Skip to content

Latest commit

 

History

History
44 lines (28 loc) · 2.14 KB

leader_election.md

File metadata and controls

44 lines (28 loc) · 2.14 KB
Raw

Leader election

With leader election you can have one responsible pod. This can be used to control that only one pod runs a batch-job or similar tasks. This is done by asking the elector pod which pod is the current leader, and comparing that to the pods hostname.

The leader election configuration does not control which pod the external service requests will be routed to.

Enable leader election

Enabling leader election in your pod is done by adding the line leaderElection: true to your nais.yaml-file. With that setting, naisd will sidecar an elector container into your pod.

When you have the elector container running in your pod, you can run a HTTP GET on the URL set in environment variable $ELECTOR_PATH to see which pod is the leader. This will return a json with the name of the leader, which you can now compare with your hostname.

Examples

Code example

// Implementation of getJSONFromUrl is left as an exercise for the reader
public boolean isLeader() {
    String electorPath = System.getenv("ELECTOR_PATH");
    JSONObject leaderJson = getJSONFromUrl(electorPath);
    String leader = leaderJson.getString("name");
    String hostname = InetAddress.getLocalHost().getHostname();

    return hostname.equals(leader);
}

cURL example

$ kubectl exec -it elector-sidecar-755b7c5795-7k2qn -c debug bash
root@elector-sidecar-755b7c5795-7k2qn:/# curl $ELECTOR_PATH
{"name":"elector-sidecar-755b7c5795-2kcm7"}

Issues

  • This is not really an issue with NAIS but: Kubernetes' leader-election-image does not support fencing, which means it does not guarantee there is only one leader. Since NAIS keeps the date and time in the pods syncronized this is not an issue. There will always be only one leader.
  • Redeploy of the deployment/pods will sometime make the non-leaders believe that the old leader still exists and is the leader. The current leader is not affected, and will be aware that the pod itself is the leader. This is not really an issue as long as you do not need to know exactly which pod is the leader. This can be resolved by deleting pods with erroneous leader election configuration.