d2f959aa1b
Using the kubernetes client in collectd on Debian leads to a leak of eventpoll objects. CentOS does not exhibit this leak. Invoking the API through python services does not leak. Newer versions of kubernetes client do not affect the leak. The leak is related to the collectd event system and its interaction with the API. The resolution is to spawn a subprocess (kubectl) and get the results from it. The kubectl results (json) are not kubernetes objects, so utility methods have been added to convert to the object types (pod, status, metadata) that are explicitly used by the cpu and memory plugin. Note: annotations are just a dictionary in API or CLI. Test Plan: Verify: kube_get_local_pods (original code) called once per second from the cpu plugin causes a leak. Verify: kube_get_local_pods (updated code) called once per second from the cpu plugin does not cause a leak. Verify: no 'new' stacktraces in the collectd logs Closes-Bug: #1998953 Signed-off-by: Al Bailey <al.bailey@windriver.com> Change-Id: I340b27926e2614e7ab55d16d961d9b680e7d1321