Troubleshooting

While Kubernetes and the ArangoDB Kubernetes operator automatically resolve a lot of issues, there are always cases where human attention is needed.

This chapter gives your tips & tricks to help you troubleshoot deployments.

Where to look

In Kubernetes all resources can be inspected using kubectl using either the get or describe command.

To get all details of the resource (both specification & status), run the following command:

kubectl get <resource-type> <resource-name> -n <namespace> -o yaml

For example, to get the entire specification and status of an ArangoDeployment resource named my-arangodb in the default namespace, run:

kubectl get ArangoDeployment my-arango -n default -o yaml
# or shorter
kubectl get arango my-arango -o yaml

Several types of resources (including all ArangoDB custom resources) support events. These events show what happened to the resource over time.

To show the events (and most important resource data) of a resource, run the following command:

kubectl describe <resource-type> <resource-name> -n <namespace>

Getting logs

Another invaluable source of information is the log of containers being run in Kubernetes. These logs are accessible through the Pods that group these containers.

To fetch the logs of the default container running in a Pod, run:

kubectl logs <pod-name> -n <namespace>
# or with follow option to keep inspecting logs while they are written
kubectl logs <pod-name> -n <namespace> -f

To inspect the logs of a specific container in Pod, add -c <container-name>. You can find the names of the containers in the Pod, using kubectl describe pod ....

Note that the ArangoDB operators are being deployed themselves as a Kubernetes Deployment with 2 replicas. This means that you have to fetch the logs of 2 Pods running those replicas.

What if

The Pods of a deployment stay in Pending state

There are two common causes for this.

  • The Pods cannot be scheduled because there are not enough nodes available. This is usually only the case with a spec.environment setting that has a value of Production.

    Solution: Add more nodes.

  • There are no PersistentVolumes available to be bound to the PersistentVolumeClaims created by the operator.

    Solution: Use kubectl get persistentvolumes to inspect the available PersistentVolumes and if needed, use the ArangoLocalStorage operator to provision PersistentVolumes.

When restarting a Node, the Pods scheduled on that node remain in Terminating state

When a Node no longer makes regular calls to the Kubernetes API server, it is marked as not available. Depending on specific settings in your Pods, Kubernetes will at some point decide to terminate the Pod. As long as the Node is not completely removed from the Kubernetes API server, Kubernetes tries to use the Node itself to terminate the Pod.

The ArangoDeployment operator recognizes this condition and tries to replace those Pods with Pods on different nodes. The exact behavior differs per type of server.

What happens when a Node with local data is broken

When a Node with PersistentVolumes hosted on that Node is broken and cannot be repaired, the data in those PersistentVolumes is lost.

If an ArangoDeployment of type Single was using one of those PersistentVolumes the database is lost and must be restored from a backup.

If an ArangoDeployment of type ActiveFailover or Cluster was using one of those PersistentVolumes, it depends on the type of server that was using the volume.

  • If an Agent was using the volume, it can be repaired as long as 2 other Agents are still healthy.
  • If a DBServer was using the volume, and the replication factor of all database collections is 2 or higher, and the remaining DB-Servers are still healthy, the cluster duplicates the remaining replicas to bring the number of replicas back to the original number.
  • If a DBServer was using the volume, and the replication factor of a database collection is 1 and happens to be stored on that DB-Server, the data is lost.
  • If a single server of an ActiveFailover deployment was using the volume, and the other single server is still healthy, the other single server becomes leader. After replacing the failed single server, the new follower synchronizes with the leader.