Merge pull request #4282 from tstromberg/apiserver-no-oom
Set apiserver oom_adj to -10 to avoid OOMing before other podspull/4296/head
commit
f74c62a667
|
@ -258,6 +258,31 @@ func (k *Bootstrapper) StartCluster(k8s config.KubernetesConfig) error {
|
|||
if err := util.RetryAfter(100, elevateKubeSystemPrivileges, time.Millisecond*500); err != nil {
|
||||
return errors.Wrap(err, "timed out waiting to elevate kube-system RBAC privileges")
|
||||
}
|
||||
|
||||
if err := k.adjustResourceLimits(); err != nil {
|
||||
glog.Warningf("unable to adjust resource limits: %v", err)
|
||||
}
|
||||
return nil
|
||||
}
|
||||
|
||||
// adjustResourceLimits makes fine adjustments to pod resources that aren't possible via kubeadm config.
|
||||
func (k *Bootstrapper) adjustResourceLimits() error {
|
||||
score, err := k.c.CombinedOutput("cat /proc/$(pgrep kube-apiserver)/oom_adj")
|
||||
if err != nil {
|
||||
return errors.Wrap(err, "oom_adj check")
|
||||
}
|
||||
glog.Infof("apiserver oom_adj: %s", score)
|
||||
// oom_adj is already a negative number
|
||||
if strings.HasPrefix(score, "-") {
|
||||
return nil
|
||||
}
|
||||
glog.Infof("adjusting apiserver oom_adj to -10")
|
||||
|
||||
// Prevent the apiserver from OOM'ing before other pods, as it is our gateway into the cluster.
|
||||
// It'd be preferable to do this via Kubernetes, but kubeadm doesn't have a way to set pod QoS.
|
||||
if err := k.c.Run("echo -10 | sudo tee /proc/$(pgrep kube-apiserver)/oom_adj"); err != nil {
|
||||
return errors.Wrap(err, "oom_adj adjust")
|
||||
}
|
||||
return nil
|
||||
}
|
||||
|
||||
|
@ -361,6 +386,10 @@ func (k *Bootstrapper) RestartCluster(k8s config.KubernetesConfig) error {
|
|||
if err := k.c.Run(fmt.Sprintf("%s phase addon all --config %s", baseCmd, configPath)); err != nil {
|
||||
return errors.Wrapf(err, "addon phase")
|
||||
}
|
||||
|
||||
if err := k.adjustResourceLimits(); err != nil {
|
||||
glog.Warningf("unable to adjust resource limits: %v", err)
|
||||
}
|
||||
return nil
|
||||
}
|
||||
|
||||
|
|
Loading…
Reference in New Issue