fix: reconcile in case pod killed abruptly (#1512)

Co-authored-by: Tamir David <tamirdavid@Tamirs-MacBook-Pro.local>
Co-authored-by: Ron Federman <73110295+RonFed@users.noreply.github.com>

odiglet/pkg/ebpf/director.go

@@ -307,7 +307,7 @@ func (d *EbpfDirector[T]) Cleanup(pod types.NamespacedName) {
 		return
 	}
 
-	log.Logger.V(0).Info("Cleaning up ebpf go instrumentation for pod", "pod", pod)
+	log.Logger.V(0).Info("Cleaning up ebpf instrumentation for pod", "pod", pod, "language", d.language)
 	delete(d.podsToDetails, pod)
 
 	// clear the pod from the workloadToPods map

odiglet/pkg/kube/instrumentation_ebpf/manager.go

@@ -37,7 +37,11 @@ func (i *podPredicate) Update(e event.UpdateEvent) bool {
 		return true
 	}
 
-	return false
+	// Sum the restart counts for both oldPod and newPod containers, then compare them.
+	// If the newPod has a higher restart count than the oldPod, we need to re-instrument it.
+	// This happens because the pod was abruptly killed, which caused an increment in the restart count.
+	// This check is required because the pod will remain running during the process kill and re-launch.
+	return GetPodSumRestarts(newPod) > GetPodSumRestarts(oldPod)
 }
 
 func (i *podPredicate) Delete(e event.DeleteEvent) bool {

odiglet/pkg/kube/instrumentation_ebpf/pods.go

@@ -123,3 +123,11 @@ func (p *PodsReconciler) getPodWorkloadObject(ctx context.Context, pod *corev1.P
 	// Pod does not necessarily have to be managed by a controller
 	return nil, nil
 }
+
+func GetPodSumRestarts(pod *corev1.Pod) int {
+	restartCount := 0
+	for _, containerStatus := range pod.Status.ContainerStatuses {
+		restartCount += int(containerStatus.RestartCount)
+	}
+	return restartCount
+}