Repository navigation
Give the cluster functions' Helm releases fixed names - #587
Conversation
|
/review |
There was a problem hiding this comment.
I think this needs one design decision before it's close. Naming each release mp-<chart> fixes #506 and matches compose-serving-stack, but most of the change keeps old release names on existing clusters, which I've asked about inline.
#594 adds an UpCloud cluster function whose autoscaler Release doesn't set an external name either, so whichever PR lands second will need the same fixed name.
The commit message says v0.4.0 named the cluster with a cluster- prefix, but I couldn't find that prefix in v0.4.0's compose-inference-cluster. Could that clause go?
Reviewed by modelplane-review d7f2116. Reply, or react 👍 or 👎, on each comment to help tune it.
provider-helm installs a Release under its external name, which defaults to the Release's generated metadata.name: the XR's name plus a hash suffix. The cluster functions compose their Helm releases (the autoscaler on EKS and Civo, the EFA DRA driver on EKS, the network operator on AKS, the CSI driver on Nebius) without setting one, so an InferenceCluster with a long name gets a release name past Helm's 53-character limit and the release fails to install. That's any name of 41 characters or more. This names each release mp-<chart>, like compose-serving-stack's. Each release runs on a cluster of its own, so a fixed name can't collide. Fixes modelplaneai#506. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: humoflife <markus.schweig@upbound.io>
807dfee to
c161eca
Compare
negz
left a comment
There was a problem hiding this comment.
Thanks for picking this up @humoflife!
Description of your changes
provider-helm installs a Release under its external name, which defaults to the Release's generated
metadata.name: the XR's name plus a hash suffix. The cluster functions compose their Helm releases (the autoscaler on EKS and Civo, the EFA DRA driver on EKS, the network operator on AKS, the CSI driver on Nebius) without setting one, so an InferenceCluster with a long name gets a release name past Helm's 53-character limit and the release never installs. That's any InferenceCluster name of 41 characters or more.This names each release
mp-<chart>, the convention compose-serving-stack already uses. Each release runs on a cluster of its own, so a fixed name can't collide. A cluster composed before this change gets a second copy of each chart under the new name.On a local control plane with a real EKS cluster in us-west-2, an InferenceCluster named
orchestration-orchestrating-cluster-second(42 characters) failed on main withRelease name is invalid: orchestration-orchestrating-cluster-second-53da71414945, and its autoscaler installed asmp-cluster-autoscalerwith this change. That run was on an earlier revision of this PR, which also kept already-installed names; the code path for a new Release is the same. The AKS, Nebius and Civo releases and the EKS EFA driver are covered by composition tests only.Fixes #506
I have:
nix flake check(or./nix.sh flake check) and made sure it passes.git commit -s.🤖 Generated with Claude Code