Charts for pageserver, safekeeper, compute

[patademahesh]

Why are no charts present for pageserver and safekeeper services? We wanted to run a complete Neon DB cluster on K8s cluster.

[kesavkolla]

Any update on this? Has anyone successfully deployed Neon DB in k8s?

[laotoutou]

Any update on this?

[SeinoYuki]

Hi.
There is a significant demand for adding storage components.
We recognize the importance of addressing this issue promptly for the development of the community.
To that end, we have submitted two proof-of-concept (PoC) pull requests to add Safekeeper, Pageserver, and MinIO.
We hope these will be merged as soon as possible, as they do not affect other charts.

The pull requests have dependencies; do not merge the #91 until the Actions (Release Charts) in the #90 are complete.

The procedure to experience this change is as follows. It is a bit complicated and needs improvement. [1]

helm repo add test-chart https://seinoyuki.github.io/helm-charts/
helm repo update
helm install neon1 test-chart/neon-pageserver
helm install neon2 test-chart/neon-safekeeper --set settings.s3.enabled=false --set settings.broker.enabled=false --set settings.s3.fullnameOverride="neon1-neon-minio" --set settings.broker.fullnameOverride="neon1-neon-storage-broker"

The following issue remains.

• Addition of README and NOTE
• Linkage between Pageserver and Storage Controller
• Linkage between MinIO and Neon Storage Scrubber
• Prepare Chart to deploy compute.
  • Neon cluster must be able to be built with one helm install. [1]
• Support sharding of Pageserver

[jcsp]

For the moment, I've added #119 to make it clearer that the helm charts don't give a full working system. We can revise that as/when it evolves.

I'm leaving this issue open as a place for folks who would like to work on other services to collaborate.

A note of caution: safekeepers rely on persistent local disk storage and kubernetes deployments are sometimes not great at protecting the availability of that (e.g. when rescheduling pods after node replacements, you need to think about ensuring the new node has a full copy of data). This makes it relatively easy to write a k8s manifest that brings up an apparently working system, but is may not be not safe to use in production.