Fix OOM status reports for Task/Batch APIs

pkg/lib/k8s/pod.go

@@ -20,7 +20,6 @@ import (
 	"bytes"
 	"context"
 	"regexp"
-	"strings"
 	"time"
 
 	"github.com/cortexlabs/cortex/pkg/lib/errors"
@@ -38,7 +37,12 @@ var _podTypeMeta = kmeta.TypeMeta{
 	Kind:       "Pod",
 }
 
-const ReasonEvicted = "Evicted"
+// pod termination reasons
+// https://github.com/kubernetes/kube-state-metrics/blob/master/docs/pod-metrics.md
+const (
+	ReasonEvicted   = "Evicted"
+	ReasonOOMKilled = "OOMKilled"
+)
 
 type PodStatus string
 
@@ -148,22 +152,14 @@ func WasPodOOMKilled(pod *kcore.Pod) bool {
 		return true
 	}
 	for _, containerStatus := range pod.Status.ContainerStatuses {
+		var reason string
 		if containerStatus.LastTerminationState.Terminated != nil {
-			exitCode := containerStatus.LastTerminationState.Terminated.ExitCode
-			reason := strings.ToLower(containerStatus.LastTerminationState.Terminated.Reason)
-			if _killStatuses[exitCode] {
-				if strings.Contains(reason, "oom") {
-					return true
-				}
-			}
+			reason = containerStatus.LastTerminationState.Terminated.Reason
 		} else if containerStatus.State.Terminated != nil {
-			exitCode := containerStatus.State.Terminated.ExitCode
-			reason := strings.ToLower(containerStatus.State.Terminated.Reason)
-			if _killStatuses[exitCode] {
-				if strings.Contains(reason, "oom") {
-					return true
-				}
-			}
+			reason = containerStatus.State.Terminated.Reason
+		}
+		if reason == ReasonOOMKilled {
+			return true
 		}
 	}
 
@@ -194,25 +190,21 @@ func GetPodStatus(pod *kcore.Pod) PodStatus {
 		}
 
 		for _, containerStatus := range pod.Status.ContainerStatuses {
+			var reason string
+			var exitCode int32
 			if containerStatus.LastTerminationState.Terminated != nil {
-				exitCode := containerStatus.LastTerminationState.Terminated.ExitCode
-				reason := strings.ToLower(containerStatus.LastTerminationState.Terminated.Reason)
-				if _killStatuses[exitCode] {
-					if strings.Contains(reason, "oom") {
-						return PodStatusKilledOOM
-					}
-					return PodStatusKilled
-				}
+				reason = containerStatus.LastTerminationState.Terminated.Reason
+				exitCode = containerStatus.LastTerminationState.Terminated.ExitCode
 			} else if containerStatus.State.Terminated != nil {
-				exitCode := containerStatus.State.Terminated.ExitCode
-				reason := strings.ToLower(containerStatus.State.Terminated.Reason)
-				if _killStatuses[exitCode] {
-					if strings.Contains(reason, "oom") {
-						return PodStatusKilledOOM
-					}
-					return PodStatusKilled
-				}
+				reason = containerStatus.State.Terminated.Reason
+				exitCode = containerStatus.State.Terminated.ExitCode
 			}
+			if reason == ReasonOOMKilled {
+				return PodStatusKilledOOM
+			} else if _killStatuses[exitCode] {
+				return PodStatusKilled
+			}
+
 		}
 		return PodStatusFailed
 	case kcore.PodRunning:
@@ -245,29 +237,25 @@ func PodStatusFromContainerStatuses(containerStatuses []kcore.ContainerStatus) P
 			numRunning++
 		} else if containerStatus.State.Terminated != nil {
 			exitCode := containerStatus.State.Terminated.ExitCode
-			reason := strings.ToLower(containerStatus.State.Terminated.Reason)
-			if exitCode == 0 {
+			reason := containerStatus.State.Terminated.Reason
+			if reason == ReasonOOMKilled {
+				numKilledOOM++
+			} else if exitCode == 0 {
 				numSucceeded++
 			} else if _killStatuses[exitCode] {
-				if strings.Contains(reason, "oom") {
-					numKilledOOM++
-				} else {
-					numKilled++
-				}
+				numKilled++
 			} else {
 				numFailed++
 			}
 		} else if containerStatus.LastTerminationState.Terminated != nil {
 			exitCode := containerStatus.LastTerminationState.Terminated.ExitCode
-			reason := strings.ToLower(containerStatus.LastTerminationState.Terminated.Reason)
-			if exitCode == 0 {
+			reason := containerStatus.LastTerminationState.Terminated.Reason
+			if reason == ReasonOOMKilled {
+				numKilledOOM++
+			} else if exitCode == 0 {
 				numSucceeded++
 			} else if _killStatuses[exitCode] {
-				if strings.Contains(reason, "oom") {
-					numKilledOOM++
-				} else {
-					numKilled++
-				}
+				numKilled++
 			} else {
 				numFailed++
 			}

pkg/operator/main.go

@@ -47,7 +47,7 @@ func main() {
 
 	telemetry.Event("operator.init", map[string]interface{}{"provider": config.Provider})
 
-	cron.Run(operator.DeleteEvictedPods, operator.ErrorHandler("delete evicted pods"), 12*time.Hour)
+	cron.Run(operator.DeleteEvictedPods, operator.ErrorHandler("delete evicted pods"), time.Hour)
 
 	switch config.Provider {
 	case types.AWSProviderType:

pkg/operator/operator/cron.go

@@ -22,6 +22,7 @@ import (
 	"github.com/cortexlabs/cortex/pkg/lib/aws"
 	"github.com/cortexlabs/cortex/pkg/lib/errors"
 	"github.com/cortexlabs/cortex/pkg/lib/k8s"
+	"github.com/cortexlabs/cortex/pkg/lib/sets/strset"
 	"github.com/cortexlabs/cortex/pkg/lib/telemetry"
 	"github.com/cortexlabs/cortex/pkg/operator/config"
 	"github.com/cortexlabs/cortex/pkg/operator/lib/logging"
@@ -30,6 +31,7 @@ import (
 )
 
 var operatorLogger = logging.GetOperatorLogger()
+var previousListOfEvictedPods = strset.New()
 
 func DeleteEvictedPods() error {
 	failedPods, err := config.K8s.ListPods(&kmeta.ListOptions{
@@ -40,14 +42,21 @@ func DeleteEvictedPods() error {
 	}
 
 	var errs []error
+	currentEvictedPods := strset.New()
 	for _, pod := range failedPods {
-		if pod.Status.Reason == k8s.ReasonEvicted {
+		if pod.Status.Reason != k8s.ReasonEvicted {
+			continue
+		}
+		if previousListOfEvictedPods.Has(pod.Name) {
 			_, err := config.K8s.DeletePod(pod.Name)
 			if err != nil {
 				errs = append(errs, err)
 			}
+			continue
 		}
+		currentEvictedPods.Add(pod.Name)
 	}
+	previousListOfEvictedPods = currentEvictedPods
 
 	if errors.HasError(errs) {
 		return errors.FirstError(errs...)

pkg/operator/resources/job/batchapi/cron.go

@@ -284,8 +284,9 @@ func reconcileInProgressJob(jobState *job.State, queueURL *string, k8sJob *kbatc
 }
 
 func checkIfJobCompleted(jobKey spec.JobKey, queueURL string, k8sJob *kbatch.Job) error {
-	if int(k8sJob.Status.Failed) > 0 {
-		return investigateJobFailure(jobKey)
+	jobFailed, err := checkForJobFailure(jobKey, k8sJob)
+	if err != nil || jobFailed {
+		return err
 	}
 
 	queueMessages, err := getQueueMetricsFromURL(queueURL)
@@ -349,19 +350,18 @@ func checkIfJobCompleted(jobKey spec.JobKey, queueURL string, k8sJob *kbatch.Job
 	return nil
 }
 
-func investigateJobFailure(jobKey spec.JobKey) error {
-	reasonFound := false
-
+func checkForJobFailure(jobKey spec.JobKey, k8sJob *kbatch.Job) (bool, error) {
 	jobLogger, err := operator.GetJobLogger(jobKey)
 	if err != nil {
-		return err
+		return false, err
 	}
 
+	reasonFound := false
 	pods, _ := config.K8s.ListPodsByLabel("jobID", jobKey.ID)
 	for _, pod := range pods {
 		if k8s.WasPodOOMKilled(&pod) {
 			jobLogger.Error("at least one worker was killed because it ran out of out of memory")
-			return errors.FirstError(
+			return true, errors.FirstError(
 				job.SetWorkerOOMStatus(jobKey),
 				deleteJobRuntimeResources(jobKey),
 			)
@@ -382,13 +382,21 @@ func investigateJobFailure(jobKey spec.JobKey) error {
 		}
 	}
 
-	if !reasonFound {
-		jobLogger.Error("workers were killed for unknown reason")
+	if int(k8sJob.Status.Failed) > 0 {
+		if !reasonFound {
+			jobLogger.Error("workers were killed for unknown reason")
+		}
+		return true, errors.FirstError(
+			job.SetWorkerErrorStatus(jobKey),
+			deleteJobRuntimeResources(jobKey),
+		)
+	} else if int(k8sJob.Status.Succeeded) == 1 && len(pods) == 0 {
+		// really unexpected situation which doesn't hurt if we check
+		return true, errors.FirstError(
+			job.SetUnexpectedErrorStatus(jobKey),
+			deleteJobRuntimeResources(jobKey),
+		)
 	}
 
-	return errors.FirstError(
-		err,
-		job.SetWorkerErrorStatus(jobKey),
-		deleteJobRuntimeResources(jobKey),
-	)
+	return false, nil
 }

pkg/operator/resources/job/taskapi/cron.go

@@ -202,27 +202,31 @@ func reconcileInProgressJob(jobState *job.State, k8sJob *kbatch.Job) (status.Job
 }
 
 func checkIfJobCompleted(jobKey spec.JobKey, k8sJob *kbatch.Job) error {
-	if int(k8sJob.Status.Failed) == 1 {
-		pods, _ := config.K8s.ListPodsByLabel("jobID", jobKey.ID)
-		for _, pod := range pods {
-			if k8s.WasPodOOMKilled(&pod) {
-				return errors.FirstError(
-					job.SetWorkerOOMStatus(jobKey),
-					deleteJobRuntimeResources(jobKey),
-				)
-			}
+	pods, _ := config.K8s.ListPodsByLabel("jobID", jobKey.ID)
+	for _, pod := range pods {
+		if k8s.WasPodOOMKilled(&pod) {
+			return errors.FirstError(
+				job.SetWorkerOOMStatus(jobKey),
+				deleteJobRuntimeResources(jobKey),
+			)
 		}
+	}
+	if int(k8sJob.Status.Failed) == 1 {
 		return errors.FirstError(
 			job.SetWorkerErrorStatus(jobKey),
 			deleteJobRuntimeResources(jobKey),
 		)
-	}
-
-	if int(k8sJob.Status.Succeeded) == 1 {
+	} else if int(k8sJob.Status.Succeeded) == 1 && len(pods) > 0 {
 		return errors.FirstError(
 			job.SetSucceededStatus(jobKey),
 			deleteJobRuntimeResources(jobKey),
 		)
+	} else if int(k8sJob.Status.Succeeded) == 1 && len(pods) == 0 {
+		// really unexpected situation which doesn't hurt if we check
+		return errors.FirstError(
+			job.SetUnexpectedErrorStatus(jobKey),
+			deleteJobRuntimeResources(jobKey),
+		)
 	}
 
 	return nil