diff --git a/cmd/sqlflowserver/main.go b/cmd/sqlflowserver/main.go
index ebeee1d78f..4a1a2d3c6e 100644
--- a/cmd/sqlflowserver/main.go
+++ b/cmd/sqlflowserver/main.go
@@ -48,7 +48,7 @@ func newServer(caCrt, caKey string) (*grpc.Server, error) {
 	return s, nil
 }
 
-func start(modelDir, caCrt, caKey string, port int) {
+func start(modelDir, caCrt, caKey string, port int, isArgoMode bool) {
 	s, err := newServer(caCrt, caKey)
 	if err != nil {
 		log.Fatalf("failed to create new gRPC Server: %v", err)
@@ -59,8 +59,12 @@ func start(modelDir, caCrt, caKey string, port int) {
 			os.Mkdir(modelDir, os.ModePerm)
 		}
 	}
+	if isArgoMode {
+		proto.RegisterSQLFlowServer(s, server.NewServer(sf.SubmitWorkflow, modelDir))
+	} else {
+		proto.RegisterSQLFlowServer(s, server.NewServer(sf.RunSQLProgram, modelDir))
+	}
 
-	proto.RegisterSQLFlowServer(s, server.NewServer(sf.RunSQLProgram, modelDir))
 	listenString := fmt.Sprintf(":%d", port)
 
 	lis, err := net.Listen("tcp", listenString)
@@ -81,6 +85,7 @@ func main() {
 	caCrt := flag.String("ca-crt", "", "CA certificate file.")
 	caKey := flag.String("ca-key", "", "CA private key file.")
 	port := flag.Int("port", 50051, "TCP port to listen on.")
+	isArgoMode := flag.Bool("argo-mode", false, "Enable Argo workflow model.")
 	flag.Parse()
-	start(*modelDir, *caCrt, *caKey, *port)
+	start(*modelDir, *caCrt, *caKey, *port, *isArgoMode)
 }
diff --git a/cmd/sqlflowserver/main_test.go b/cmd/sqlflowserver/main_test.go
index 6ef875b22d..28603f8bff 100644
--- a/cmd/sqlflowserver/main_test.go
+++ b/cmd/sqlflowserver/main_test.go
@@ -265,7 +265,7 @@ func TestEnd2EndMySQL(t *testing.T) {
 		t.Fatalf("failed to generate CA pair %v", err)
 	}
 
-	go start(modelDir, caCrt, caKey, unitestPort)
+	go start(modelDir, caCrt, caKey, unitestPort, false)
 	waitPortReady(fmt.Sprintf("localhost:%d", unitestPort), 0)
 	err = prepareTestData(dbConnStr)
 	if err != nil {
@@ -359,7 +359,7 @@ func TestEnd2EndHive(t *testing.T) {
 		t.Skip("Skipping hive tests")
 	}
 	dbConnStr = "hive://root:root@127.0.0.1:10000/iris?auth=NOSASL"
-	go start(modelDir, caCrt, caKey, unitestPort)
+	go start(modelDir, caCrt, caKey, unitestPort, false)
 	waitPortReady(fmt.Sprintf("localhost:%d", unitestPort), 0)
 	err = prepareTestData(dbConnStr)
 	if err != nil {
@@ -396,7 +396,7 @@ func TestEnd2EndMaxCompute(t *testing.T) {
 	SK := os.Getenv("MAXCOMPUTE_SK")
 	endpoint := os.Getenv("MAXCOMPUTE_ENDPOINT")
 	dbConnStr = fmt.Sprintf("maxcompute://%s:%s@%s", AK, SK, endpoint)
-	go start(modelDir, caCrt, caKey, unitestPort)
+	go start(modelDir, caCrt, caKey, unitestPort, false)
 	waitPortReady(fmt.Sprintf("localhost:%d", unitestPort), 0)
 	err = prepareTestData(dbConnStr)
 	if err != nil {
@@ -440,7 +440,7 @@ func TestEnd2EndMaxComputeALPS(t *testing.T) {
 		t.Fatalf("prepare test dataset failed: %v", err)
 	}
 
-	go start(modelDir, caCrt, caKey, unitestPort)
+	go start(modelDir, caCrt, caKey, unitestPort, false)
 	waitPortReady(fmt.Sprintf("localhost:%d", unitestPort), 0)
 
 	t.Run("CaseTrainALPS", CaseTrainALPS)
@@ -479,7 +479,7 @@ func TestEnd2EndMaxComputeElasticDL(t *testing.T) {
 		t.Fatalf("prepare test dataset failed: %v", err)
 	}
 
-	go start(modelDir, caCrt, caKey, unitestPort)
+	go start(modelDir, caCrt, caKey, unitestPort, false)
 	waitPortReady(fmt.Sprintf("localhost:%d", unitestPort), 0)
 
 	t.Run("CaseTrainElasticDL", CaseTrainElasticDL)
diff --git a/pkg/server/proto/sqlflow.proto b/pkg/server/proto/sqlflow.proto
index 71d8a4fa10..b301662231 100644
--- a/pkg/server/proto/sqlflow.proto
+++ b/pkg/server/proto/sqlflow.proto
@@ -10,27 +10,29 @@ service SQLFlow {
     // SQL statements like `SELECT ...`, `DESCRIBE ...` returns a rowset.
     // The rowset might be big. In such cases, Query returns a stream
     // of RunResponse
+    // 
+    // SQLFlow implements the Run interface with two mode:
     //
+    // 1. Local model
+    // The SQLFlow server execute the SQL statments on the local host.
+    // 
     // SQL statements like `USE database`, `DELETE` returns only a success
     // message.
     //
     // SQL statement like `SELECT ... TO TRAIN/PREDICT ...` returns a stream of
     // messages which indicates the training/predicting progress
+    // 
+    // 2. Argo Workflow mode
+    // The SQLFlow server submites an Argo workflow into a Kubernetes cluster,
+    // and returns a stream of messages indicates the WorkFlow ID and the 
+    // submiting progress.
+    //
+    // The SQLFlow gRPC client should fetch the logs of the workflow by
+    // calling the Fetch interface in a polling manner.
     rpc Run (Request) returns (stream Response);
 
-    // Submit a SQLFlow Job which contains a SQL program to SQLFlow server.
-    //
-    // A SQL program contains one or more SQL statments.
-    // Each of these SQL statments can be a standard SQL like:
-    // `SELECT ... FROM ...;`, `DESCRIBE ...`,
-    // or an extended SQLFlow SQL like:
-    // `SELECT ... TO TRAIN/PREDICT/EXPLAIN ...`.
-    //
-    // Submit returns a Job message which contains the SQLFlow Job ID.
-    rpc Submit (Request) returns (Job);
-    
     // Fetch fetchs the SQLFlow job status and logs in a polling manner.
-    rpc Fetch (Job) returns(JobStatus);
+    rpc Fetch (Job) returns (JobStatus);
 }
 
 message Job {
@@ -78,6 +80,7 @@ message Response {
         Row row = 2;
         Message message = 3;
         EndOfExecution eoe = 4;
+        Job job = 5;
     }
 }
 
diff --git a/pkg/server/sqlflowserver.go b/pkg/server/sqlflowserver.go
index 864a78d484..7436b6b2ff 100644
--- a/pkg/server/sqlflowserver.go
+++ b/pkg/server/sqlflowserver.go
@@ -34,7 +34,8 @@ import (
 )
 
 // NewServer returns a server instance
-func NewServer(run func(string, *sf.DB, string, *pb.Session) *sf.PipeReader, modelDir string) *Server {
+func NewServer(run func(string, *sf.DB, string, *pb.Session) *sf.PipeReader,
+	modelDir string) *Server {
 	return &Server{run: run, modelDir: modelDir}
 }
 
@@ -44,12 +45,6 @@ type Server struct {
 	modelDir string
 }
 
-// Submit implements `rpc Submit (Request) returns (Job)`
-func (s *Server) Submit(ctx context.Context, in *pb.Request) (*pb.Job, error) {
-	job := &pb.Job{}
-	return job, nil
-}
-
 // Fetch implements `rpc Fetch (Job) returns(JobStatus)`
 func (s *Server) Fetch(ctx context.Context, job *pb.Job) (*pb.JobStatus, error) {
 	js := &pb.JobStatus{}
@@ -59,6 +54,7 @@ func (s *Server) Fetch(ctx context.Context, job *pb.Job) (*pb.JobStatus, error)
 // Run implements `rpc Run (Request) returns (stream Response)`
 func (s *Server) Run(req *pb.Request, stream pb.SQLFlow_RunServer) error {
 	var db *sf.DB
+
 	var err error
 	if db, err = sf.NewDB(req.Session.DbConnStr); err != nil {
 		return fmt.Errorf("create DB failed: %v", err)
@@ -82,6 +78,9 @@ func (s *Server) Run(req *pb.Request, stream pb.SQLFlow_RunServer) error {
 			res, err = encodeRow(s)
 		case string:
 			res, err = encodeMessage(s)
+		case sf.WorkflowJob:
+			job := r.(sf.WorkflowJob)
+			res = &pb.Response{Response: &pb.Response_Job{Job: &pb.Job{Id: job.JobID}}}
 		case sf.EndOfExecution:
 			// if sqlStatements have only one field, do **NOT** return EndOfExecution message.
 			if len(sqlStatements) > 1 {
diff --git a/pkg/sql/executor_ir.go b/pkg/sql/executor_ir.go
index 0bb6ca9c47..9a04ce8a47 100644
--- a/pkg/sql/executor_ir.go
+++ b/pkg/sql/executor_ir.go
@@ -40,6 +40,11 @@ type EndOfExecution struct {
 	Statement string
 }
 
+// WorkflowJob indicates the Argo Workflow ID
+type WorkflowJob struct {
+	JobID string
+}
+
 var envSubmitter = os.Getenv("SQLFLOW_submitter")
 
 // SubmitterType is the type of SQLFlow submitter
@@ -88,6 +93,41 @@ func RunSQLProgram(sqlProgram string, db *DB, modelDir string, session *pb.Sessi
 	return rd
 }
 
+// SubmitWorkflow submits an Argo workflow
+func SubmitWorkflow(sqlProgram string, db *DB, modelDir string, session *pb.Session) *PipeReader {
+	rd, wr := Pipe()
+	go func() {
+		defer wr.Close()
+		err := submitWorkflow(wr, sqlProgram, db, modelDir, session)
+		if err != nil {
+			log.Errorf("submit Workflow error: %v", err)
+		}
+	}()
+	return rd
+}
+
+func submitWorkflow(wr *PipeWriter, sqlProgram string, db *DB, modelDir string, session *pb.Session) error {
+	sqls, err := parse(db.driverName, sqlProgram)
+	if err != nil {
+		return err
+	}
+
+	connStr := fmt.Sprintf("%s://%s", db.driverName, db.dataSourceName)
+	_, err = programToIR(sqls, connStr, modelDir, true, false)
+	if err != nil {
+		return err
+	}
+
+	// TODO(yancey1989):
+	// 1. call codegen_couler.go to genearte Couler program.
+	// 2. compile Couler program into Argo YAML.
+	// 3. submit Argo YAML and fetch the workflow ID.
+
+	return wr.Write(WorkflowJob{
+		JobID: "sqlflow-workflow",
+	})
+}
+
 func runSQLProgram(wr *PipeWriter, sqlProgram string, db *DB, modelDir string, session *pb.Session) error {
 	sqls, err := parse(db.driverName, sqlProgram)
 	if err != nil {
@@ -95,7 +135,7 @@ func runSQLProgram(wr *PipeWriter, sqlProgram string, db *DB, modelDir string, s
 	}
 
 	connStr := fmt.Sprintf("%s://%s", db.driverName, db.dataSourceName)
-	programIR, err := programToIR(sqls, connStr, modelDir, submitter() != SubmitterPAI)
+	programIR, err := programToIR(sqls, connStr, modelDir, submitter() != SubmitterPAI, true /*enableFeatureDerivation = true*/)
 	if err != nil {
 		return err
 	}
diff --git a/pkg/sql/executor_ir_test.go b/pkg/sql/executor_ir_test.go
index 5cef0c7820..1dcd1f7567 100644
--- a/pkg/sql/executor_ir_test.go
+++ b/pkg/sql/executor_ir_test.go
@@ -224,3 +224,20 @@ func TestLogChanWriter_Write(t *testing.T) {
 	_, more := <-c
 	a.False(more)
 }
+
+func TestSubmitWorkflow(t *testing.T) {
+	a := assert.New(t)
+	modelDir := ""
+	a.NotPanics(func() {
+		rd := SubmitWorkflow(testXGBoostTrainSelectIris, testDB, modelDir, getDefaultSession())
+		for r := range rd.ReadAll() {
+			switch r.(type) {
+			case WorkflowJob:
+				job := r.(WorkflowJob)
+				a.Equal(job.JobID, "sqlflow-workflow")
+			default:
+				a.Fail("SubmitWorkflow should return JobID")
+			}
+		}
+	})
+}
diff --git a/pkg/sql/ir_generator.go b/pkg/sql/ir_generator.go
index 7fbeb4287a..1c1f128c04 100644
--- a/pkg/sql/ir_generator.go
+++ b/pkg/sql/ir_generator.go
@@ -716,13 +716,19 @@ func parseResultTable(intoStatement string) (string, string, error) {
 }
 
 // programToIR generate a list of IRs from a SQL program
-func programToIR(sqls []statementParseResult, connStr, modelDir string, getTrainIRFromModel bool) (ir.SQLProgram, error) {
+func programToIR(sqls []statementParseResult, connStr, modelDir string, getTrainIRFromModel bool, enableFeatureDerivation bool) (ir.SQLProgram, error) {
 	IRs := ir.SQLProgram{}
 	for _, sql := range sqls {
 		if sql.extended != nil {
 			parsed := sql.extended
 			if parsed.train {
-				ir, err := generateTrainIRWithInferredColumns(parsed, connStr)
+				var ir *ir.TrainClause
+				var err error
+				if enableFeatureDerivation {
+					ir, err = generateTrainIRWithInferredColumns(parsed, connStr)
+				} else {
+					ir, err = generateTrainIR(parsed, connStr)
+				}
 				if err != nil {
 					return nil, err
 				}