From 17992ad52c6c0401add7cd88d7b9cd4c4d94ac35 Mon Sep 17 00:00:00 2001 From: ChethanUK Date: Tue, 17 Nov 2026 12:40:40 +0800 Subject: [PATCH] Document that EKS GPU and CPU pools can't scale up from zero The upstream cluster-autoscaler's AWS provider doesn't model the DRA ResourceSlices a node would publish, so an EKS node group at zero never wakes for a pod whose GPU or CPU claim binds through DRA. The AKS and CoreWeave XRDs already warn about the same limit on minNodeCount; EKSCluster didn't, and InferenceCluster's minNodeCount had no description at all. The EKSCluster reference example and the Llama 3.1 8B EKS recipe went further and showed a GPU pool that scales to zero. This adds the warning to EKSCluster and InferenceCluster minNodeCount, regenerates the models and schema lock, and keeps the EKS examples at one node. InferenceCluster lists AKS for GPU pools only, matching AKSCluster. It doesn't reject minNodeCount: 0, since the issue hasn't settled which providers that should cover. Towards #345. Signed-off-by: ChethanUK --- apis/eksclusters/definition.yaml | 8 ++++++-- apis/inferenceclusters/definition.yaml | 7 +++++++ .../recipes/llama-3.1-8b/inference-cluster-eks.yaml | 2 +- docs/manifests/reference/eksclusters.yaml | 5 ++++- schemas/.lock.json | 2 +- .../models/ai/modelplane/inferencecluster/v1alpha1.py | 3 +++ .../ai/modelplane/infrastructure/ekscluster/v1alpha1.py | 2 +- 7 files changed, 23 insertions(+), 6 deletions(-) diff --git a/apis/eksclusters/definition.yaml b/apis/eksclusters/definition.yaml index e3718703b..2abeaac9b 100644 --- a/apis/eksclusters/definition.yaml +++ b/apis/eksclusters/definition.yaml @@ -162,8 +162,12 @@ spec: type: integer default: 0 description: >- - Minimum number of nodes for autoscaling. Set to 1 or - higher for groups that must always be available. + Minimum number of nodes for autoscaling. Set it to 1 + or higher for groups that must always be available, + and for GPU and CPU groups: the upstream cluster + autoscaler's AWS provider can't scale a group whose + devices bind through DRA up from zero nodes, so a pod + scheduled to an empty group waits forever. minimum: 0 maximum: 1000 maxNodeCount: diff --git a/apis/inferenceclusters/definition.yaml b/apis/inferenceclusters/definition.yaml index da4b159df..0249c9381 100644 --- a/apis/inferenceclusters/definition.yaml +++ b/apis/inferenceclusters/definition.yaml @@ -528,6 +528,13 @@ spec: minimum: 0 minNodeCount: type: integer + description: >- + Minimum node count for autoscaling. Keep GPU pools on + EKS, AKS and CoreWeave, and CPU pools on EKS and + CoreWeave, at 1 or higher: their cluster autoscalers + can't scale a pool whose devices bind through DRA up + from zero nodes, so a pod scheduled to an empty pool + waits forever. minimum: 0 maxNodeCount: type: integer diff --git a/docs/manifests/recipes/llama-3.1-8b/inference-cluster-eks.yaml b/docs/manifests/recipes/llama-3.1-8b/inference-cluster-eks.yaml index 8ce473601..c31e23d04 100644 --- a/docs/manifests/recipes/llama-3.1-8b/inference-cluster-eks.yaml +++ b/docs/manifests/recipes/llama-3.1-8b/inference-cluster-eks.yaml @@ -19,5 +19,5 @@ spec: nodeCount: 1 zones: - us-west-2a - minNodeCount: 0 + minNodeCount: 1 maxNodeCount: 4 diff --git a/docs/manifests/reference/eksclusters.yaml b/docs/manifests/reference/eksclusters.yaml index dadc5eedb..3ef68892b 100644 --- a/docs/manifests/reference/eksclusters.yaml +++ b/docs/manifests/reference/eksclusters.yaml @@ -15,7 +15,10 @@ spec: role: GPU instanceType: g6.12xlarge diskSizeGb: 200 - nodeCount: 0 + # GPU groups need at least one node: the cluster autoscaler cannot + # scale a group whose GPUs bind through DRA up from zero. + nodeCount: 1 + minNodeCount: 1 maxNodeCount: 8 zones: [us-west-2a, us-west-2b] gpu: diff --git a/schemas/.lock.json b/schemas/.lock.json index f5e193428..fa75db8dc 100644 --- a/schemas/.lock.json +++ b/schemas/.lock.json @@ -1,6 +1,6 @@ { "packages": { - "fs://apis": "7765fe0a737410e9a3f3504455b088e8d2c9ba91fb33046a56acb5441f1ce741", + "fs://apis": "44bfccfab3ee8fdaccac6139473f1f96bb3f1a7d21fc3709b962bb6f3599b962", "git://https://github.com/crossplane/crossplane/cluster/crds": "14c53aac914d393b5bcc00962fbc90d4be5d35bc", "git://https://github.com/modelplaneai/cluster-agent/cluster/charts/cluster-agent/files/crds": "a833d3f6f7d9a6ac6083e7453375b58b65ee741a", "xpkg://xpkg.upbound.io/upbound/provider-aws-ec2:v2.9.4": "sha256:7a0febdfb8da2d59a9fdcdbc2d476811f00566fcb941880313fc2034b0ca6177", diff --git a/schemas/python/models/ai/modelplane/inferencecluster/v1alpha1.py b/schemas/python/models/ai/modelplane/inferencecluster/v1alpha1.py index 09fcebc51..7a806c981 100644 --- a/schemas/python/models/ai/modelplane/inferencecluster/v1alpha1.py +++ b/schemas/python/models/ai/modelplane/inferencecluster/v1alpha1.py @@ -270,6 +270,9 @@ class NodePool(BaseModel): Maximum node count for autoscaling. Omit for fixed-size pools. """ minNodeCount: conint(ge=0) | None = None + """ + Minimum node count for autoscaling. Keep GPU pools on EKS, AKS and CoreWeave, and CPU pools on EKS and CoreWeave, at 1 or higher: their cluster autoscalers can't scale a pool whose devices bind through DRA up from zero nodes, so a pod scheduled to an empty pool waits forever. + """ name: constr(pattern=r'^[a-z0-9]([-a-z0-9]*[a-z0-9])?$', max_length=40) """ Name of the pool. Its nodes carry it as the value of the modelplane.ai/pool label, and the cluster agent observes the pool through an object of this name. diff --git a/schemas/python/models/ai/modelplane/infrastructure/ekscluster/v1alpha1.py b/schemas/python/models/ai/modelplane/infrastructure/ekscluster/v1alpha1.py index 37ce4eefa..0d12af18c 100644 --- a/schemas/python/models/ai/modelplane/infrastructure/ekscluster/v1alpha1.py +++ b/schemas/python/models/ai/modelplane/infrastructure/ekscluster/v1alpha1.py @@ -117,7 +117,7 @@ class NodePool(BaseModel): """ minNodeCount: conint(ge=0, le=1000) | None = 0 """ - Minimum number of nodes for autoscaling. Set to 1 or higher for groups that must always be available. + Minimum number of nodes for autoscaling. Set it to 1 or higher for groups that must always be available, and for GPU and CPU groups: the upstream cluster autoscaler's AWS provider can't scale a group whose devices bind through DRA up from zero nodes, so a pod scheduled to an empty group waits forever. """ name: constr(min_length=1, max_length=40) """