Karpenter support #71
Replies: 4 comments 3 replies
-
Awesome! would love to see it come into play. cc @wzshiming |
Beta Was this translation helpful? Give feedback.
-
Awesome! FYI: It looks like karpenter has been using kwok for performance testing. |
Beta Was this translation helpful? Give feedback.
-
The kwok provider included with Karpenter only supports a hard-coded list of node types that don't actually correspond to any "real" node types. If we want to run a scheduling simulation with Karpenter in the same way that we do with Cluster Autoscaler, we somehow need to tell Karpenter what the "real" list of node types it can draw from is... One idea would be to just make Karpenter read from the Another point to consider (this could be an issue with Cluster Autoscaler, too, but probably more-so with Karpenter) is that SimKube and/or KWOK and/or the autoscaler needs to know something about the "simulated" availability of an instance type. Ideally we'd like to be able to simulate scenarios where the "best" or "desired" instance type is unavailable so that we can see what our autoscaler does in response. I'm not 100% sure how to implement this yet. |
Beta Was this translation helpful? Give feedback.
-
The karpenter project now lets you define custom instance types for the KWOK provider, going to close this out. |
Beta Was this translation helpful? Give feedback.
-
Right now we interact with Cluster Autoscaler via an externalgrpc cloud provider. It would be really great if we can also support Karpenter, since there is a lot of interest and excitement about this project right now. If we end up moving to KWOK (#70) we will just write a KWOK provider for karpenter, otherwise we might have to build something inside SimKube itself. This is a tracking issue for this project.
Beta Was this translation helpful? Give feedback.
All reactions