Killing Jobs

A Job can be killed after it is started, either manually, through a timeout, or other means (such as parallel completion).

Using the CLI

The easiest way to kill a Job execution is to use the furiko CLI tool:

furiko kill <job-name>

Kill Timestamp

Each Job has a killTimestamp in its spec, which specifies the timestamp before the JobController tries to kill all of its active tasks.

To kill a Job explicitly, set killTimestamp to the current time or earlier:

apiVersion: execution.furiko.io/v1alpha1
kind: Job
spec:
  killTimestamp: "2021-01-01T00:00:00Z"

You can also set a time in the future to kill the Job by setting a future time. The JobController will only start killing active tasks once killTimestamp is passed.

Setting a kill timestamp does not imply that the Job will be terminated by time that the kill timestamp has been reached. It simply means that

Automatic Termination

A Job's task may be automatically terminated from several sources, including but not limited to:

• Task pending timeout
• Completion of other parallel tasks

Termination Process

The JobController uses deletion to kill tasks, and optionally reaches for force deletion as a last resort if the task is still not deleted after a timeout.

Task Deletion

The JobController uses normal resource deletion to kill a task.

If the pod task executor is used, normal graceful termination will apply and container termination lifecycle hooks will be triggered. To adjust the grace period, use terminationGracePeriodSeconds. For more information, see Termination of Pods on the Kubernetes documentation.

Force Deletion

When a Pod is running on a Node that is no longer available, above methods to kill the Job will have no effect until the Node comes back up. As a last resort, the controller may opt to use force deletion to kill the task instead, which does not block any forward progress of the JobConfig.

For more details and implications of using such a termination method, see the Force Deletion documentation.