Hack to shuffel host in evacuation request

internal/scheduling/nova/external_scheduler_api.go

@@ -11,6 +11,7 @@ import (
 	"fmt"
 	"io"
 	"log/slog"
+	"math/rand"
 	"net/http"
 
 	api "github.com/cobaltcore-dev/cortex/api/external/nova"
@@ -24,6 +25,15 @@ import (
 	"sigs.k8s.io/controller-runtime/pkg/metrics"
 )
 
+// Custom configuration for the Nova external scheduler api.
+type HTTPAPIConfig struct {
+	// Number of top hosts to shuffle for evacuation requests. Defaults to 3.
+	EvacuationShuffleK int `json:"evacuationShuffleK,omitempty"`
+	// NovaLimitHostsToRequest, if true, will filter the Nova scheduler response
+	// to only include hosts that were in the original request.
+	NovaLimitHostsToRequest bool `json:"novaLimitHostsToRequest,omitempty"`
+}
+
 type HTTPAPIDelegate interface {
 	// Process the decision from the API. Should create and return the updated decision.
 	ProcessNewDecisionFromAPI(ctx context.Context, decision *v1alpha1.Decision) error
@@ -34,12 +44,6 @@ type HTTPAPI interface {
 	Init(*http.ServeMux)
 }
 
-type HTTPAPIConfig struct {
-	// NovaLimitHostsToRequest, if true, will filter the Nova scheduler response
-	// to only include hosts that were in the original request.
-	NovaLimitHostsToRequest bool `json:"novaLimitHostsToRequest,omitempty"`
-}
-
 type httpAPI struct {
 	monitor  scheduling.APIMonitor
 	delegate HTTPAPIDelegate
@@ -116,6 +120,26 @@ func (httpAPI *httpAPI) inferPipelineName(requestData api.ExternalSchedulerReque
 	}
 }
 
+// shuffleTopHosts randomly reorders the first k hosts if the request
+// is an evacuation. This helps distribute evacuated VMs across multiple hosts
+// rather than concentrating them on the single "best" host.
+func shuffleTopHosts(hosts []string, k int) []string {
+	if k <= 0 {
+		k = 3
+	}
+	n := min(k, len(hosts))
+	if n <= 1 {
+		return hosts
+	}
+	result := make([]string, len(hosts))
+	copy(result, hosts)
+	rand.Shuffle(n, func(i, j int) {
+		result[i], result[j] = result[j], result[i]
+	})
+	slog.Info("shuffled top hosts for evacuation", "k", n, "hosts", result[:n])
+	return result
+}
+
 // Limit the external scheduler response to the hosts provided in the  external
 // scheduler request. i.e. don't provide new hosts that weren't in the request,
 // since the Nova scheduler won't know how to handle them.
@@ -235,6 +259,12 @@ func (httpAPI *httpAPI) NovaExternalScheduler(w http.ResponseWriter, r *http.Req
 		slog.Info("limited hosts to request",
 			"hosts", hosts, "originalHosts", decision.Status.Result.OrderedHosts)
 	}
+	// This is a hack to address the problem that Nova only uses the first host in hosts for evacuation requests.
+	// Only for evacuation we shuffle the first k hosts to ensure that we do not get stuck on a single host
+	intent, err := requestData.GetIntent()
+	if err == nil && intent == api.EvacuateIntent {
+		hosts = shuffleTopHosts(hosts, httpAPI.config.EvacuationShuffleK)
+	}
 	response := api.ExternalSchedulerResponse{Hosts: hosts}
 	w.Header().Set("Content-Type", "application/json")
 	if err = json.NewEncoder(w).Encode(response); err != nil {

internal/scheduling/nova/external_scheduler_api_test.go

@@ -506,6 +506,88 @@ func TestLimitHostsToRequest(t *testing.T) {
 	}
 }
 
+func TestShuffleTopHosts(t *testing.T) {
+	tests := []struct {
+		name              string
+		hosts             []string
+		k                 int
+		unchangedTailFrom int // index from which hosts should be unchanged (-1 if all can change)
+	}{
+		{
+			name:  "empty hosts returns empty",
+			hosts: []string{},
+			k:     3,
+		},
+		{
+			name:              "single host returns unchanged",
+			hosts:             []string{"host1"},
+			k:                 3,
+			unchangedTailFrom: 0,
+		},
+		{
+			name:              "shuffles only first k hosts",
+			hosts:             []string{"host1", "host2", "host3", "host4", "host5"},
+			k:                 3,
+			unchangedTailFrom: 3,
+		},
+		{
+			name:              "k=0 uses default of 3",
+			hosts:             []string{"host1", "host2", "host3", "host4", "host5"},
+			k:                 0,
+			unchangedTailFrom: 3,
+		},
+		{
+			name:              "negative k uses default of 3",
+			hosts:             []string{"host1", "host2", "host3", "host4", "host5"},
+			k:                 -1,
+			unchangedTailFrom: 3,
+		},
+		{
+			name:              "k larger than hosts shuffles all",
+			hosts:             []string{"host1", "host2"},
+			k:                 10,
+			unchangedTailFrom: -1,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			original := make([]string, len(tt.hosts))
+			copy(original, tt.hosts)
+
+			result := shuffleTopHosts(tt.hosts, tt.k)
+
+			if len(result) != len(tt.hosts) {
+				t.Fatalf("expected %d hosts, got %d", len(tt.hosts), len(result))
+			}
+			// Verify original slice not modified
+			for i, h := range original {
+				if tt.hosts[i] != h {
+					t.Errorf("original slice modified at %d: expected %s, got %s", i, h, tt.hosts[i])
+				}
+			}
+			// Verify tail unchanged
+			if tt.unchangedTailFrom >= 0 {
+				for i := tt.unchangedTailFrom; i < len(original); i++ {
+					if result[i] != original[i] {
+						t.Errorf("expected host[%d] = %s unchanged, got %s", i, original[i], result[i])
+					}
+				}
+			}
+			// Verify all hosts present
+			hostSet := make(map[string]bool)
+			for _, h := range result {
+				hostSet[h] = true
+			}
+			for _, h := range original {
+				if !hostSet[h] {
+					t.Errorf("host %s missing from result", h)
+				}
+			}
+		})
+	}
+}
+
 func TestHTTPAPI_inferPipelineName(t *testing.T) {
 	delegate := &mockHTTPAPIDelegate{}
 	api := NewAPI(HTTPAPIConfig{}, delegate).(*httpAPI)

internal/scheduling/nova/integration_test.go

@@ -283,7 +283,9 @@ func NewIntegrationTestServer(t *testing.T, pipelineConfig PipelineConfig, objec
 	controller.PipelineConfigs[testPipeline.Name] = testPipeline
 
 	// Create the HTTP API with the controller as delegate - skip metrics registration
+	// Set EvacuationShuffleK=1 to disable shuffle for deterministic test results
 	api := &httpAPI{
+		config:   HTTPAPIConfig{EvacuationShuffleK: 1},
 		monitor:  lib.NewSchedulerMonitor(), // Create new monitor but don't register
 		delegate: controller,
 	}