test for model deletion steps

Author: domsolutions

…/model/explicit_model_deployment.feature …es/model/custom_model_deployment.featuretests/integration/godog/features/model/explicit_model_deployment.feature renamed to tests/integration/godog/features/model/custom_model_deployment.feature tests/integration/godog/features/model/explicit_model_deployment.feature renamed to tests/integration/godog/features/model/custom_model_deployment.feature

@@ -1,24 +1,25 @@
-@ModelDeployment @Functional @Models @Explicit
+@ModelDeployment @Functional @Models @CustomModelSpec
 Feature: Explicit Model deployment
   I deploy a custom model spec, wait for model to be deployed to the servers
-  and send an inference request to that model
+  and send an inference request to that model and expect a successful response.
+  I then delete the model and send inference requests and expect them to fail.
 
   Scenario: Load model and send inference request to envoy
-    Given I deploy model spec:
+    Given I deploy model spec with timeout "10s":
     """
     apiVersion: mlops.seldon.io/v1alpha1
     kind: Model
     metadata:
-      name: iris
+      name: alpha-1
     spec:
       replicas: 1
       requirements:
       - sklearn
       - mlserver
       storageUri: gs://seldon-models/scv2/samples/mlserver_1.3.5/iris-sklearn
     """
-    When the model "iris" should eventually become Ready with timeout "20s"
-    Then send HTTP inference request with timeout "20s" to model "iris" with payload:
+    When the model "alpha-1" should eventually become Ready with timeout "20s"
+    Then send HTTP inference request with timeout "20s" to model "alpha-1" with payload:
     """
     {
         "inputs": [
@@ -51,7 +52,7 @@ Feature: Explicit Model deployment
     }
   ] }
     """
-    Then send gRPC inference request with timeout "20s" to model "iris" with payload:
+    Then send gRPC inference request with timeout "20s" to model "alpha-1" with payload:
     """
     {
         "inputs": [
@@ -82,4 +83,35 @@ Feature: Explicit Model deployment
       "contents": {"int64_contents" : [2]}
     }
   ] }
-    """
+    """
+    Then delete the model "alpha-1" with timeout "10s"
+    Then send HTTP inference request with timeout "20s" to model "alpha-1" with payload:
+    """
+    {
+        "inputs": [
+          {
+            "name": "predict",
+            "shape": [1, 4],
+            "datatype": "FP32",
+            "data": [[1, 2, 3, 4]]
+          }
+        ]
+    }
+    """
+    And expect http response status code "404"
+    Then send gRPC inference request with timeout "20s" to model "alpha-1" with payload:
+    """
+    {
+        "inputs": [
+          {
+            "name": "predict",
+            "shape": [1, 4],
+            "datatype": "FP32",
+            "contents": {
+              "int64_contents" : [1, 2, 3, 4]
+            }
+          }
+        ]
+    }
+    """
+    And expect gRPC response error to contain "Unimplemented"

tests/integration/godog/steps/custom_model_steps.go

@@ -0,0 +1,114 @@
+/*
+Copyright (c) 2024 Seldon Technologies Ltd.
+
+Use of this software is governed BY
+(1) the license included in the LICENSE file or
+(2) if the license included in the LICENSE file is the Business Source License 1.1,
+the Change License after the Change Date as each is defined in accordance with the LICENSE file.
+*/
+
+package steps
+
+import (
+	"context"
+	"errors"
+	"fmt"
+
+	"github.com/seldonio/seldon-core/operator/v2/apis/mlops/v1alpha1"
+	k8serrors "k8s.io/apimachinery/pkg/api/errors"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/watch"
+)
+
+// deleteModel we have to wait for model to be deleted, as there is a finalizer attached so the scheduler can confirm
+// when model has been unloaded from inference pod, model-gw, dataflow-engine, pipeline-gw and controller will remove
+// finalizer so deletion can complete.
+func (m *Model) deleteModel(ctx context.Context, model string) error {
+	modelCR, err := m.k8sClient.MlopsV1alpha1().Models(m.namespace).Get(ctx, model, metav1.GetOptions{})
+	if err != nil {
+		if k8serrors.IsNotFound(err) {
+			return fmt.Errorf("model %s can't be deleted, does not exist", model)
+		}
+		return fmt.Errorf("failed to get model %s", model)
+	}
+
+	if err := m.k8sClient.MlopsV1alpha1().Models(m.namespace).Delete(ctx, model, metav1.DeleteOptions{}); err != nil {
+		return fmt.Errorf("failed deleting model: %w", err)
+	}
+	m.log.Debugf("Delete request for model %s sent", model)
+
+	watcher, err := m.k8sClient.MlopsV1alpha1().Models(m.namespace).Watch(ctx, metav1.ListOptions{
+		FieldSelector:   fmt.Sprintf("metadata.name=%s", model),
+		ResourceVersion: modelCR.ResourceVersion,
+	})
+	if err != nil {
+		return fmt.Errorf("failed watching model: %w", err)
+	}
+	defer watcher.Stop()
+
+	m.log.Debugf("Waiting for %s model deletion confirmation", model)
+
+	for {
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case event, ok := <-watcher.ResultChan():
+			if !ok {
+				return errors.New("watcher channel closed")
+			}
+			if event.Type == watch.Error {
+				return fmt.Errorf("watch error: %v", event.Object)
+			}
+			if event.Type == watch.Deleted {
+				return nil
+			}
+		}
+	}
+}
+
+func (m *Model) waitForModelReady(ctx context.Context, model string) error {
+	foundModel, err := m.k8sClient.MlopsV1alpha1().Models(m.namespace).Get(ctx, model, metav1.GetOptions{})
+	if err != nil {
+		return fmt.Errorf("failed getting model: %w", err)
+	}
+
+	if foundModel.Status.IsReady() {
+		return nil
+	}
+
+	watcher, err := m.k8sClient.MlopsV1alpha1().Models(m.namespace).Watch(ctx, metav1.ListOptions{
+		FieldSelector:   fmt.Sprintf("metadata.name=%s", model),
+		ResourceVersion: foundModel.ResourceVersion,
+		Watch:           true,
+	})
+	if err != nil {
+		return fmt.Errorf("failed subscribed to watch model: %w", err)
+	}
+	defer watcher.Stop()
+
+	for {
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case event, ok := <-watcher.ResultChan():
+			if !ok {
+				return fmt.Errorf("watch channel closed")
+			}
+
+			if event.Type == watch.Error {
+				return fmt.Errorf("watch error: %v", event.Object)
+			}
+
+			if event.Type == watch.Added || event.Type == watch.Modified {
+				model := event.Object.(*v1alpha1.Model)
+				if model.Status.IsReady() {
+					return nil
+				}
+			}
+
+			if event.Type == watch.Deleted {
+				return fmt.Errorf("resource was deleted")
+			}
+		}
+	}
+}

tests/integration/godog/steps/explicit_model_steps.go

@@ -61,10 +61,10 @@ func (i *inference) sendGRPCModelInferenceRequest(ctx context.Context, model str
 	ctx = metadata.NewOutgoingContext(context.Background(), md)
 	resp, err := i.grpc.ModelInfer(ctx, msg)
 	if err != nil {
-		return fmt.Errorf("could not send grpc model inference: %w", err)
+		i.lastGRPCResponse.err = err
 	}
 
-	i.lastGRPCResponse = resp
+	i.lastGRPCResponse.response = resp
 	return nil
 }
 
@@ -126,12 +126,24 @@ func jsonContainsObjectSubset(jsonStr, needleStr string) (bool, error) {
 	return containsSubset(needle, hay), nil
 }
 
+func (i *inference) gRPCRespContainsError(err string) error {
+	if i.lastGRPCResponse.err == nil {
+		return errors.New("no gRPC response error found")
+	}
+
+	if strings.Contains(i.lastGRPCResponse.err.Error(), err) {
+		return nil
+	}
+
+	return fmt.Errorf("error %s does not contain %s", i.lastGRPCResponse.err.Error(), err)
+}
+
 func (i *inference) gRPCRespCheckBodyContainsJSON(expectJSON *godog.DocString) error {
-	if i.lastGRPCResponse == nil {
+	if i.lastGRPCResponse.response == nil {
 		return errors.New("no gRPC response found")
 	}
 
-	gotJson, err := json.Marshal(i.lastGRPCResponse)
+	gotJson, err := json.Marshal(i.lastGRPCResponse.response)
 	if err != nil {
 		return fmt.Errorf("could not marshal gRPC json: %w", err)
 	}