We have some jobs that take several hours to run, and you can sometimes detect it is going to fail early so it can be useful to cancel the job. But while the job disappears from the job list, it doesn't free up the runner and the task is left running on the machine hosting the runner. The only way to kill it seems to be to connect to the runner and kill the process it was running.
How does the runner kills jobs? Gitlab-runner is running on a Linux machine and not inside a Docker container.