chore(controller): ignore k8s job fail reason and refine task finish time

server/controller/src/main/java/ai/starwhale/mlops/schedule/impl/k8s/reporting/JobEventHandler.java

@@ -139,23 +139,16 @@ private void reportRunStatus(V1Job job) {
             log.warn("no pod stop time found for job {}, use now", jobName);
         }
 
-        // prefer using the pod failed reason, it has more details
+        // use the pod failed reason, it has more details
+        // we have disabled the job retry, so the job failed reason is not useful
         var podFailedReason = getPodFailedReason(jobName);
 
-        var failedReason = "";
-        if (StringUtils.hasText(jobFailedReason)) {
-            failedReason = "job failed: " + jobFailedReason;
-        }
-        if (StringUtils.hasText(podFailedReason)) {
-            failedReason = failedReason + "\npod failed: " + podFailedReason;
-        }
-
         var report = ReportedRun.builder()
                 .id(runId)
                 .status(runStatus)
                 .startTimeMillis(startTime)
                 .stopTimeMillis(stopTime)
-                .failedReason(StringUtils.hasText(failedReason) ? failedReason : null)
+                .failedReason(StringUtils.hasText(podFailedReason) ? podFailedReason : null)
                 .build();
         runReportReceiver.receive(report);
     }

server/controller/src/main/java/ai/starwhale/mlops/schedule/reporting/listener/impl/RunUpdateListenerForStatus.java

@@ -95,6 +95,8 @@ public void onRunUpdate(Run run) {
         retryNum = null == retryNum ? 0 : retryNum;
         Integer userRetryLimit = task.getStep().getSpec().getBackoffLimit();
         Integer backOffLimit = userRetryLimit == null ? this.backOffLimit : userRetryLimit;
+
+        var updateFinishedTime = true;
         if (
                 run.getStatus() == RunStatus.FAILED
                         && retryNum < backOffLimit
@@ -108,6 +110,8 @@ public void onRunUpdate(Run run) {
                 Task ot = ((WatchableTask) task).unwrap();
                 ot.updateStatus(TaskStatus.READY);
             }
+            // do not update finished time for retrying task
+            updateFinishedTime = false;
         } else {
             taskNewStatus = taskStatusMachine.transfer(task.getStatus(), run.getStatus());
         }
@@ -118,7 +122,7 @@ public void onRunUpdate(Run run) {
         if (run.getStartTime() != null) {
             taskMapper.updateTaskStartedTimeIfNotSet(run.getTaskId(), new Date(run.getStartTime()));
         }
-        if (run.getFinishTime() != null) {
+        if (run.getFinishTime() != null && updateFinishedTime) {
             taskMapper.updateTaskFinishedTimeIfNotSet(run.getTaskId(), new Date(run.getFinishTime()));
         }
         if (run.getIp() != null) {

server/controller/src/test/java/ai/starwhale/mlops/schedule/impl/k8s/JobEventHandlerTest.java

@@ -101,6 +101,7 @@ public void testOnAddFail() {
                 .stopTimeMillis(endTime.toInstant().toEpochMilli())
                 .build();
         verify(runReportReceiver).receive(expected);
+        reset(runReportReceiver);
 
         // test with reason and message
         var con = new V1JobCondition()
@@ -112,7 +113,7 @@ public void testOnAddFail() {
                 .status(RunStatus.FAILED)
                 .startTimeMillis(startTime.toInstant().toEpochMilli())
                 .stopTimeMillis(endTime.toInstant().toEpochMilli())
-                .failedReason("job failed: reason, message")
+                .failedReason(null)
                 .build();
         verify(runReportReceiver).receive(expected2);
 
@@ -128,7 +129,7 @@ public void testOnAddFail() {
                 .status(RunStatus.FAILED)
                 .startTimeMillis(startTime.toInstant().toEpochMilli())
                 .stopTimeMillis(endTime.toInstant().toEpochMilli())
-                .failedReason("job failed: reason, message\npod failed: foo, bar\nfoo, bar")
+                .failedReason("foo, bar\nfoo, bar")
                 .build();
         verify(runReportReceiver).receive(expected3);
     }

server/controller/src/test/java/ai/starwhale/mlops/schedule/reporting/listener/impl/RunUpdateListenerForStatusTest.java

@@ -18,6 +18,7 @@
 
 import static org.mockito.ArgumentMatchers.any;
 import static org.mockito.Mockito.mock;
+import static org.mockito.Mockito.never;
 import static org.mockito.Mockito.times;
 import static org.mockito.Mockito.verify;
 import static org.mockito.Mockito.when;
@@ -32,6 +33,7 @@
 import ai.starwhale.mlops.domain.task.status.TaskStatus;
 import ai.starwhale.mlops.domain.task.status.TaskStatusMachine;
 import ai.starwhale.mlops.schedule.SwTaskScheduler;
+import java.util.Date;
 import org.junit.jupiter.api.BeforeEach;
 import org.junit.jupiter.api.Test;
 import org.mockito.Mockito;
@@ -103,4 +105,32 @@ public void testOnUpdate() {
 
     }
 
+    @Test
+    public void testFinishTime() {
+        var task = Task.builder()
+                .retryNum(0)
+                .step(Step.builder().spec(StepSpec.builder().backoffLimit(2).build()).build())
+                .currentRun(Run.builder().id(1L).status(RunStatus.RUNNING).build())
+                .build();
+        when(hotJobHolder.taskWithId(1L)).thenReturn(task);
+        var reportdRun = Run.builder()
+                .id(1L)
+                .taskId(1L)
+                .status(RunStatus.FAILED)
+                .finishTime(42L)
+                .ip("1.1.1.1")
+                .build();
+        runReportReceiver.onRunUpdate(reportdRun);
+        verify(this.taskMapper, never()).updateTaskFinishedTimeIfNotSet(any(), any());
+
+        // fail again
+        task.getCurrentRun().setStatus(RunStatus.RUNNING);
+        runReportReceiver.onRunUpdate(reportdRun);
+        verify(this.taskMapper, never()).updateTaskFinishedTimeIfNotSet(any(), any());
+
+        // the last attempt
+        task.getCurrentRun().setStatus(RunStatus.RUNNING);
+        runReportReceiver.onRunUpdate(reportdRun);
+        verify(this.taskMapper, times(1)).updateTaskFinishedTimeIfNotSet(1L, new Date(42L));
+    }
 }