fix: change upsert duplicate PK behavior from dedup to error (#45997)

issue: #44320

Replace the DeduplicateFieldData function with CheckDuplicatePkExist
that returns an error when duplicate primary keys are detected in the
same batch, instead of silently deduplicating.

Changes:
- Replace DeduplicateFieldData with CheckDuplicatePkExist in util.go
- Update upsertTask.PreExecute to return error on duplicate PKs
- Simplify helper function from findLastOccurrenceIndices to
hasDuplicates
- Update unit tests to verify the new error behavior
- Add Python integration tests for duplicate PK error cases

Signed-off-by: Wei Liu <wei.liu@zilliz.com>

internal/proxy/task_upsert.go

@@ -1078,18 +1078,13 @@ func (it *upsertTask) PreExecute(ctx context.Context) error {
 		log.Warn("fail to get primary field schema", zap.Error(err))
 		return err
 	}
-	deduplicatedFieldsData, newNumRows, err := DeduplicateFieldData(primaryFieldSchema, it.req.GetFieldsData(), schema)
+	duplicate, err := CheckDuplicatePkExist(primaryFieldSchema, it.req.GetFieldsData())
 	if err != nil {
-		log.Warn("fail to deduplicate upsert data", zap.Error(err))
+		log.Warn("fail to check duplicate primary keys", zap.Error(err))
+		return err
 	}
-
+	if duplicate {
-	// dedup won't decrease numOfRows to 0
+		return merr.WrapErrParameterInvalidMsg("duplicate primary keys are not allowed in the same batch")
-	if newNumRows > 0 && newNumRows != it.req.NumRows {
-		log.Info("upsert data deduplicated",
-			zap.Uint32("original_num_rows", it.req.NumRows),
-			zap.Uint32("deduplicated_num_rows", newNumRows))
-		it.req.FieldsData = deduplicatedFieldsData
-		it.req.NumRows = newNumRows
 	}
 
 	it.upsertMsg = &msgstream.UpsertMsg{

internal/proxy/task_upsert_test.go

@@ -35,7 +35,6 @@ import (
 	"github.com/milvus-io/milvus/internal/proxy/shardclient"
 	"github.com/milvus-io/milvus/internal/util/function/embedding"
 	"github.com/milvus-io/milvus/internal/util/segcore"
-	"github.com/milvus-io/milvus/pkg/v2/common"
 	"github.com/milvus-io/milvus/pkg/v2/mq/msgstream"
 	"github.com/milvus-io/milvus/pkg/v2/proto/planpb"
 	"github.com/milvus-io/milvus/pkg/v2/proto/rootcoordpb"
@@ -1467,402 +1466,137 @@ func TestGenNullableFieldData_GeometryAndTimestamptz(t *testing.T) {
 	})
 }
 
-func TestUpsertTask_PlanNamespace_AfterPreExecute(t *testing.T) {
+func TestUpsertTask_DuplicatePK_Int64(t *testing.T) {
-	mockey.PatchConvey("TestUpsertTask_PlanNamespace_AfterPreExecute", t, func() {
+	schema := &schemapb.CollectionSchema{
-		// Setup global meta cache and common mocks
+		Name: "test_duplicate_pk",
-		globalMetaCache = &MetaCache{}
+		Fields: []*schemapb.FieldSchema{
-		mockey.Mock(GetReplicateID).Return("", nil).Build()
+			{FieldID: 100, Name: "id", IsPrimaryKey: true, DataType: schemapb.DataType_Int64},
-		mockey.Mock((*MetaCache).GetCollectionID).Return(int64(1001), nil).Build()
+			{FieldID: 101, Name: "value", DataType: schemapb.DataType_Int32},
-		mockey.Mock((*MetaCache).GetCollectionInfo).Return(&collectionInfo{updateTimestamp: 12345}, nil).Build()
+		},
-		mockey.Mock((*MetaCache).GetPartitionInfo).Return(&partitionInfo{name: "_default"}, nil).Build()
+	}
-		mockey.Mock((*MetaCache).GetPartitionID).Return(int64(1002), nil).Build()
-		mockey.Mock(isPartitionKeyMode).Return(false, nil).Build()
-		mockey.Mock(validatePartitionTag).Return(nil).Build()
 
-		// Schema with namespace enabled
+	// Data with duplicate primary keys: 1, 2, 1 (duplicate)
-		mockey.Mock((*MetaCache).GetCollectionSchema).To(func(_ *MetaCache, _ context.Context, _ string, _ string) (*schemaInfo, error) {
+	fieldsData := []*schemapb.FieldData{
-			info := createTestSchema()
-			info.CollectionSchema.Properties = append(info.CollectionSchema.Properties, &commonpb.KeyValuePair{Key: common.NamespaceEnabledKey, Value: "true"})
-			return info, nil
-		}).Build()
-
-		// Capture plan to verify namespace
-		var capturedPlan *planpb.PlanNode
-		mockey.Mock(planparserv2.CreateRequeryPlan).To(func(_ *schemapb.FieldSchema, _ *schemapb.IDs) *planpb.PlanNode {
-			capturedPlan = &planpb.PlanNode{}
-			return capturedPlan
-		}).Build()
-
-		// Mock query to return a valid result for queryPreExecute merge path
-		mockey.Mock((*Proxy).query).Return(&milvuspb.QueryResults{
-			Status: merr.Success(),
-			FieldsData: []*schemapb.FieldData{
 		{
 			FieldName: "id",
 			FieldId:   100,
 			Type:      schemapb.DataType_Int64,
-					Field:     &schemapb.FieldData_Scalars{Scalars: &schemapb.ScalarField{Data: &schemapb.ScalarField_LongData{LongData: &schemapb.LongArray{Data: []int64{1, 2}}}}},
-				},
-				{
-					FieldName: "name",
-					FieldId:   102,
-					Type:      schemapb.DataType_VarChar,
-					Field:     &schemapb.FieldData_Scalars{Scalars: &schemapb.ScalarField{Data: &schemapb.ScalarField_StringData{StringData: &schemapb.StringArray{Data: []string{"old1", "old2"}}}}},
-				},
-				{
-					FieldName: "vector",
-					FieldId:   101,
-					Type:      schemapb.DataType_FloatVector,
-					Field:     &schemapb.FieldData_Vectors{Vectors: &schemapb.VectorField{Dim: 128, Data: &schemapb.VectorField_FloatVector{FloatVector: &schemapb.FloatArray{Data: make([]float32, 256)}}}},
-				},
-			},
-		}, segcore.StorageCost{}, nil).Build()
-
-		// Build task
-		task := createTestUpdateTask()
-		ns := "ns-1"
-		task.req.PartialUpdate = true
-		task.req.Namespace = &ns
-
-		// Skip insert/delete heavy logic
-		mockey.Mock((*upsertTask).insertPreExecute).Return(nil).Build()
-		mockey.Mock((*upsertTask).deletePreExecute).Return(nil).Build()
-
-		err := task.PreExecute(context.Background())
-		assert.NoError(t, err)
-		assert.NotNil(t, capturedPlan)
-		assert.NotNil(t, capturedPlan.Namespace)
-		assert.Equal(t, *task.req.Namespace, *capturedPlan.Namespace)
-	})
-}
-
-func TestUpsertTask_Deduplicate_Int64PK(t *testing.T) {
-	// Test deduplication with Int64 primary key
-	primaryFieldSchema := &schemapb.FieldSchema{
-		Name:         "id",
-		FieldID:      100,
-		DataType:     schemapb.DataType_Int64,
-		IsPrimaryKey: true,
-	}
-
-	collSchema := &schemapb.CollectionSchema{
-		Fields: []*schemapb.FieldSchema{
-			primaryFieldSchema,
-			{
-				Name:     "float_field",
-				FieldID:  101,
-				DataType: schemapb.DataType_Float,
-			},
-		},
-	}
-	schema := newSchemaInfo(collSchema)
-
-	// Create field data with duplicate IDs: [1, 2, 3, 2, 1]
-	// Expected to keep last occurrence of each: [3, 2, 1] (indices 2, 3, 4)
-	fieldsData := []*schemapb.FieldData{
-		{
-			FieldName: "id",
-			Type:      schemapb.DataType_Int64,
 			Field: &schemapb.FieldData_Scalars{
 				Scalars: &schemapb.ScalarField{
 					Data: &schemapb.ScalarField_LongData{
-						LongData: &schemapb.LongArray{
+						LongData: &schemapb.LongArray{Data: []int64{1, 2, 1}},
-							Data: []int64{1, 2, 3, 2, 1},
-						},
 					},
 				},
 			},
 		},
 		{
-			FieldName: "float_field",
+			FieldName: "value",
-			Type:      schemapb.DataType_Float,
+			FieldId:   101,
+			Type:      schemapb.DataType_Int32,
 			Field: &schemapb.FieldData_Scalars{
 				Scalars: &schemapb.ScalarField{
-					Data: &schemapb.ScalarField_FloatData{
+					Data: &schemapb.ScalarField_IntData{
-						FloatData: &schemapb.FloatArray{
+						IntData: &schemapb.IntArray{Data: []int32{100, 200, 300}},
-							Data: []float32{1.1, 2.2, 3.3, 2.4, 1.5},
-						},
 					},
 				},
 			},
 		},
 	}
 
-	deduplicatedFields, newNumRows, err := DeduplicateFieldData(primaryFieldSchema, fieldsData, schema)
+	// Test CheckDuplicatePkExist directly
+	primaryFieldSchema, err := typeutil.GetPrimaryFieldSchema(schema)
 	assert.NoError(t, err)
-	assert.Equal(t, uint32(3), newNumRows)
+	hasDuplicate, err := CheckDuplicatePkExist(primaryFieldSchema, fieldsData)
-	assert.Equal(t, 2, len(deduplicatedFields))
+	assert.NoError(t, err)
-
+	assert.True(t, hasDuplicate, "should detect duplicate primary keys")
-	// Check deduplicated primary keys
-	pkField := deduplicatedFields[0]
-	pkData := pkField.GetScalars().GetLongData().GetData()
-	assert.Equal(t, 3, len(pkData))
-	assert.Equal(t, []int64{3, 2, 1}, pkData)
-
-	// Check corresponding float values (should be 3.3, 2.4, 1.5)
-	floatField := deduplicatedFields[1]
-	floatData := floatField.GetScalars().GetFloatData().GetData()
-	assert.Equal(t, 3, len(floatData))
-	assert.Equal(t, []float32{3.3, 2.4, 1.5}, floatData)
 }
 
-func TestUpsertTask_Deduplicate_VarCharPK(t *testing.T) {
+func TestUpsertTask_DuplicatePK_VarChar(t *testing.T) {
-	// Test deduplication with VarChar primary key
+	schema := &schemapb.CollectionSchema{
-	primaryFieldSchema := &schemapb.FieldSchema{
+		Name: "test_duplicate_pk_varchar",
-		Name:         "id",
-		FieldID:      100,
-		DataType:     schemapb.DataType_VarChar,
-		IsPrimaryKey: true,
-	}
-
-	collSchema := &schemapb.CollectionSchema{
 		Fields: []*schemapb.FieldSchema{
-			primaryFieldSchema,
+			{FieldID: 100, Name: "id", IsPrimaryKey: true, DataType: schemapb.DataType_VarChar, TypeParams: []*commonpb.KeyValuePair{{Key: "max_length", Value: "100"}}},
-			{
+			{FieldID: 101, Name: "value", DataType: schemapb.DataType_Int32},
-				Name:     "int_field",
-				FieldID:  101,
-				DataType: schemapb.DataType_Int64,
-			},
 		},
 	}
-	schema := newSchemaInfo(collSchema)
 
-	// Create field data with duplicate IDs: ["a", "b", "c", "b", "a"]
+	// Data with duplicate primary keys: "a", "b", "a" (duplicate)
-	// Expected to keep last occurrence of each: ["c", "b", "a"] (indices 2, 3, 4)
 	fieldsData := []*schemapb.FieldData{
 		{
 			FieldName: "id",
+			FieldId:   100,
 			Type:      schemapb.DataType_VarChar,
 			Field: &schemapb.FieldData_Scalars{
 				Scalars: &schemapb.ScalarField{
 					Data: &schemapb.ScalarField_StringData{
-						StringData: &schemapb.StringArray{
+						StringData: &schemapb.StringArray{Data: []string{"a", "b", "a"}},
-							Data: []string{"a", "b", "c", "b", "a"},
-						},
 					},
 				},
 			},
 		},
 		{
-			FieldName: "int_field",
+			FieldName: "value",
-			Type:      schemapb.DataType_Int64,
+			FieldId:   101,
+			Type:      schemapb.DataType_Int32,
 			Field: &schemapb.FieldData_Scalars{
 				Scalars: &schemapb.ScalarField{
-					Data: &schemapb.ScalarField_LongData{
+					Data: &schemapb.ScalarField_IntData{
-						LongData: &schemapb.LongArray{
+						IntData: &schemapb.IntArray{Data: []int32{100, 200, 300}},
-							Data: []int64{100, 200, 300, 201, 101},
-						},
 					},
 				},
 			},
 		},
 	}
 
-	deduplicatedFields, newNumRows, err := DeduplicateFieldData(primaryFieldSchema, fieldsData, schema)
+	// Test CheckDuplicatePkExist directly
+	primaryFieldSchema, err := typeutil.GetPrimaryFieldSchema(schema)
 	assert.NoError(t, err)
-	assert.Equal(t, uint32(3), newNumRows)
+	hasDuplicate, err := CheckDuplicatePkExist(primaryFieldSchema, fieldsData)
-	assert.Equal(t, 2, len(deduplicatedFields))
+	assert.NoError(t, err)
-
+	assert.True(t, hasDuplicate, "should detect duplicate primary keys")
-	// Check deduplicated primary keys
-	pkField := deduplicatedFields[0]
-	pkData := pkField.GetScalars().GetStringData().GetData()
-	assert.Equal(t, 3, len(pkData))
-	assert.Equal(t, []string{"c", "b", "a"}, pkData)
-
-	// Check corresponding int64 values (should be 300, 201, 101)
-	int64Field := deduplicatedFields[1]
-	int64Data := int64Field.GetScalars().GetLongData().GetData()
-	assert.Equal(t, 3, len(int64Data))
-	assert.Equal(t, []int64{300, 201, 101}, int64Data)
 }
 
-func TestUpsertTask_Deduplicate_NoDuplicates(t *testing.T) {
+func TestUpsertTask_NoDuplicatePK(t *testing.T) {
-	// Test with no duplicates - should return original data
+	schema := &schemapb.CollectionSchema{
-	primaryFieldSchema := &schemapb.FieldSchema{
+		Name: "test_no_duplicate_pk",
-		Name:         "id",
-		FieldID:      100,
-		DataType:     schemapb.DataType_Int64,
-		IsPrimaryKey: true,
-	}
-
-	collSchema := &schemapb.CollectionSchema{
 		Fields: []*schemapb.FieldSchema{
-			primaryFieldSchema,
+			{FieldID: 100, Name: "id", IsPrimaryKey: true, DataType: schemapb.DataType_Int64},
+			{FieldID: 101, Name: "value", DataType: schemapb.DataType_Int32},
 		},
 	}
-	schema := newSchemaInfo(collSchema)
 
+	// Data with unique primary keys: 1, 2, 3
 	fieldsData := []*schemapb.FieldData{
 		{
 			FieldName: "id",
+			FieldId:   100,
 			Type:      schemapb.DataType_Int64,
 			Field: &schemapb.FieldData_Scalars{
 				Scalars: &schemapb.ScalarField{
 					Data: &schemapb.ScalarField_LongData{
-						LongData: &schemapb.LongArray{
+						LongData: &schemapb.LongArray{Data: []int64{1, 2, 3}},
-							Data: []int64{1, 2, 3, 4, 5},
 					},
 				},
 			},
 		},
-		},
-	}
-
-	deduplicatedFields, newNumRows, err := DeduplicateFieldData(primaryFieldSchema, fieldsData, schema)
-	assert.NoError(t, err)
-	assert.Equal(t, uint32(5), newNumRows)
-	assert.Equal(t, 1, len(deduplicatedFields))
-
-	// Should be unchanged
-	pkField := deduplicatedFields[0]
-	pkData := pkField.GetScalars().GetLongData().GetData()
-	assert.Equal(t, []int64{1, 2, 3, 4, 5}, pkData)
-}
-
-func TestUpsertTask_Deduplicate_WithVector(t *testing.T) {
-	// Test deduplication with vector field
-	primaryFieldSchema := &schemapb.FieldSchema{
-		Name:         "id",
-		FieldID:      100,
-		DataType:     schemapb.DataType_Int64,
-		IsPrimaryKey: true,
-	}
-
-	collSchema := &schemapb.CollectionSchema{
-		Fields: []*schemapb.FieldSchema{
-			primaryFieldSchema,
 		{
-				Name:     "vector",
+			FieldName: "value",
-				FieldID:  101,
+			FieldId:   101,
-				DataType: schemapb.DataType_FloatVector,
+			Type:      schemapb.DataType_Int32,
-			},
-		},
-	}
-	schema := newSchemaInfo(collSchema)
-
-	dim := 4
-	// Create field data with duplicate IDs: [1, 2, 1]
-	// Expected to keep indices [1, 2] (last occurrence of 2, last occurrence of 1)
-	fieldsData := []*schemapb.FieldData{
-		{
-			FieldName: "id",
-			Type:      schemapb.DataType_Int64,
 			Field: &schemapb.FieldData_Scalars{
 				Scalars: &schemapb.ScalarField{
-					Data: &schemapb.ScalarField_LongData{
+					Data: &schemapb.ScalarField_IntData{
-						LongData: &schemapb.LongArray{
+						IntData: &schemapb.IntArray{Data: []int32{100, 200, 300}},
-							Data: []int64{1, 2, 1},
-						},
-					},
-				},
-			},
-		},
-		{
-			FieldName: "vector",
-			Type:      schemapb.DataType_FloatVector,
-			Field: &schemapb.FieldData_Vectors{
-				Vectors: &schemapb.VectorField{
-					Dim: int64(dim),
-					Data: &schemapb.VectorField_FloatVector{
-						FloatVector: &schemapb.FloatArray{
-							Data: []float32{
-								1.0, 1.1, 1.2, 1.3, // vector for ID 1 (first occurrence)
-								2.0, 2.1, 2.2, 2.3, // vector for ID 2
-								1.4, 1.5, 1.6, 1.7, // vector for ID 1 (second occurrence - keep this)
-							},
-						},
 					},
 				},
 			},
 		},
 	}
 
-	deduplicatedFields, newNumRows, err := DeduplicateFieldData(primaryFieldSchema, fieldsData, schema)
+	// Call CheckDuplicatePkExist directly to verify no duplicate error
+	primaryFieldSchema, err := typeutil.GetPrimaryFieldSchema(schema)
 	assert.NoError(t, err)
-	assert.Equal(t, uint32(2), newNumRows)
+	hasDuplicate, err := CheckDuplicatePkExist(primaryFieldSchema, fieldsData)
-	assert.Equal(t, 2, len(deduplicatedFields))
-
-	// Check deduplicated primary keys
-	pkField := deduplicatedFields[0]
-	pkData := pkField.GetScalars().GetLongData().GetData()
-	assert.Equal(t, 2, len(pkData))
-	assert.Equal(t, []int64{2, 1}, pkData)
-
-	// Check corresponding vector (should keep vectors for ID 2 and ID 1's last occurrence)
-	vectorField := deduplicatedFields[1]
-	vectorData := vectorField.GetVectors().GetFloatVector().GetData()
-	assert.Equal(t, 8, len(vectorData)) // 2 vectors * 4 dimensions
-	expectedVector := []float32{
-		2.0, 2.1, 2.2, 2.3, // vector for ID 2
-		1.4, 1.5, 1.6, 1.7, // vector for ID 1 (last occurrence)
-	}
-	assert.Equal(t, expectedVector, vectorData)
-}
-
-func TestUpsertTask_Deduplicate_EmptyData(t *testing.T) {
-	// Test with empty data
-	primaryFieldSchema := &schemapb.FieldSchema{
-		Name:         "id",
-		FieldID:      100,
-		DataType:     schemapb.DataType_Int64,
-		IsPrimaryKey: true,
-	}
-
-	collSchema := &schemapb.CollectionSchema{
-		Fields: []*schemapb.FieldSchema{
-			primaryFieldSchema,
-		},
-	}
-	schema := newSchemaInfo(collSchema)
-
-	fieldsData := []*schemapb.FieldData{}
-
-	deduplicatedFields, newNumRows, err := DeduplicateFieldData(primaryFieldSchema, fieldsData, schema)
 	assert.NoError(t, err)
-	assert.Equal(t, uint32(0), newNumRows)
+	assert.False(t, hasDuplicate, "should not have duplicate primary keys")
-	assert.Equal(t, 0, len(deduplicatedFields))
-}
-
-func TestUpsertTask_Deduplicate_MissingPrimaryKey(t *testing.T) {
-	// Test with missing primary key field
-	primaryFieldSchema := &schemapb.FieldSchema{
-		Name:         "id",
-		FieldID:      100,
-		DataType:     schemapb.DataType_Int64,
-		IsPrimaryKey: true,
-	}
-
-	collSchema := &schemapb.CollectionSchema{
-		Fields: []*schemapb.FieldSchema{
-			primaryFieldSchema,
-			{
-				Name:     "other_field",
-				FieldID:  101,
-				DataType: schemapb.DataType_Float,
-			},
-		},
-	}
-	schema := newSchemaInfo(collSchema)
-
-	fieldsData := []*schemapb.FieldData{
-		{
-			FieldName: "other_field",
-			Type:      schemapb.DataType_Float,
-			Field: &schemapb.FieldData_Scalars{
-				Scalars: &schemapb.ScalarField{
-					Data: &schemapb.ScalarField_FloatData{
-						FloatData: &schemapb.FloatArray{
-							Data: []float32{1.1, 2.2},
-						},
-					},
-				},
-			},
-		},
-	}
-
-	_, _, err := DeduplicateFieldData(primaryFieldSchema, fieldsData, schema)
-	assert.Error(t, err)
-	// validateFieldDataColumns will fail first due to column count mismatch
-	// or the function will fail when trying to find primary key
-	assert.True(t, err != nil)
 }

internal/proxy/util.go

@@ -1049,31 +1049,25 @@ func parsePrimaryFieldData2IDs(fieldData *schemapb.FieldData) (*schemapb.IDs, er
 	return primaryData, nil
 }
 
-// findLastOccurrenceIndices finds indices of last occurrences for each unique ID
+// hasDuplicates checks if there are any duplicate values in the slice.
-func findLastOccurrenceIndices[T comparable](ids []T) []int {
+// Returns true immediately when the first duplicate is found (early exit).
-	lastOccurrence := make(map[T]int, len(ids))
+func hasDuplicates[T comparable](ids []T) bool {
-	for idx, id := range ids {
+	seen := make(map[T]struct{}, len(ids))
-		lastOccurrence[id] = idx
+	for _, id := range ids {
+		if _, exists := seen[id]; exists {
+			return true
+		}
+		seen[id] = struct{}{}
+	}
+	return false
 }
 
-	keepIndices := make([]int, 0, len(lastOccurrence))
+// CheckDuplicatePkExist checks if there are duplicate primary keys in the field data.
-	for idx, id := range ids {
+// Returns (true, nil) if duplicates exist, (false, nil) if no duplicates.
-		if lastOccurrence[id] == idx {
+// Returns (false, error) if there's an error during checking.
-			keepIndices = append(keepIndices, idx)
+func CheckDuplicatePkExist(primaryFieldSchema *schemapb.FieldSchema, fieldsData []*schemapb.FieldData) (bool, error) {
-		}
-	}
-	return keepIndices
-}
-
-// DeduplicateFieldData removes duplicate primary keys from field data,
-// keeping the last occurrence of each ID
-func DeduplicateFieldData(primaryFieldSchema *schemapb.FieldSchema, fieldsData []*schemapb.FieldData, schema *schemaInfo) ([]*schemapb.FieldData, uint32, error) {
 	if len(fieldsData) == 0 {
-		return fieldsData, 0, nil
+		return false, nil
-	}
-
-	if err := fillFieldPropertiesOnly(fieldsData, schema); err != nil {
-		return nil, 0, err
 	}
 
 	// find primary field data
@@ -1086,64 +1080,26 @@ func DeduplicateFieldData(primaryFieldSchema *schemapb.FieldSchema, fieldsData [
 	}
 
 	if primaryFieldData == nil {
-		return nil, 0, merr.WrapErrParameterInvalidMsg(fmt.Sprintf("must assign pk when upsert, primary field: %v", primaryFieldSchema.GetName()))
+		return false, merr.WrapErrParameterInvalidMsg(fmt.Sprintf("must assign pk when upsert, primary field: %v", primaryFieldSchema.GetName()))
 	}
 
-	// get row count
+	// check for duplicates based on primary key type
-	var numRows int
 	switch primaryFieldData.Field.(type) {
 	case *schemapb.FieldData_Scalars:
 		scalarField := primaryFieldData.GetScalars()
 		switch scalarField.Data.(type) {
 		case *schemapb.ScalarField_LongData:
-			numRows = len(scalarField.GetLongData().GetData())
-		case *schemapb.ScalarField_StringData:
-			numRows = len(scalarField.GetStringData().GetData())
-		default:
-			return nil, 0, merr.WrapErrParameterInvalidMsg("unsupported primary key type")
-		}
-	default:
-		return nil, 0, merr.WrapErrParameterInvalidMsg("primary field must be scalar type")
-	}
-
-	if numRows == 0 {
-		return fieldsData, 0, nil
-	}
-
-	// build map to track last occurrence of each primary key
-	var keepIndices []int
-	switch primaryFieldData.Field.(type) {
-	case *schemapb.FieldData_Scalars:
-		scalarField := primaryFieldData.GetScalars()
-		switch scalarField.Data.(type) {
-		case *schemapb.ScalarField_LongData:
-			// for Int64 primary keys
 			intIDs := scalarField.GetLongData().GetData()
-			keepIndices = findLastOccurrenceIndices(intIDs)
+			return hasDuplicates(intIDs), nil
-
 		case *schemapb.ScalarField_StringData:
-			// for VarChar primary keys
 			strIDs := scalarField.GetStringData().GetData()
-			keepIndices = findLastOccurrenceIndices(strIDs)
+			return hasDuplicates(strIDs), nil
+		default:
+			return false, merr.WrapErrParameterInvalidMsg("unsupported primary key type")
 		}
+	default:
+		return false, merr.WrapErrParameterInvalidMsg("primary field must be scalar type")
 	}
-
-	// if no duplicates found, return original data
-	if len(keepIndices) == numRows {
-		return fieldsData, uint32(numRows), nil
-	}
-
-	log.Info("duplicate primary keys detected in upsert request, deduplicating",
-		zap.Int("original_rows", numRows),
-		zap.Int("deduplicated_rows", len(keepIndices)))
-
-	// use typeutil.AppendFieldData to rebuild field data with deduplicated rows
-	result := typeutil.PrepareResultFieldData(fieldsData, int64(len(keepIndices)))
-	for _, idx := range keepIndices {
-		typeutil.AppendFieldData(result, fieldsData, int64(idx))
-	}
-
-	return result, uint32(len(keepIndices)), nil
 }
 
 // autoGenPrimaryFieldData generate primary data when autoID == true
@@ -1214,12 +1170,12 @@ func validateFieldDataColumns(columns []*schemapb.FieldData, schema *schemaInfo)
 	expectColumnNum := 0
 
 	// Count expected columns
-	for _, field := range schema.GetFields() {
+	for _, field := range schema.CollectionSchema.GetFields() {
 		if !typeutil.IsBM25FunctionOutputField(field, schema.CollectionSchema) {
 			expectColumnNum++
 		}
 	}
-	for _, structField := range schema.GetStructArrayFields() {
+	for _, structField := range schema.CollectionSchema.GetStructArrayFields() {
 		expectColumnNum += len(structField.GetFields())
 	}
 

internal/proxy/util_test.go

@@ -4866,3 +4866,147 @@ func TestGetStorageCost(t *testing.T) {
 		assert.True(t, ok)
 	})
 }
+
+func TestCheckDuplicatePkExist_Int64PK(t *testing.T) {
+	primaryFieldSchema := &schemapb.FieldSchema{
+		Name:     "id",
+		FieldID:  100,
+		DataType: schemapb.DataType_Int64,
+	}
+
+	t.Run("with duplicates", func(t *testing.T) {
+		fieldsData := []*schemapb.FieldData{
+			{
+				FieldName: "id",
+				Type:      schemapb.DataType_Int64,
+				Field: &schemapb.FieldData_Scalars{
+					Scalars: &schemapb.ScalarField{
+						Data: &schemapb.ScalarField_LongData{
+							LongData: &schemapb.LongArray{
+								Data: []int64{1, 2, 3, 1, 4, 2}, // duplicates: 1, 2
+							},
+						},
+					},
+				},
+			},
+		}
+
+		hasDup, err := CheckDuplicatePkExist(primaryFieldSchema, fieldsData)
+		assert.NoError(t, err)
+		assert.True(t, hasDup)
+	})
+
+	t.Run("without duplicates", func(t *testing.T) {
+		fieldsData := []*schemapb.FieldData{
+			{
+				FieldName: "id",
+				Type:      schemapb.DataType_Int64,
+				Field: &schemapb.FieldData_Scalars{
+					Scalars: &schemapb.ScalarField{
+						Data: &schemapb.ScalarField_LongData{
+							LongData: &schemapb.LongArray{
+								Data: []int64{1, 2, 3, 4, 5},
+							},
+						},
+					},
+				},
+			},
+		}
+
+		hasDup, err := CheckDuplicatePkExist(primaryFieldSchema, fieldsData)
+		assert.NoError(t, err)
+		assert.False(t, hasDup)
+	})
+}
+
+func TestCheckDuplicatePkExist_VarCharPK(t *testing.T) {
+	primaryFieldSchema := &schemapb.FieldSchema{
+		Name:     "id",
+		FieldID:  100,
+		DataType: schemapb.DataType_VarChar,
+	}
+
+	t.Run("with duplicates", func(t *testing.T) {
+		fieldsData := []*schemapb.FieldData{
+			{
+				FieldName: "id",
+				Type:      schemapb.DataType_VarChar,
+				Field: &schemapb.FieldData_Scalars{
+					Scalars: &schemapb.ScalarField{
+						Data: &schemapb.ScalarField_StringData{
+							StringData: &schemapb.StringArray{
+								Data: []string{"a", "b", "c", "a", "d"}, // duplicate: "a"
+							},
+						},
+					},
+				},
+			},
+		}
+
+		hasDup, err := CheckDuplicatePkExist(primaryFieldSchema, fieldsData)
+		assert.NoError(t, err)
+		assert.True(t, hasDup)
+	})
+
+	t.Run("without duplicates", func(t *testing.T) {
+		fieldsData := []*schemapb.FieldData{
+			{
+				FieldName: "id",
+				Type:      schemapb.DataType_VarChar,
+				Field: &schemapb.FieldData_Scalars{
+					Scalars: &schemapb.ScalarField{
+						Data: &schemapb.ScalarField_StringData{
+							StringData: &schemapb.StringArray{
+								Data: []string{"a", "b", "c", "d", "e"},
+							},
+						},
+					},
+				},
+			},
+		}
+
+		hasDup, err := CheckDuplicatePkExist(primaryFieldSchema, fieldsData)
+		assert.NoError(t, err)
+		assert.False(t, hasDup)
+	})
+}
+
+func TestCheckDuplicatePkExist_EmptyData(t *testing.T) {
+	primaryFieldSchema := &schemapb.FieldSchema{
+		Name:     "id",
+		FieldID:  100,
+		DataType: schemapb.DataType_Int64,
+	}
+
+	hasDup, err := CheckDuplicatePkExist(primaryFieldSchema, []*schemapb.FieldData{})
+	assert.NoError(t, err)
+	assert.False(t, hasDup)
+}
+
+func TestCheckDuplicatePkExist_MissingPrimaryKey(t *testing.T) {
+	primaryFieldSchema := &schemapb.FieldSchema{
+		Name:     "id",
+		FieldID:  100,
+		DataType: schemapb.DataType_Int64,
+	}
+
+	fieldsData := []*schemapb.FieldData{
+		{
+			FieldName: "other_field",
+			Type:      schemapb.DataType_Int64,
+			Field: &schemapb.FieldData_Scalars{
+				Scalars: &schemapb.ScalarField{
+					Data: &schemapb.ScalarField_LongData{
+						LongData: &schemapb.LongArray{
+							Data: []int64{1, 2, 3},
+						},
+					},
+				},
+			},
+		},
+	}
+
+	hasDup, err := CheckDuplicatePkExist(primaryFieldSchema, fieldsData)
+	assert.Error(t, err)
+	assert.False(t, hasDup)
+}

tests/go_client/testcases/upsert_test.go

@@ -436,7 +436,7 @@ func TestUpsertAutoID(t *testing.T) {
 	// upsert without pks -> error
 	vecColumn = hp.GenColumnData(nb, entity.FieldTypeFloatVector, *hp.TNewDataOption())
 	_, err = mc.Upsert(ctx, client.NewColumnBasedInsertOption(schema.CollectionName).WithColumns(vecColumn))
-	common.CheckErr(t, err, false, "has no corresponding fieldData pass in: invalid parameter")
+	common.CheckErr(t, err, false, "must assign pk when upsert")
 }
 
 // test upsert autoId collection

tests/python_client/milvus_client/test_milvus_client_partial_update.py

@@ -1221,14 +1221,13 @@ class TestMilvusClientPartialUpdateValid(TestMilvusClientV2Base):
     @pytest.mark.tags(CaseLabel.L1)
     def test_milvus_client_partial_update_same_pk_same_field(self):
         """
-        target:  Test PU will success and query will success
+        target: Test partial update on an existing pk with the same field will success
         method:
             1. Create a collection
             2. Insert rows
-            3. Upsert the rows with same pk and same field
+            3. Upsert a single row with existing pk and same field (partial update)
-            4. Query the rows
+            4. Query the row to verify the update
-            5. Upsert the rows with same pk and different field
+        expected: All steps should success, and the field value should be updated
-        expected: Step 2 -> 4 should success 5 should fail
         """
         # step 1: create collection
         client = self._client()
@@ -1248,16 +1247,19 @@ class TestMilvusClientPartialUpdateValid(TestMilvusClientV2Base):
         rows = cf.gen_row_data_by_schema(nb=default_nb, schema=schema)
         self.upsert(client, collection_name, rows, partial_update=True)
 
-        # step 3: Upsert the rows with same pk and same field
+        # step 3: Upsert a single row with existing pk=0 and update the same field
-        new_rows = [{default_primary_key_field_name: 0, 
+        updated_value = 99999
-                    default_int32_field_name: i} for i in range(default_nb)]
+        new_row = {default_primary_key_field_name: 0,
-        self.upsert(client, collection_name, new_rows, partial_update=True)
+                   default_int32_field_name: updated_value}
+        self.upsert(client, collection_name, [new_row], partial_update=True)
 
-        # step 4: Query the rows
+        # step 4: Query the row to verify the update
+        expected_row = {default_primary_key_field_name: 0,
+                        default_int32_field_name: updated_value}
         result = self.query(client, collection_name, filter=f"{default_primary_key_field_name} == 0",
                    check_task=CheckTasks.check_query_results,
                    output_fields=[default_int32_field_name],
-                   check_items={exp_res: [new_rows[-1]],
+                   check_items={exp_res: [expected_row],
                                 "pk_name": default_primary_key_field_name})[0]
         assert len(result) == 1
 

tests/python_client/milvus_client/test_milvus_client_upsert.py

@@ -371,6 +371,72 @@ class TestMilvusClientUpsertInvalid(TestMilvusClientV2Base):
                     check_task=CheckTasks.err_res,
                     check_items=error)
 
+    @pytest.mark.tags(CaseLabel.L1)
+    def test_milvus_client_upsert_duplicate_pk_int64(self):
+        """
+        target: test upsert with duplicate primary keys (Int64)
+        method:
+            1. create collection with Int64 primary key
+            2. upsert data with duplicate primary keys in the same batch
+        expected: raise error - duplicate primary keys are not allowed
+        """
+        client = self._client()
+        collection_name = cf.gen_collection_name_by_testcase_name()
+        # 1. create collection
+        self.create_collection(client, collection_name, default_dim, consistency_level="Strong")
+        # 2. upsert with duplicate PKs: 1, 2, 1 (duplicate)
+        rng = np.random.default_rng(seed=19530)
+        rows = [
+            {default_primary_key_field_name: 1, default_vector_field_name: list(rng.random((1, default_dim))[0]),
+             default_float_field_name: 1.0, default_string_field_name: "first"},
+            {default_primary_key_field_name: 2, default_vector_field_name: list(rng.random((1, default_dim))[0]),
+             default_float_field_name: 2.0, default_string_field_name: "second"},
+            {default_primary_key_field_name: 1, default_vector_field_name: list(rng.random((1, default_dim))[0]),
+             default_float_field_name: 1.1, default_string_field_name: "duplicate"},
+        ]
+        error = {ct.err_code: 1100,
+                 ct.err_msg: "duplicate primary keys are not allowed in the same batch"}
+        self.upsert(client, collection_name, rows,
+                    check_task=CheckTasks.err_res, check_items=error)
+        self.drop_collection(client, collection_name)
+
+    @pytest.mark.tags(CaseLabel.L1)
+    def test_milvus_client_upsert_duplicate_pk_varchar(self):
+        """
+        target: test upsert with duplicate primary keys (VarChar)
+        method:
+            1. create collection with VarChar primary key
+            2. upsert data with duplicate primary keys in the same batch
+        expected: raise error - duplicate primary keys are not allowed
+        """
+        client = self._client()
+        collection_name = cf.gen_collection_name_by_testcase_name()
+        dim = default_dim
+        # 1. create collection with VarChar primary key
+        schema = self.create_schema(client, enable_dynamic_field=False)[0]
+        schema.add_field(default_primary_key_field_name, DataType.VARCHAR, max_length=64, is_primary=True,
+                         auto_id=False)
+        schema.add_field(default_vector_field_name, DataType.FLOAT_VECTOR, dim=dim)
+        schema.add_field(default_float_field_name, DataType.FLOAT)
+        index_params = self.prepare_index_params(client)[0]
+        index_params.add_index(default_vector_field_name, metric_type="COSINE")
+        self.create_collection(client, collection_name, dimension=dim, schema=schema, index_params=index_params)
+        # 2. upsert with duplicate PKs: "a", "b", "a" (duplicate)
+        rng = np.random.default_rng(seed=19530)
+        rows = [
+            {default_primary_key_field_name: "a", default_vector_field_name: list(rng.random((1, dim))[0]),
+             default_float_field_name: 1.0},
+            {default_primary_key_field_name: "b", default_vector_field_name: list(rng.random((1, dim))[0]),
+             default_float_field_name: 2.0},
+            {default_primary_key_field_name: "a", default_vector_field_name: list(rng.random((1, dim))[0]),
+             default_float_field_name: 1.1},
+        ]
+        error = {ct.err_code: 1100,
+                 ct.err_msg: "duplicate primary keys are not allowed in the same batch"}
+        self.upsert(client, collection_name, rows,
+                    check_task=CheckTasks.err_res, check_items=error)
+        self.drop_collection(client, collection_name)
+
 
 class TestMilvusClientUpsertValid(TestMilvusClientV2Base):
     """ Test case of search interface """
@@ -551,342 +617,3 @@ class TestMilvusClientUpsertValid(TestMilvusClientV2Base):
             self.drop_partition(client, collection_name, partition_name)
         if self.has_collection(client, collection_name)[0]:
             self.drop_collection(client, collection_name)
-
-
-class TestMilvusClientUpsertDedup(TestMilvusClientV2Base):
-    """Test case for upsert deduplication functionality"""
-
-    @pytest.fixture(scope="function", params=["COSINE", "L2"])
-    def metric_type(self, request):
-        yield request.param
-
-    @pytest.mark.tags(CaseLabel.L1)
-    def test_milvus_client_upsert_dedup_int64_pk(self):
-        """
-        target: test upsert with duplicate int64 primary keys in same batch
-        method: 
-            1. create collection with int64 primary key
-            2. upsert data with duplicate primary keys [1, 2, 3, 2, 1]
-            3. query to verify only last occurrence is kept
-        expected: only 3 unique records exist, with data from last occurrence
-        """
-        client = self._client()
-        collection_name = cf.gen_collection_name_by_testcase_name()
-        
-        # 1. create collection
-        self.create_collection(client, collection_name, default_dim, consistency_level="Strong")
-        
-        # 2. upsert data with duplicate PKs: [1, 2, 3, 2, 1]
-        # Expected: keep last occurrence -> [3, 2, 1] at indices [2, 3, 4]
-        rng = np.random.default_rng(seed=19530)
-        rows = [
-            {default_primary_key_field_name: 1, default_vector_field_name: list(rng.random((1, default_dim))[0]),
-             default_float_field_name: 1.0, default_string_field_name: "str_1_first"},
-            {default_primary_key_field_name: 2, default_vector_field_name: list(rng.random((1, default_dim))[0]),
-             default_float_field_name: 2.0, default_string_field_name: "str_2_first"},
-            {default_primary_key_field_name: 3, default_vector_field_name: list(rng.random((1, default_dim))[0]),
-             default_float_field_name: 3.0, default_string_field_name: "str_3"},
-            {default_primary_key_field_name: 2, default_vector_field_name: list(rng.random((1, default_dim))[0]),
-             default_float_field_name: 2.1, default_string_field_name: "str_2_last"},
-            {default_primary_key_field_name: 1, default_vector_field_name: list(rng.random((1, default_dim))[0]),
-             default_float_field_name: 1.1, default_string_field_name: "str_1_last"},
-        ]
-        
-        results = self.upsert(client, collection_name, rows)[0]
-        # After deduplication, should only have 3 records
-        assert results['upsert_count'] == 3
-        
-        # 3. query to verify deduplication - should have only 3 unique records
-        query_results = self.query(client, collection_name, filter="id >= 0")[0]
-        assert len(query_results) == 3
-        
-        # Verify that last occurrence data is kept
-        id_to_data = {item['id']: item for item in query_results}
-        assert 1 in id_to_data
-        assert 2 in id_to_data
-        assert 3 in id_to_data
-        
-        # Check that data from last occurrence is preserved
-        assert id_to_data[1]['float'] == 1.1
-        assert id_to_data[1]['varchar'] == "str_1_last"
-        assert id_to_data[2]['float'] == 2.1
-        assert id_to_data[2]['varchar'] == "str_2_last"
-        assert id_to_data[3]['float'] == 3.0
-        assert id_to_data[3]['varchar'] == "str_3"
-        
-        self.drop_collection(client, collection_name)
-
-    @pytest.mark.tags(CaseLabel.L1)
-    def test_milvus_client_upsert_dedup_varchar_pk(self):
-        """
-        target: test upsert with duplicate varchar primary keys in same batch
-        method: 
-            1. create collection with varchar primary key
-            2. upsert data with duplicate primary keys ["a", "b", "c", "b", "a"]
-            3. query to verify only last occurrence is kept
-        expected: only 3 unique records exist, with data from last occurrence
-        """
-        client = self._client()
-        collection_name = cf.gen_collection_name_by_testcase_name()
-        
-        # 1. create collection with varchar primary key
-        schema = self.create_schema(client, enable_dynamic_field=True)[0]
-        schema.add_field("id", DataType.VARCHAR, max_length=64, is_primary=True, auto_id=False)
-        schema.add_field(default_vector_field_name, DataType.FLOAT_VECTOR, dim=default_dim)
-        schema.add_field("age", DataType.INT64)
-        index_params = self.prepare_index_params(client)[0]
-        index_params.add_index(default_vector_field_name, metric_type="COSINE")
-        self.create_collection(client, collection_name, default_dim, schema=schema, 
-                               index_params=index_params, consistency_level="Strong")
-        
-        # 2. upsert data with duplicate PKs: ["a", "b", "c", "b", "a"]
-        # Expected: keep last occurrence -> ["c", "b", "a"] at indices [2, 3, 4]
-        rng = np.random.default_rng(seed=19530)
-        rows = [
-            {"id": "a", default_vector_field_name: list(rng.random((1, default_dim))[0]), 
-             "age": 10},
-            {"id": "b", default_vector_field_name: list(rng.random((1, default_dim))[0]), 
-             "age": 20},
-            {"id": "c", default_vector_field_name: list(rng.random((1, default_dim))[0]), 
-             "age": 30},
-            {"id": "b", default_vector_field_name: list(rng.random((1, default_dim))[0]), 
-             "age": 21},
-            {"id": "a", default_vector_field_name: list(rng.random((1, default_dim))[0]), 
-             "age": 11},
-        ]
-        
-        results = self.upsert(client, collection_name, rows)[0]
-        # After deduplication, should only have 3 records
-        assert results['upsert_count'] == 3
-        
-        # 3. query to verify deduplication
-        query_results = self.query(client, collection_name, filter='id in ["a", "b", "c"]')[0]
-        assert len(query_results) == 3
-        
-        # Verify that last occurrence data is kept
-        id_to_data = {item['id']: item for item in query_results}
-        assert "a" in id_to_data
-        assert "b" in id_to_data
-        assert "c" in id_to_data
-        
-        # Check that data from last occurrence is preserved
-        assert id_to_data["a"]["age"] == 11
-        assert id_to_data["b"]["age"] == 21
-        assert id_to_data["c"]["age"] == 30
-        
-        self.drop_collection(client, collection_name)
-
-    @pytest.mark.tags(CaseLabel.L1)
-    def test_milvus_client_upsert_dedup_all_duplicates(self):
-        """
-        target: test upsert when all records have same primary key
-        method: 
-            1. create collection
-            2. upsert 5 records with same primary key
-            3. query to verify only 1 record exists
-        expected: only 1 record exists with data from last occurrence
-        """
-        client = self._client()
-        collection_name = cf.gen_collection_name_by_testcase_name()
-        
-        # 1. create collection
-        self.create_collection(client, collection_name, default_dim, consistency_level="Strong")
-        
-        # 2. upsert data where all have same PK (id=1)
-        rng = np.random.default_rng(seed=19530)
-        rows = [
-            {default_primary_key_field_name: 1, default_vector_field_name: list(rng.random((1, default_dim))[0]),
-             default_float_field_name: i * 1.0, default_string_field_name: f"version_{i}"} 
-            for i in range(5)
-        ]
-        
-        results = self.upsert(client, collection_name, rows)[0]
-        # After deduplication, should only have 1 record
-        assert results['upsert_count'] == 1
-        
-        # 3. query to verify only 1 record exists
-        query_results = self.query(client, collection_name, filter="id == 1")[0]
-        assert len(query_results) == 1
-        
-        # Verify it's the last occurrence (i=4)
-        assert query_results[0]['float'] == 4.0
-        assert query_results[0]['varchar'] == "version_4"
-        
-        self.drop_collection(client, collection_name)
-
-    @pytest.mark.tags(CaseLabel.L1)
-    def test_milvus_client_upsert_dedup_no_duplicates(self):
-        """
-        target: test upsert with no duplicate primary keys
-        method: 
-            1. create collection
-            2. upsert data with unique primary keys
-            3. query to verify all records exist
-        expected: all records exist as-is
-        """
-        client = self._client()
-        collection_name = cf.gen_collection_name_by_testcase_name()
-        
-        # 1. create collection
-        self.create_collection(client, collection_name, default_dim, consistency_level="Strong")
-        
-        # 2. upsert data with unique PKs
-        rng = np.random.default_rng(seed=19530)
-        nb = 10
-        rows = [
-            {default_primary_key_field_name: i, default_vector_field_name: list(rng.random((1, default_dim))[0]),
-             default_float_field_name: i * 1.0, default_string_field_name: str(i)} 
-            for i in range(nb)
-        ]
-        
-        results = self.upsert(client, collection_name, rows)[0]
-        # No deduplication should occur
-        assert results['upsert_count'] == nb
-        
-        # 3. query to verify all records exist
-        query_results = self.query(client, collection_name, filter=f"id >= 0")[0]
-        assert len(query_results) == nb
-        
-        self.drop_collection(client, collection_name)
-
-    @pytest.mark.tags(CaseLabel.L2)
-    def test_milvus_client_upsert_dedup_large_batch(self):
-        """
-        target: test upsert deduplication with large batch
-        method: 
-            1. create collection
-            2. upsert large batch with 50% duplicate primary keys
-            3. query to verify correct number of records
-        expected: only unique records exist
-        """
-        client = self._client()
-        collection_name = cf.gen_collection_name_by_testcase_name()
-        
-        # 1. create collection
-        self.create_collection(client, collection_name, default_dim, consistency_level="Strong")
-        
-        # 2. upsert large batch where each ID appears twice
-        rng = np.random.default_rng(seed=19530)
-        nb = 500
-        unique_ids = nb // 2  # 250 unique IDs
-        
-        rows = []
-        for i in range(nb):
-            pk = i % unique_ids  # This creates duplicates: 0,1,2...249,0,1,2...249
-            rows.append({
-                default_primary_key_field_name: pk, 
-                default_vector_field_name: list(rng.random((1, default_dim))[0]),
-                default_float_field_name: float(i),  # Different value for each row
-                default_string_field_name: f"batch_{i}"
-            })
-        
-        results = self.upsert(client, collection_name, rows)[0]
-        # After deduplication, should only have unique_ids records
-        assert results['upsert_count'] == unique_ids
-        
-        # 3. query to verify correct number of records
-        query_results = self.query(client, collection_name, filter=f"id >= 0", limit=1000)[0]
-        assert len(query_results) == unique_ids
-        
-        # Verify that last occurrence is kept (should have higher float values)
-        for item in query_results:
-            pk = item['id']
-            # Last occurrence of pk is at index (pk + unique_ids)
-            expected_float = float(pk + unique_ids)
-            assert item['float'] == expected_float
-            assert item['varchar'] == f"batch_{pk + unique_ids}"
-        
-        self.drop_collection(client, collection_name)
-
-    @pytest.mark.tags(CaseLabel.L1)
-    def test_milvus_client_upsert_dedup_with_partition(self):
-        """
-        target: test upsert deduplication works correctly with partitions
-        method: 
-            1. create collection with partition
-            2. upsert data with duplicates to specific partition
-            3. query to verify deduplication in partition
-        expected: deduplication works within partition
-        """
-        client = self._client()
-        collection_name = cf.gen_collection_name_by_testcase_name()
-        partition_name = cf.gen_unique_str("partition")
-        
-        # 1. create collection and partition
-        self.create_collection(client, collection_name, default_dim, consistency_level="Strong")
-        self.create_partition(client, collection_name, partition_name)
-        
-        # 2. upsert data with duplicates to partition
-        rng = np.random.default_rng(seed=19530)
-        rows = [
-            {default_primary_key_field_name: 1, default_vector_field_name: list(rng.random((1, default_dim))[0]),
-             default_float_field_name: 1.0, default_string_field_name: "first"},
-            {default_primary_key_field_name: 2, default_vector_field_name: list(rng.random((1, default_dim))[0]),
-             default_float_field_name: 2.0, default_string_field_name: "unique"},
-            {default_primary_key_field_name: 1, default_vector_field_name: list(rng.random((1, default_dim))[0]),
-             default_float_field_name: 1.1, default_string_field_name: "last"},
-        ]
-        
-        results = self.upsert(client, collection_name, rows, partition_name=partition_name)[0]
-        assert results['upsert_count'] == 2
-        
-        # 3. query partition to verify deduplication
-        query_results = self.query(client, collection_name, filter="id >= 0", 
-                                   partition_names=[partition_name])[0]
-        assert len(query_results) == 2
-        
-        # Verify correct data
-        id_to_data = {item['id']: item for item in query_results}
-        assert id_to_data[1]['float'] == 1.1
-        assert id_to_data[1]['varchar'] == "last"
-        assert id_to_data[2]['float'] == 2.0
-        assert id_to_data[2]['varchar'] == "unique"
-        
-        self.drop_collection(client, collection_name)
-
-    @pytest.mark.tags(CaseLabel.L1)
-    def test_milvus_client_upsert_dedup_with_vectors(self):
-        """
-        target: test upsert deduplication preserves correct vector data
-        method: 
-            1. create collection
-            2. upsert data with duplicate PKs but different vectors
-            3. search to verify correct vector is preserved
-        expected: vector from last occurrence is preserved
-        """
-        client = self._client()
-        collection_name = cf.gen_collection_name_by_testcase_name()
-        
-        # 1. create collection
-        self.create_collection(client, collection_name, default_dim, consistency_level="Strong")
-        
-        # 2. upsert data with duplicate PK=1 but different vectors
-        # Create distinctly different vectors for easy verification
-        first_vector = [1.0] * default_dim  # All 1.0
-        last_vector = [2.0] * default_dim   # All 2.0
-        
-        rows = [
-            {default_primary_key_field_name: 1, default_vector_field_name: first_vector,
-             default_float_field_name: 1.0, default_string_field_name: "first"},
-            {default_primary_key_field_name: 2, default_vector_field_name: [0.5] * default_dim,
-             default_float_field_name: 2.0, default_string_field_name: "unique"},
-            {default_primary_key_field_name: 1, default_vector_field_name: last_vector,
-             default_float_field_name: 1.1, default_string_field_name: "last"},
-        ]
-        
-        results = self.upsert(client, collection_name, rows)[0]
-        assert results['upsert_count'] == 2
-        
-        # 3. query to get vector data
-        query_results = self.query(client, collection_name, filter="id == 1", 
-                                   output_fields=["id", "vector", "float", "varchar"])[0]
-        assert len(query_results) == 1
-        
-        # Verify it's the last occurrence with last_vector
-        result = query_results[0]
-        assert result['float'] == 1.1
-        assert result['varchar'] == "last"
-        # Vector should be last_vector (all 2.0)
-        assert all(abs(v - 2.0) < 0.001 for v in result['vector'])
-        
-        self.drop_collection(client, collection_name)