Skip to main content

Incident: Service disruption 2023-04-06 22:54 - 2023-04-07 08:22

Outage fediverse.foundation services 2023-04-06 - postmortem

Summary

Type
Incident
Summary
Unplanned automatic update of the metallb OKD loadbalancer components
Impact
Full outage of all services on OKD cluster
Duration 9:28h
Status

Resolved

Reporter(s)
@damadmai, @paula
Responder(s)
@b2c, @fossie
Internal reference
/boards/team/kxr1rdhhqtb47muhr3kucn5mxe/bnqecfsesxi8x3pkdm6ahbqf1gy/v48sbw7m8hfn95y9u78fezy3s6o/c5ciqgju3ajnu7q74n3qrnf9dih


Service graphs

2023-04-07T12-39-21.952Z.png

2023-04-07T12-51-36.895Z.png

2023-04-07T12-52-26.828Z.png

Symptoms

All services hosted on OKD unreachable from the internet although containers and middle/backend services up and running. Applications expose "connection reset" behaviour although public VIP reachable. Ingress container up, but shows no incoming connections. Hardware network components functional, routing working as expected.

These symptoms point to a failure of an OKD infrastructure component.

Further investigation of the metallb environment shows multiple restarts of containers:

NAME                                                  READY   STATUS    RESTARTS         AGE
controller-65698489c4-8kk49                           1/1     Running   1                6d12h
metallb-operator-controller-manager-f4f5b9865-2c6nl   1/1     Running   1                6d12h
metallb-operator-webhook-server-bf55c447b-x5nnv       1/1     Running   1                6d12h
speaker-67k7x                                         1/1     Running   7 (6d12h ago)    28d
speaker-6v5lv                                         1/1     Running   6 (6d12h ago)    28d
speaker-rp4k8                                         1/1     Running   7 (6d12h ago)    28d
speaker-stnkv                                         1/1     Running   92 (5m35s ago)   28d

Logs of the speaker pods expose errors to read  configMap information from the namespace causing them to crash and get restarted ad infinitum:

W0407 06:01:21.839577 1 reflector.go:424] pkg/mod/k8s.io/client-go@v0.26.0/tools/cache/reflector.go:169:
  failed to list *v1.ConfigMap: configmaps is forbidden: User "system:serviceaccount:metallb-system:speaker"
  cannot list resource "configmaps" in API group "" in the namespace "metallb-system"

Workaround

Edit the clusterroles.authorization.openshift.io and add the permission to read configmap objects to the service account metallb-system:speaker

Root cause

The root cause has been identified in the deployment of the metallb services, which use containers from the main tag of the repository, causing nightly builds of the development branch of the containers ending up getting deployed uncontrollably to the OKD cluster. Also version differences of the running containers are possible, leading to an undefined metallb cluster state.

* https://quay.io/repository/metallb/speaker?tab=tags 

Resolution

Change the tag of the deployments to a stable release version (v0.13)

Follow-up tasks

  • Update documentation
  • Implement monitoring/alerting solution: /boards/team/kxr1rdhhqtb47muhr3kucn5mxe/bnqecfsesxi8x3pkdm6ahbqf1gy/v48sbw7m8hfn95y9u78fezy3s6o/cc1fh1fw1qi8mmguuje55zr1i3a