You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

statusSync.go 6.7 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216
  1. package status
  2. import (
  3. "errors"
  4. "fmt"
  5. "github.com/zeromicro/go-zero/core/logx"
  6. "gitlink.org.cn/JointCloud/pcm-coordinator/internal/config"
  7. "gitlink.org.cn/JointCloud/pcm-coordinator/internal/scheduler/database"
  8. "gitlink.org.cn/JointCloud/pcm-coordinator/internal/scheduler/service/collector"
  9. "gitlink.org.cn/JointCloud/pcm-coordinator/internal/scheduler/service/utils/jcs"
  10. "gitlink.org.cn/JointCloud/pcm-coordinator/internal/types"
  11. "gitlink.org.cn/JointCloud/pcm-coordinator/pkg/constants"
  12. "gitlink.org.cn/JointCloud/pcm-coordinator/pkg/models"
  13. "google.golang.org/grpc/codes"
  14. "google.golang.org/grpc/status"
  15. "net/http"
  16. "strconv"
  17. "sync"
  18. )
  19. type TaskStatus struct {
  20. taskSyncLock sync.Mutex
  21. aiStorages *database.AiStorage
  22. aiCollectorAdapterMap map[string]map[string]collector.AiCollector
  23. config *config.Config
  24. }
  25. func NewTaskStatus(storage *database.AiStorage, aiCollectorAdapterMap map[string]map[string]collector.AiCollector, config *config.Config) *TaskStatus {
  26. return &TaskStatus{
  27. aiStorages: storage,
  28. aiCollectorAdapterMap: aiCollectorAdapterMap,
  29. config: config,
  30. }
  31. }
  32. func (s *TaskStatus) UpdateAiTaskStatus(tasklist []*types.TaskModel) {
  33. s.taskSyncLock.Lock()
  34. defer s.taskSyncLock.Unlock()
  35. list := make([]*types.TaskModel, len(tasklist))
  36. copy(list, tasklist)
  37. for i := len(list) - 1; i >= 0; i-- {
  38. if list[i].AdapterTypeDict != "1" || list[i].Status == constants.Succeeded || list[i].Status == constants.Failed {
  39. list = append(list[:i], list[i+1:]...)
  40. }
  41. }
  42. if len(list) == 0 {
  43. return
  44. }
  45. for _, task := range list {
  46. aiTaskList, err := s.aiStorages.GetAiTaskListById(task.Id)
  47. if err != nil {
  48. panic(err.Error())
  49. }
  50. if len(aiTaskList) == 0 {
  51. continue
  52. }
  53. go s.updateAiTask(aiTaskList)
  54. }
  55. }
  56. func (s *TaskStatus) UpdateTaskStatus(tasklist []*types.TaskModel) {
  57. s.taskSyncLock.Lock()
  58. defer s.taskSyncLock.Unlock()
  59. list := make([]*types.TaskModel, len(tasklist))
  60. copy(list, tasklist)
  61. for i := len(list) - 1; i >= 0; i-- {
  62. if list[i].AdapterTypeDict != "1" || list[i].Status == constants.Succeeded || list[i].Status == constants.Failed || list[i].Status == constants.Cancelled {
  63. list = append(list[:i], list[i+1:]...)
  64. }
  65. }
  66. if len(list) == 0 {
  67. return
  68. }
  69. for _, task := range list {
  70. aiTask, err := s.aiStorages.GetAiTaskListById(task.Id)
  71. if err != nil {
  72. panic(err.Error())
  73. }
  74. if len(aiTask) == 0 {
  75. continue
  76. }
  77. logx.Errorf("############ Report Status Message Before switch %s", task.Status)
  78. if len(aiTask) == 1 {
  79. logx.Errorf("############ Report Status Message Switch %s", aiTask[0].Status)
  80. switch aiTask[0].Status {
  81. case constants.Completed:
  82. task.Status = constants.Succeeded
  83. logx.Errorf("############ Report Status Message Before Sending %s", task.Status)
  84. _ = s.reportStatusMessages(task, aiTask[0])
  85. case constants.Failed:
  86. task.Status = constants.Failed
  87. logx.Errorf("############ Report Status Message Before Sending %s", task.Status)
  88. _ = s.reportStatusMessages(task, aiTask[0])
  89. default:
  90. task.Status = aiTask[0].Status
  91. }
  92. task.StartTime = aiTask[0].StartTime
  93. task.EndTime = aiTask[0].EndTime
  94. err := s.aiStorages.UpdateTask(task)
  95. if err != nil {
  96. panic(err.Error())
  97. }
  98. }
  99. }
  100. }
  101. func (s *TaskStatus) updateAiTask(aiTaskList []*models.TaskAi) {
  102. var wg sync.WaitGroup
  103. for _, aitask := range aiTaskList {
  104. t := aitask
  105. if t.Status == constants.Completed || t.Status == constants.Failed || t.JobId == "" || t.Status == constants.Cancelled {
  106. continue
  107. }
  108. wg.Add(1)
  109. go func() {
  110. h := http.Request{}
  111. trainingTask, err := s.aiCollectorAdapterMap[strconv.FormatInt(t.AdapterId, 10)][strconv.FormatInt(t.ClusterId, 10)].GetTrainingTask(h.Context(), t.JobId)
  112. if err != nil {
  113. if status.Code(err) == codes.DeadlineExceeded {
  114. msg := fmt.Sprintf("###UpdateAiTaskStatus###, AiTaskId: %v, clusterId: %v , JobId: %v, error: %v \n", t.Id, t.ClusterId, t.JobId, err.Error())
  115. logx.Errorf(errors.New(msg).Error())
  116. wg.Done()
  117. return
  118. }
  119. msg := fmt.Sprintf("###UpdateAiTaskStatus###, AiTaskId: %v, clusterId: %v , JobId: %v, error: %v \n", t.Id, t.ClusterId, t.JobId, err.Error())
  120. logx.Errorf(errors.New(msg).Error())
  121. wg.Done()
  122. return
  123. }
  124. if trainingTask == nil {
  125. wg.Done()
  126. return
  127. }
  128. switch trainingTask.Status {
  129. case constants.Running:
  130. if t.Status != trainingTask.Status {
  131. s.aiStorages.AddNoticeInfo(strconv.FormatInt(t.AdapterId, 10), t.AdapterName, strconv.FormatInt(t.ClusterId, 10), t.ClusterName, t.Name, "running", "任务运行中")
  132. t.Status = trainingTask.Status
  133. }
  134. case constants.Failed:
  135. if t.Status != trainingTask.Status {
  136. s.aiStorages.AddNoticeInfo(strconv.FormatInt(t.AdapterId, 10), t.AdapterName, strconv.FormatInt(t.ClusterId, 10), t.ClusterName, t.Name, "failed", "任务失败")
  137. t.Status = trainingTask.Status
  138. }
  139. case constants.Completed:
  140. if t.Status != trainingTask.Status {
  141. s.aiStorages.AddNoticeInfo(strconv.FormatInt(t.AdapterId, 10), t.AdapterName, strconv.FormatInt(t.ClusterId, 10), t.ClusterName, t.Name, "completed", "任务完成")
  142. t.Status = trainingTask.Status
  143. }
  144. default:
  145. if t.Status != trainingTask.Status {
  146. s.aiStorages.AddNoticeInfo(strconv.FormatInt(t.AdapterId, 10), t.AdapterName, strconv.FormatInt(t.ClusterId, 10), t.ClusterName, t.Name, "pending", "任务pending")
  147. t.Status = trainingTask.Status
  148. }
  149. }
  150. t.StartTime = trainingTask.Start
  151. t.EndTime = trainingTask.End
  152. err = s.aiStorages.UpdateAiTask(t)
  153. if err != nil {
  154. msg := fmt.Sprintf("###UpdateAiTaskStatus###, AiTaskId: %v, clusterId: %v , JobId: %v, error: %v \n", t.Id, t.ClusterId, t.JobId, err.Error())
  155. logx.Errorf(errors.New(msg).Error())
  156. wg.Done()
  157. return
  158. }
  159. wg.Done()
  160. }()
  161. }
  162. wg.Wait()
  163. }
  164. func (s *TaskStatus) reportStatusMessages(task *types.TaskModel, aiTask *models.TaskAi) error {
  165. report := &jcs.JobStatusReportReq{
  166. TaskName: task.Name,
  167. TaskID: strconv.FormatInt(task.Id, 10),
  168. Messages: make([]*jcs.ReportMessage, 0),
  169. }
  170. //add report msg
  171. var output string
  172. switch aiTask.ClusterName {
  173. case "openI":
  174. output = aiTask.JobId
  175. case "鹏城云脑II-modelarts":
  176. output = aiTask.Output
  177. }
  178. jobMsg := &jcs.ReportMessage{
  179. Status: true,
  180. Message: "",
  181. ClusterID: strconv.FormatInt(aiTask.ClusterId, 10),
  182. Output: output,
  183. }
  184. report.Messages = append(report.Messages, jobMsg)
  185. _ = jcs.StatusReport(s.config.JcsMiddleware.JobStatusReportUrl, report)
  186. err := jcs.TempSaveReportToTask(s.aiStorages, task, report)
  187. if err != nil {
  188. return err
  189. }
  190. return nil
  191. }

PCM is positioned as Software stack over Cloud, aiming to build the standards and ecology of heterogeneous cloud collaboration for JCC in a non intrusive and autonomous peer-to-peer manner.