Uploaded image for project: 'OpenShift Bugs'
  1. OpenShift Bugs
  2. OCPBUGS-29713

Excessive node status updates causing high control plane CPU

XMLWordPrintable

    • +
    • No
    • MCO Sprint 249, MCO Sprint 250
    • 2
    • Rejected
    • False
    • Hide

      None

      Show
      None
    • Hide
      * Previously, the default value of the `nodeStatusUpdateFrequency` parameter was changed from `0s` to `10s`. This change inadvertently caused the `nodeStatusReportFrequency` to increase significantly, because the value was linked to the `nodeStatusReportFrequency` value. This resulted in high CPU usage on control plane operators and the API server. This fix manually sets the `nodeStatusReportFrequency` value to `5m`, which prevents this high CPU usage. (link:https://issues.redhat.com/browse/OCPBUGS-29713[*OCPBUGS-29713*])
      Show
      * Previously, the default value of the `nodeStatusUpdateFrequency` parameter was changed from `0s` to `10s`. This change inadvertently caused the `nodeStatusReportFrequency` to increase significantly, because the value was linked to the `nodeStatusReportFrequency` value. This resulted in high CPU usage on control plane operators and the API server. This fix manually sets the `nodeStatusReportFrequency` value to `5m`, which prevents this high CPU usage. (link: https://issues.redhat.com/browse/OCPBUGS-29713 [* OCPBUGS-29713 *])
    • Bug Fix
    • Done

      Description of problem:

      OCPBUGS-29424 revealed that setting the node status update frequency in kubelet (introduced with OCPBUGS-15583) causes a lot of control plane CPU. 

The reason is the increased frequency of kubelet node status updates will trigger second order effects in all control plane operators that usually trigger on node changes (api server, etcd, PDB guard pod controllers, or any other static pod based machinery).

Reverting the code in OCPBUGS-15583, or manually setting the report/status frequency to 0s causes the CPU to drop immediately.

      Version-Release number of selected component (if applicable):

      Versions where OCPBUGS-15583 was backported. This includes 4.16, 4.15.0, 4.14.8, 4.13.33, and the next 4.12.z likely 4.12.51.

      How reproducible:

      always

      Steps to Reproduce:

      1. create a cluster that contains a fix for OCPBUGS-15583
2. observe the apiserver metrics (eg rate(apiserver_request_total[5m])), those should show abnormal values for pod/configmap GET
    alternatively the rate of node updates is increaed (rate(apiserver_request_total{resource="nodes", subresource="status", verb="PATCH"}[1m])) 

    

      Actual results:

      the node status updates every 10s, which causes high CPU usage on control plane operators and apiserver

      Expected results:

      the node status should not update that frequently, meaning the control plane CPU usage should go down again

      Additional info:

      slack thread with the node team:
https://redhat-internal.slack.com/archives/C02CZNQHGN8/p1708429189987849
    

              cdoern@redhat.com Charles Doern
              tjungblu@redhat.com Thomas Jungblut
              Sergio Regidor de la Rosa Sergio Regidor de la Rosa
              Votes:
              0 Vote for this issue
              Watchers:
              11 Start watching this issue

                Created:
                Updated:
                Resolved: