Druid Helm Chart

[drcrallen]

Helm is an easy to use deployment system for running groups of services on kubernetes (k8s). This ask is that a Helm Chart (the thing that defines groups of apps) be developed for easy Druid deployment.

The thing which k8s is not terribly good at is a common historical setup where two tiers of historical servers are replicas of each other. I don't know how to get Stateful Sets setup in k8s to accommodate such a setup. But other nodes with no local state like brokers, overlords, coordinators, even routers should be pretty easy to setup charts for.

Easy:

• Broker
• Overlord
• Coordinator
• Router
• Zookeeper

Hard:

• Historical
• Middle manager

[josephglanville]

Regarding the "common setup" ff you wanted to create 2 distinct tiers of historical servers you would most likely create 2 different StatefulSets, i.e statefulset/historical-hot, statefulset/historical-cold or something. Because discovery is done via Zookeeper rather than k8s labels etc no real care has to be taken when doing this, it should "just work".

We are running Druid on k8s (without Helm) so I would be interested in contributing to this effort but don't have the bandwidth to lead it right now.

[drcrallen]

@josephglanville the problem with 2 stateful sets is that one doesn't know what the other is doing. So I'm not sure there's a way to say "don't do any operations on this set while the other doesn't meet stateful requirements". Its kind of a... stateful set of stateful sets

[josephglanville]

Yeah such a feature would need to be in Helm/Tiller most likely. There is also the problem that PodDisruptionBudget can't respect cross statefulset constrains which means that automated cluster operations could prove problematic.

[andrusha]

It might be worth looking into CRD and Operators:
https://kubernetes.io/docs/tasks/access-kubernetes-api/extend-api-custom-resource-definitions/
https://coreos.com/blog/introducing-operators.html
https://github.com/coreos/etcd-operator

They allow you to instrument your deployments with custom code.

[tekn0ir]

Also consider adding support for k8s as full resource manager, i.e. replacing zookeeper for service discovery etc. It would add neat features like restart strategies for example.

[townsendbt]

Hi. I'm wondering if this issue is being considered any time soon?
Thanks, Brian.

[maver1ck]

@andrusha Going into operator will make this task much more difficult. Simple helm chart will be the best starting point.

[maver1ck]

helm/charts#9314
PS. There is no option for multi-tier configuration in this Chart. And I don't want to add it in initial PR.

[rahulsingh303]

Checkout this one https://github.com/rahulsingh303/druid-helm . It works for me , tested on both minikube and kubernetes cluster.

[maver1ck]

@rahulsingh303 I do a quick look at your chart but I think my PR is superior (i.e. it's using StatefulSets for historical and middlemanagers) and follows the rules of creating helm chart.

[rahulsingh303]

@maver1ck I was trained at Google and learnt that stateful sets are not needed unless you really need it for case where you want to bind your pv to a particular server (may be because you have a license or some other business reason) and hence they recommended not to use as no one uses in 99% of cases . Also my friend this is a community and not about who is superior but rather who can contribute and make a better reusable artifact :) .

[StardustXiaoT]

incubator druid helm charts have no router configurations, how do I add router to the druid cluster?

[rhamonantunes]

@StardustXiaoT Did you manage to do that? I have the same problem!