on #202? The use case is clusters with no persistent storage which right now are pretty much all out of the box do-it-yourself dev cluster like minikube, kube-solo, kube-cluster and I'm suggesting having both flags per dependency and a global, not either or. So more flexible.
Another one for this list:
Ensure repository (not just tag) is configurable. https://github.com/kubernetes/charts/pull/276/files
Echoing what @mgoodness said above, app
labels are problematic (and long) when they include the release, I've also noticed that many charts don't have an app
label either. We've needed to add component
labels to various charts also so that we can select pieces to use with kubectl port-forward
, etc. I also add the release
label to selectors for specificity there.
Example changes for the prometheus server service:
@@ -2,7 +2,7 @@
kind: Service
metadata:
labels:
- app: {{ template "fullname" . }}
+ app: "{{ .Chart.Name }}"
chart: "{{ .Chart.Name }}-{{ .Chart.Version }}"
component: "{{ .Values.server.name }}"
heritage: "{{ .Release.Service }}"
@@ -15,6 +15,7 @@
protocol: TCP
targetPort: 9090
selector:
- app: {{ template "fullname" . }}
+ app: "{{ .Chart.Name }}"
+ release: "{{ .Release.Name }}"
component: "{{ .Values.server.name }}"
type: "{{ .Values.server.serviceType }}"
I'd like to know what the best practices are for indicating whether values are required or optional.
In addition to my previous suggestions, I've taken to changing the fullname
template in my internal charts. Rather than having it generate RELEASE_NAME-CHART_NAME
, it's flipped to CHART_NAME-RELEASE_NAME
. If a chart handles multiple components, I'll do something like CHART_NAME-COMPONENT-RELEASE_NAME
. I find it's easier to view multiple releases of the same chart in kubectl when using that pattern.
Pod names should include both the release name and chart name. When using a chart as a dependency its good to be able to distinguish pods that belong to the same release.
@prydonius,
Do we have a macro for this yet or is the _helper template for "fullname" the best we have right now?
@viglesiasce the fullname template is the best we have for now, see https://github.com/kubernetes/helm/issues/1186 - looks like it's slated for 2.3.0
would it make sense to say:
it is preferred to include defaults in a packaged
values.yaml
over using thedefault
function in the manifest templates?
it seems having default
everywhere in a template makes it less readable, and a secure default values.yaml
needs to be provided according to the technical requirements anyway?
(plus, you can see a lot of documentation is added to most stable chart values.yaml
- which is a single place to look instead of that information being spread over the files in templates folder)
I've tried to take into account the comments in this thread, as well as Bitnami's documentation into conventions for my company - https://gist.github.com/so0k/f927a4b60003cedd101a0911757c605a
Checksum your configs in an annotation on your deployments so that config changes can automatically update your pods:
annotations:
checksum/config: {{ include (print $.Chart.Name "/templates/configmap/spinnaker-config.yaml") . | sha256sum }}
@viglesiasce k8s configmaps mounted in volumes are updated and fire inotify... I think the above is more applicable for secrets (I'm not sure about configmaps referenced in env vars)
@so0k I think exceptions can (should?) be made for applications that monitor filesystem events and reload their configuration. But for the majority of cases, it's helpful to force a re-deploy when ConfigMaps and/or Secrets change.
Standardize (to the extent possible) how Ingresses are created. It's the wild west out there!
some I've hit recently:
Need a way to set nodeSelectors on templates. Otherwise, arrangement is completely random, which fails on nonuniform clusters.
For svc's, I would like a way to set the external_ip on them so I can access them outside the cluster without having to setup an ingress.
For Helm Charts to really take off as a standard way of managing k8s deployments, I think it is important for chart consumers to understand the usage patterns any Best Practices are designed to work best with.
[...] Pod names should include both the release name and chart name. When using a chart as a dependency its good to be able to distinguish pods that belong to the same release.
I sort of understand using this pattern with deployment/rs/rc/pod/etc names; though, I must say that I have gotten used to using label filters rather than relying on kubectl's default get
output.
Specifically, with services, names such as chartname-flinging-monkey
don't work. (I also have taken to putting the chartname before the release name) Even with specifically named releases, names end up similar to 'component-dev-feature-YYZ420`.
e.g.: I have a telegraf daemonset that exposes its statsd service to the cluster. One of the modifications I made to the telegraf chart was to set .metadata.name
to match the exposed service port.
Before:
dig +search +short _statsd._udp.telegraf-ds-prod-monitor.monitors SRV
after
dig +search +short _statsd._udp.statsd.monitors SRV
It gets even uglier when dealing with service configuration data via the container environment. (not terribly useful for inter-namespace work but stil)
With that said, I'm also interested in understanding the production context that this pattern comes from. The pattern seems to run counter to maintaining loose coupling between resources and components. I'm fairly certain my perception is due to not understanding and not any lack of wizardry on the part of the implementors.
Best practices for naming of template partials is discussed here: https://github.com/kubernetes/helm/issues/1688
Exposing the resource-policy
annotation on PV/PVC is covered here: https://github.com/kubernetes/helm/issues/1875
The summary provided by @so0k gist actually summarise most of the things we talked here + what I have seen in multiple charts in kubernetes repo.
In our startup, right now, we are converting our kubernetes yaml files by writing a lot of charts. We are planning to use @so0k gist for now.
Also, we follow "if you want best practices followed by most people, you should provide practical implementation & common use case examples" in our company.
For that, we are thinking about using "includes" or "blocks" for common labels and names patterns such as app, heritage, release, chart, component. Maybe templates for container-name, service-name... . So that we don't need to remember or copy-paste each time to follow the best practices.
I'd like to make small adjustments (haven't applied these back to the gist):
Prefix the files with the name of the Chart component
only makes sense if there is more than 1 component, to make a more re-usable chart skeleton... apps with a single component should have an agreed prefix (none or app-
?)fullname
template to be useful sometimes... I think my understanding of the way the subchart is referenced was a bit wrong and you still need to re-define the scoped fullname template in the parent chart anywayvalues.yaml
(and this is preferred over sprinkling | default
in the templates). This came from the originally wrong assumption within the company to keep secrets in values.yaml
and thus gitignore it, we've since adopted the practice of always including a values.yaml
and gitignoring a .secrets.yaml
For helm,
Are everyone @viglesiasce @mgoodness @technosophos is on the same page about improving up on @so0k gist and adding it to the docs as best_practices.md?
@so0k What do you think about providing includes, block or templates for best practices you summarised?
Sorry for bugging you all with many messages, I wanted helm charts move forward before we move everything to helm charts
includes
/ blocks
/ templates
for best practices such as the labels sounds good - there must already be charts out there that follow this practice?
ps, I personally also like the following snippet I saw in one of @ohaiwalt charts:
This converts snake-cased maps in the values.yaml
to environment variables:
# values.yaml
config:
kelsey-rating: "pretty dope"
env: gcpug
asset-pipeline: false
# manifest template fragment
env:
{{- range $key, $value := .Values.config }}
- name: {{ $key | upper | replace "-" "_" }}
value: {{ $value | quote }}
{{- end }}
# rendered template fragment:
env:
- name: ASSET_PIPELINE
value: "false"
- name: ENV
value: "gcpug"
- name: KELSEY_RATING
value: "pretty dope"
Example chart from a workshop I gave last week.
...to make a more re-usable chart skeleton... apps with a single component should have an agreed prefix (none or app- ?)
None IMHO, because multi component charts are app as a whole too.
I do find the default fullname template to be useful sometimes... I think my understanding of the way the subchart is referenced was a bit wrong and you still need to re-define the scoped fullname template in the parent chart anyway
I saw charts with multi components have scoped fullname s for naming components yamls with their own names, like here.
For subcharts, I didn't tried but subcharts fullname template will use parents chart .Chart and .Release, so it should be same fullname in parent and sub charts as I remember from docs.
My argumentation for nested hierarchies kind of falls away when you always have to document all options in values.yaml (and this is preferred over sprinkling | default in the templates). ...
I don't like default
in the templates too, Defaults should be given in values.yaml
for having them on one place.
@so0k
I am just writing followings for demonstration purposes
for deploy.yaml
...
metadata:
name: {{ template "deployment-name" .}}
labels:
{{ include "labels" . | indent 4 }}
spec:
template:
metadata:
labels:
{{ include "labels" . | indent 8 }}
spec:
containers:
- name: {{ template "container-name" .}}
...
@so0k Also, the implication here is we should have labels (app, heritage...) and names (name, fullname and different patterns) same as best practices at the _helpers.tpl
.
Chart developer can add their own labels directly to the yaml templates just over the {{ include "labels"...
. We don't limit them, we just make it easy to follow conventions.
@technosophos added best practices guide at kubernetes/helm#2007 for helm 2.3
I hadn't seen it before - thanks for the link!
So we can PR on those documents to add sample templates for default labels?
That's also news to me. Give us maintainers a chance to review and decide if we want to adopt those as the best practices for this repo. We'll also want to consider all the suggestions/requests made in this thread.
For first point of @mgoodness, I invite everyone to kubernetes/helm#2007 for reviewing current state of the guide. I believe we have still time for little changes until 2.3 on that PR.
Then we should shape&move suggestions here by discussing over a new PR at kubernetes/helm.
Unsuprisingly, @so0k gist mostly overlap with best practices guide but also includes additional suggestions that can be added as additional parts to the guide. I believe it is a good starting point for our new PR.
@so0k if you are willing to work on this, we can work together on that new PR and everyone is welcome.
Our plan is maybe:
If anyone(esp. @so0k) want to join me (actively), I want to work on this together
After best practices shaped more clearly, we can add suggestive comments and best practices to the default starter pack with another PR in the future.
IMHO, default starter pack should reflect what is written on the best practices guide closely.
@so0k That is better place to discuss including one template for recommended labels I mentioned before
Hey everyone, where is the best place to address yaml indentation inconsistencies across the stable charts?
For example, is there a standard for YAML array indentation?
- One
- Two
vs
- One
- Two
Both are legit, but most docs I see across projects (eg Symfony) use the second. Does charts have a standard?
I was talking with @prydonius about YAML linting - he noted https://yamllint.readthedocs.io/en/latest/rules.html#module-yamllint.rules.hyphens (personally, I am an advocate for indentation: {spaces: 2, indent-sequences: consistent}
, but really whatever the standard, as long as it’s consistent). Perhaps we could do as an automated PR check in the charts repo? But in either case, it seems this would be a good helper command in helm.
I noticed this was raised this comment https://github.com/kubernetes/charts/issues/44#issuecomment-256472424 and also in https://github.com/kubernetes/helm/issues/1663. I sought this out as a response to https://github.com/kubernetes/charts/pull/920. I'll be happy to follow up wherever is best, if anyone has a good idea of where to go with this?
Should this be closed with the introduction of https://docs.helm.sh/chart_best_practices/#the-chart-best-practices-guide?
@bacongobbler That makes sense. And we can continue to discuss these topics in separate issues/PRs against https://github.com/kubernetes/helm/tree/master/docs/chart_best_practices?
Issues go stale after 90d of inactivity.
Mark the issue as fresh with /remove-lifecycle stale
.
Stale issues rot after an additional 30d of inactivity and eventually close.
Prevent issues from auto-closing with an /lifecycle frozen
comment.
If this issue is safe to close now please do so with /close
.
Send feedback to sig-testing, kubernetes/test-infra and/or @fejta
.
/lifecycle stale
Stale issues rot after 30d of inactivity.
Mark the issue as fresh with /remove-lifecycle rotten
.
Rotten issues close after an additional 30d of inactivity.
If this issue is safe to close now please do so with /close
.
Send feedback to sig-testing, kubernetes/test-infra and/or fejta. /lifecycle rotten /remove-lifecycle stale
/remove-lifecycle rotten /remove-lifecycle stale
Issues go stale after 90d of inactivity.
Mark the issue as fresh with /remove-lifecycle stale
.
Stale issues rot after an additional 30d of inactivity and eventually close.
If this issue is safe to close now please do so with /close
.
Send feedback to sig-testing, kubernetes/test-infra and/or fejta. /lifecycle stale
We need a place where we can establish and document best practices for chart creation. This should then be reference from the README in order to point users to this repositories conventions.