fix: Fix GPU instance filtering when --instance-selector-gpus=0 (#8506)

Fix GPU instance filtering when --instance-selector-gpus=0

- Add workaround for AWS API bug where g6f instances incorrectly report GPU count as 0
- Filter out GPU instances using eksctl's local instance types map when GPUs=0 is specified

The AWS DescribeInstanceTypes API incorrectly reports GPU count as 0 for g6f family
instances (g6f.large, g6f.xlarge, g6f.2xlarge, g6f.4xlarge) even though they have
1 L4 GPU each. This caused the instance selector to include these GPU instances
when users specified --instance-selector-gpus=0 (expecting only non-GPU instances).

This fix adds an additional filtering step that uses eksctl's accurate local
instance types map to remove GPU instances when GPUs=0 is specified, while
preserving existing behavior for all other cases.

Author: sapphirew

pkg/eks/nodegroup_service.go

@@ -20,6 +20,7 @@ import (
 	"github.com/weaveworks/eksctl/pkg/cfn/manager"
 	"github.com/weaveworks/eksctl/pkg/outposts"
 	"github.com/weaveworks/eksctl/pkg/ssh"
+	"github.com/weaveworks/eksctl/pkg/utils/instance"
 	"github.com/weaveworks/eksctl/pkg/utils/tasks"
 )
 
@@ -240,6 +241,22 @@ func (n *NodeGroupService) expandInstanceSelector(ins *api.InstanceSelector, azs
 		return nil, errors.New("instance selector criteria matched no instances; consider broadening your criteria so that more instance types are returned")
 	}
 
+	// Workaround for AWS API bug: filter out GPU instances when GPUs=0 is specified
+	// AWS incorrectly reports GPU count as 0 for some GPU instances (e.g., g6f family)
+	if ins.GPUs != nil && *ins.GPUs == 0 {
+		filteredTypes := make([]string, 0, len(instanceTypes))
+		for _, instanceType := range instanceTypes {
+			if !instance.IsGPUInstanceType(instanceType) {
+				filteredTypes = append(filteredTypes, instanceType)
+			}
+		}
+		instanceTypes = filteredTypes
+
+		if len(instanceTypes) == 0 {
+			return nil, errors.New("instance selector criteria matched no non-GPU instances; consider broadening your criteria")
+		}
+	}
+
 	return instanceTypes, nil
 }
 

pkg/eks/nodegroup_service_gpu_test.go

@@ -0,0 +1,97 @@
+package eks
+
+import (
+	"context"
+	"testing"
+
+	"github.com/aws/amazon-ec2-instance-selector/v3/pkg/selector"
+	"github.com/stretchr/testify/assert"
+
+	api "github.com/weaveworks/eksctl/pkg/apis/eksctl.io/v1alpha5"
+)
+
+// MockInstanceSelector for testing
+type MockInstanceSelector struct {
+	returnInstances []string
+	returnError     error
+}
+
+func (m *MockInstanceSelector) Filter(ctx context.Context, filters selector.Filters) ([]string, error) {
+	return m.returnInstances, m.returnError
+}
+
+func TestExpandInstanceSelector_GPUFiltering(t *testing.T) {
+	tests := []struct {
+		name                string
+		instanceSelector    *api.InstanceSelector
+		mockReturnInstances []string
+		expectedInstances   []string
+		expectError         bool
+	}{
+		{
+			name: "filters out GPU instances when GPUs=0",
+			instanceSelector: &api.InstanceSelector{
+				VCPUs: 8,
+				GPUs:  newIntPtr(0),
+			},
+			mockReturnInstances: []string{"m5.2xlarge", "g6f.2xlarge", "c5.2xlarge", "g4dn.xlarge"},
+			expectedInstances:   []string{"m5.2xlarge", "c5.2xlarge"},
+			expectError:         false,
+		},
+		{
+			name: "includes GPU instances when GPUs=1",
+			instanceSelector: &api.InstanceSelector{
+				VCPUs: 8,
+				GPUs:  newIntPtr(1),
+			},
+			mockReturnInstances: []string{"m5.2xlarge", "g6f.2xlarge", "c5.2xlarge", "g4dn.xlarge"},
+			expectedInstances:   []string{"m5.2xlarge", "g6f.2xlarge", "c5.2xlarge", "g4dn.xlarge"},
+			expectError:         false,
+		},
+		{
+			name: "no filtering when GPUs is nil",
+			instanceSelector: &api.InstanceSelector{
+				VCPUs: 8,
+			},
+			mockReturnInstances: []string{"m5.2xlarge", "g6f.2xlarge", "c5.2xlarge"},
+			expectedInstances:   []string{"m5.2xlarge", "g6f.2xlarge", "c5.2xlarge"},
+			expectError:         false,
+		},
+		{
+			name: "error when all instances are filtered out",
+			instanceSelector: &api.InstanceSelector{
+				VCPUs: 8,
+				GPUs:  newIntPtr(0),
+			},
+			mockReturnInstances: []string{"g6f.2xlarge", "g4dn.xlarge"},
+			expectedInstances:   nil,
+			expectError:         true,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			mockSelector := &MockInstanceSelector{
+				returnInstances: tt.mockReturnInstances,
+			}
+
+			service := &NodeGroupService{
+				instanceSelector: mockSelector,
+			}
+
+			result, err := service.expandInstanceSelector(tt.instanceSelector, []string{"us-west-2a"})
+
+			if tt.expectError {
+				assert.Error(t, err)
+				assert.Nil(t, result)
+			} else {
+				assert.NoError(t, err)
+				assert.Equal(t, tt.expectedInstances, result)
+			}
+		})
+	}
+}
+
+func newIntPtr(i int) *int {
+	return &i
+}