Merge pull request 'fix startall stopall bugs' (#304) from tzwang/pcm-coordinator:master into master

Former-commit-id: bd27fac1804472b4ade48b3775bdb3715b1f4ccb

internal/logic/inference/startallbydeploytaskidlogic.go

@@ -3,12 +3,15 @@ package inference
 import (
 	"context"
 	"errors"
+	"fmt"
 	"github.com/zeromicro/go-zero/core/logx"
 	"gitlink.org.cn/JointCloud/pcm-coordinator/internal/scheduler/service/inference"
 	"gitlink.org.cn/JointCloud/pcm-coordinator/internal/storeLink"
 	"gitlink.org.cn/JointCloud/pcm-coordinator/internal/svc"
 	"gitlink.org.cn/JointCloud/pcm-coordinator/internal/types"
+	"gitlink.org.cn/JointCloud/pcm-coordinator/pkg/models"
 	"strconv"
+	"sync"
 )
 
 type StartAllByDeployTaskIdLogic struct {
@@ -35,17 +38,8 @@ func (l *StartAllByDeployTaskIdLogic) StartAllByDeployTaskId(req *types.StartAll
 		return nil, err
 	}
 
-	for _, ins := range list {
+	if len(list) == 0 {
-		in, err := l.svcCtx.Scheduler.AiService.InferenceAdapterMap[strconv.FormatInt(ins.AdapterId, 10)][strconv.FormatInt(ins.ClusterId, 10)].GetInferDeployInstance(l.ctx, ins.InstanceId)
+		return nil, errors.New("instances are empty")
-		if err != nil {
-			return nil, err
-		}
-		if checkStopStatus(in) {
-			success := l.svcCtx.Scheduler.AiService.InferenceAdapterMap[strconv.FormatInt(ins.AdapterId, 10)][strconv.FormatInt(ins.ClusterId, 10)].StartInferDeployInstance(l.ctx, ins.InstanceId)
-			if !success {
-				return nil, errors.New(ins.InstanceName + " start failed")
-			}
-		}
 	}
 
 	err = l.svcCtx.Scheduler.AiStorages.UpdateDeployTaskById(id)
@@ -53,9 +47,91 @@ func (l *StartAllByDeployTaskIdLogic) StartAllByDeployTaskId(req *types.StartAll
 		return nil, err
 	}
 
+	err = l.startAll(list)
+	if err != nil {
+		return nil, err
+	}
+
 	return resp, nil
 }
 
+func (l *StartAllByDeployTaskIdLogic) startAll(list []*models.AiInferDeployInstance) error {
+	var wg sync.WaitGroup
+	var errCh = make(chan interface{}, len(list))
+	var errs []interface{}
+	buf := make(chan bool, 2)
+
+	for _, instance := range list {
+		wg.Add(1)
+		ins := instance
+		buf <- true
+		go func() {
+			in, err := l.svcCtx.Scheduler.AiService.InferenceAdapterMap[strconv.FormatInt(ins.AdapterId, 10)][strconv.FormatInt(ins.ClusterId, 10)].GetInferDeployInstance(l.ctx, ins.InstanceId)
+			if err != nil {
+				e := struct {
+					errTyp       uint8
+					err          error
+					instanceName string
+					clusterName  string
+				}{
+					errTyp:       1,
+					err:          err,
+					instanceName: ins.InstanceName,
+					clusterName:  ins.ClusterName,
+				}
+				errCh <- e
+				wg.Done()
+				<-buf
+			}
+			if checkStopStatus(in) {
+				success := l.svcCtx.Scheduler.AiService.InferenceAdapterMap[strconv.FormatInt(ins.AdapterId, 10)][strconv.FormatInt(ins.ClusterId, 10)].StartInferDeployInstance(l.ctx, ins.InstanceId)
+				if !success {
+					e := struct {
+						errTyp       uint8
+						err          error
+						instanceName string
+						clusterName  string
+					}{
+						errTyp:       2,
+						err:          err,
+						instanceName: ins.InstanceName,
+						clusterName:  ins.ClusterName,
+					}
+					errCh <- e
+					wg.Done()
+					<-buf
+				}
+			}
+		}()
+		<-buf
+	}
+
+	for e := range errCh {
+		errs = append(errs, e)
+	}
+
+	if len(errs) != 0 {
+		var msg string
+		for _, err := range errs {
+			e := (err).(struct {
+				errTyp       uint8
+				err          error
+				instanceName string
+				clusterName  string
+			})
+			switch e.errTyp {
+			case 1:
+				msg += fmt.Sprintf("GetInstance Failed # clusterName: %v , instanceName: %v , error: %v \n", e.clusterName, e.instanceName, e.err.Error())
+			case 2:
+				msg += fmt.Sprintf("StartInstance Failed # clusterName: %v , instanceName: %v , error: %v \n", e.clusterName, e.instanceName, e.err.Error())
+			}
+		}
+		return errors.New(msg)
+	}
+
+	return nil
+}
+
 func checkStopStatus(in *inference.DeployInstance) bool {
 	switch in.ClusterType {
 	case storeLink.TYPE_OCTOPUS:

internal/logic/inference/stopallbydeploytaskidlogic.go

@@ -3,11 +3,14 @@ package inference
 import (
 	"context"
 	"errors"
+	"fmt"
 	"gitlink.org.cn/JointCloud/pcm-coordinator/internal/scheduler/service/inference"
 	"gitlink.org.cn/JointCloud/pcm-coordinator/internal/storeLink"
 	"gitlink.org.cn/JointCloud/pcm-coordinator/internal/svc"
 	"gitlink.org.cn/JointCloud/pcm-coordinator/internal/types"
+	"gitlink.org.cn/JointCloud/pcm-coordinator/pkg/models"
 	"strconv"
+	"sync"
 
 	"github.com/zeromicro/go-zero/core/logx"
 )
@@ -36,17 +39,8 @@ func (l *StopAllByDeployTaskIdLogic) StopAllByDeployTaskId(req *types.StopAllByD
 		return nil, err
 	}
 
-	for _, ins := range list {
+	if len(list) == 0 {
-		in, err := l.svcCtx.Scheduler.AiService.InferenceAdapterMap[strconv.FormatInt(ins.AdapterId, 10)][strconv.FormatInt(ins.ClusterId, 10)].GetInferDeployInstance(l.ctx, ins.InstanceId)
+		return nil, errors.New("instances are empty")
-		if err != nil {
-			return nil, err
-		}
-		if checkStatus(in) {
-			success := l.svcCtx.Scheduler.AiService.InferenceAdapterMap[strconv.FormatInt(ins.AdapterId, 10)][strconv.FormatInt(ins.ClusterId, 10)].StopInferDeployInstance(l.ctx, ins.InstanceId)
-			if !success {
-				return nil, errors.New(ins.InstanceName + " stop failed")
-			}
-		}
 	}
 
 	err = l.svcCtx.Scheduler.AiStorages.UpdateDeployTaskById(id)
@@ -54,9 +48,91 @@ func (l *StopAllByDeployTaskIdLogic) StopAllByDeployTaskId(req *types.StopAllByD
 		return nil, err
 	}
 
+	err = l.stopAll(list)
+	if err != nil {
+		return nil, err
+	}
+
 	return resp, nil
 }
 
+func (l *StopAllByDeployTaskIdLogic) stopAll(list []*models.AiInferDeployInstance) error {
+	var wg sync.WaitGroup
+	var errCh = make(chan interface{}, len(list))
+	var errs []interface{}
+	buf := make(chan bool, 2)
+
+	for _, instance := range list {
+		wg.Add(1)
+		ins := instance
+		buf <- true
+		go func() {
+			in, err := l.svcCtx.Scheduler.AiService.InferenceAdapterMap[strconv.FormatInt(ins.AdapterId, 10)][strconv.FormatInt(ins.ClusterId, 10)].GetInferDeployInstance(l.ctx, ins.InstanceId)
+			if err != nil {
+				e := struct {
+					errTyp       uint8
+					err          error
+					instanceName string
+					clusterName  string
+				}{
+					errTyp:       1,
+					err:          err,
+					instanceName: ins.InstanceName,
+					clusterName:  ins.ClusterName,
+				}
+				errCh <- e
+				wg.Done()
+				<-buf
+			}
+			if checkStatus(in) {
+				success := l.svcCtx.Scheduler.AiService.InferenceAdapterMap[strconv.FormatInt(ins.AdapterId, 10)][strconv.FormatInt(ins.ClusterId, 10)].StopInferDeployInstance(l.ctx, ins.InstanceId)
+				if !success {
+					e := struct {
+						errTyp       uint8
+						err          error
+						instanceName string
+						clusterName  string
+					}{
+						errTyp:       2,
+						err:          err,
+						instanceName: ins.InstanceName,
+						clusterName:  ins.ClusterName,
+					}
+					errCh <- e
+					wg.Done()
+					<-buf
+				}
+			}
+		}()
+		<-buf
+	}
+
+	for e := range errCh {
+		errs = append(errs, e)
+	}
+
+	if len(errs) != 0 {
+		var msg string
+		for _, err := range errs {
+			e := (err).(struct {
+				errTyp       uint8
+				err          error
+				instanceName string
+				clusterName  string
+			})
+			switch e.errTyp {
+			case 1:
+				msg += fmt.Sprintf("GetInstance Failed # clusterName: %v , instanceName: %v , error: %v \n", e.clusterName, e.instanceName, e.err.Error())
+			case 2:
+				msg += fmt.Sprintf("StopInstance Failed # clusterName: %v , instanceName: %v , error: %v \n", e.clusterName, e.instanceName, e.err.Error())
+			}
+		}
+		return errors.New(msg)
+	}
+
+	return nil
+}
+
 func checkStatus(in *inference.DeployInstance) bool {
 	switch in.ClusterType {
 	case storeLink.TYPE_OCTOPUS: