stackdriver-metadata-agent-cluster-level gets OOMK

I updated a GKE cluster from 1.13 to 1.15.9-gke.12. In the process I switched from legacy logging to Stackdriver Kubernetes Engine Monitoring. Now I have the problem that the stackdriver-metadata-agent-cluster-level pod keeps restarting because it gets OOMKilled.

The memory seems to be just fine though.

The logs also look just fine (same as the logs of a newly created cluster):

I0305 08:32:33.436613       1 log_spam.go:42] Command line arguments:
I0305 08:32:33.436726       1 log_spam.go:44]  argv[0]: '/k8s_metadata'
I0305 08:32:33.436753       1 log_spam.go:44]  argv[1]: '-logtostderr'
I0305 08:32:33.436779       1 log_spam.go:44]  argv[2]: '-v=1'
I0305 08:32:33.436818       1 log_spam.go:46] Process id 1
I0305 08:32:33.436859       1 log_spam.go:50] Current working directory /
I0305 08:32:33.436901       1 log_spam.go:52] Built on Jun 27 20:15:21 (1561666521)
 at gcm-agent-dev-releaser@ikle14.prod.google.com:/google/src/files/255462966/depot/branches/gcm_k8s_metadata_release_branch/255450506.1/OVERLAY_READONLY/google3
 as //cloud/monitoring/agents/k8s_metadata:k8s_metadata
 with gc go1.12.5 for linux/amd64
 from changelist 255462966 with baseline 255450506 in a mint client based on //depot/branches/gcm_k8s_metadata_release_branch/255450506.1/google3
Build label: gcm_k8s_metadata_20190627a_RC00
Build tool: Blaze, release blaze-2019.06.17-2 (mainline @253503028)
Build target: //cloud/monitoring/agents/k8s_metadata:k8s_metadata
I0305 08:32:33.437188       1 trace.go:784] Starting tracingd dapper tracing
I0305 08:32:33.437315       1 trace.go:898] Failed loading config; disabling tracing: open /export/hda3/trace_data/trace_config.proto: no such file or directory
W0305 08:32:33.536093       1 client_config.go:549] Neither --kubeconfig nor --master was specified.  Using the inClusterConfig.  This might not work.
I0305 08:32:33.936066       1 main.go:134] Initiating watch for { v1 nodes} resources
I0305 08:32:33.936169       1 main.go:134] Initiating watch for { v1 pods} resources
I0305 08:32:33.936231       1 main.go:134] Initiating watch for {batch v1beta1 cronjobs} resources
I0305 08:32:33.936297       1 main.go:134] Initiating watch for {apps v1 daemonsets} resources
I0305 08:32:33.936361       1 main.go:134] Initiating watch for {extensions v1beta1 daemonsets} resources
I0305 08:32:33.936420       1 main.go:134] Initiating watch for {apps v1 deployments} resources
I0305 08:32:33.936489       1 main.go:134] Initiating watch for {extensions v1beta1 deployments} resources
I0305 08:32:33.936552       1 main.go:134] Initiating watch for { v1 endpoints} resources
I0305 08:32:33.936627       1 main.go:134] Initiating watch for {extensions v1beta1 ingresses} resources
I0305 08:32:33.936698       1 main.go:134] Initiating watch for {batch v1 jobs} resources
I0305 08:32:33.936777       1 main.go:134] Initiating watch for { v1 namespaces} resources
I0305 08:32:33.936841       1 main.go:134] Initiating watch for {apps v1 replicasets} resources
I0305 08:32:33.936897       1 main.go:134] Initiating watch for {extensions v1beta1 replicasets} resources
I0305 08:32:33.936986       1 main.go:134] Initiating watch for { v1 replicationcontrollers} resources
I0305 08:32:33.937067       1 main.go:134] Initiating watch for { v1 services} resources
I0305 08:32:33.937135       1 main.go:134] Initiating watch for {apps v1 statefulsets} resources
I0305 08:32:33.937157       1 main.go:142] All resources are being watched, agent has started successfully
I0305 08:32:33.937168       1 main.go:145] No statusz port provided; not starting a server
I0305 08:32:37.134913       1 binarylog.go:95] Starting disk-based binary logging
I0305 08:32:37.134965       1 binarylog.go:265] rpc: flushed binary log to ""

I already tried to disable the logging and reenable it without success. It keeps restarting all the time (more or less every minute).

Does anybody have the same experience?

回答1:

The issue is being caused because the LIMIT set on the metadata-agent deployment is too low on resources so the POD is being killed (OOM killed) since the POD requires more memory to properly work.

There is a workaround for this issue until it is fixed.

You can overwrite the base resources in the configmap of the metadata-agent with:

kubectl edit cm -n kube-system metadata-agent-config

Setting baseMemory: 50Mi should be enough, if it doesn't work use higher value 100Mi or 200Mi.

So metadata-agent-config configmap should look something like this:

apiVersion: v1
data:
  NannyConfiguration: |-
    apiVersion: nannyconfig/v1alpha1
    kind: NannyConfiguration
    baseMemory: 50Mi
kind: ConfigMap

Note also that You need to restart the deployment, as the config map doesn't get picked up automatically:

kubectl delete deployment -n kube-system stackdriver-metadata-agent-cluster-level

For more details look into addon-resizer Documentation.

回答2:

I was about to open a support ticket with GCP, but they have this notice:

Description We are experiencing issue with Fluentd crashlooping in Google Kubernetes Engine where master version is 1.14 or 1.15, when gVisor is enabled. The fix is targeted for a release aiming to begin on 17 April 2020. We will provide more updates as the date gets closer. We will provide an update by Thursday, 2020-04-09 14:30 US/Pacific with current details. We apologize to all who are affected by the disruption.

Start time April 2, 2020 at 10:58:24 AM GMT-7

End time Steps to reproduce Fluentd crashloops in GKE clusters could lead to missing logs.

Workaround Upgrade Google Kubernetes Engine cluster masters to version 1.16+.

Affected products Other