Cases14 cases

9 to 13 of 14

1. IL-0191 A label selector accidentally added to the Loki distributor Service matched only the new pods of a rollout, which took all write traffic, were OOM-killed and stalled the rollout

   Grafana Labs2022

   Symptoms

   During a rollout of the Grafana Loki distributor Deployment in one Grafana Cloud Logs environment, the write path went down for about 25 minutes. The newly started distributor pods were OOM-killed over and over, never passed their health and readiness checks, and the rollout stopped making progress.

   Service outageOOMKilledLoad imbalanceProbe failures

   Trigger

   Step 1 of moving the Loki ring off a separately deployed key-value store onto embedded memberlist added a new label, gossip_ring_member: 'true', to the distributor pods and Service, together with a small config change, and the change required a rollout of the distributors.

   Root cause

   Configuration change

   Component

   Configuration templatingService discoveryContainer memory limit

   Contributing factor

   Unreviewed defaultsIgnored early signalUnrepresentative staging

   Detection

   The first clue was kubectl get pods and the kube_deployment_status_replicas_updated metric showing that only 16 of 140 distributor pods had been updated.

   Remediation

   • scaled the distributor Deployment out horizontally so the load spread over more pods with the new label and the rollout could continue
   • reverted the migration step to end the outage
   • set service_ignored_labels:: [$._config.gossip_member_label] and passed it to k.util.serviceFor so the label is no longer copied into the Service selector

   Action items

   • Update the function that generates a Service from a Deployment or StatefulSet to handle edge cases, including this one

   Lessons

   A Service whose selector is built by copying every label of a workload changes which pods it matches as soon as a new label is added. During a rollout it then matches only the new pods, which receive all the traffic before there are enough of them. A few errors during a rollout in other environments hid the problem, because those environments gave distributors more memory and could absorb the shift. Manifest diffs kept for every pull request made it possible to see that the Service selector had changed.

   View case
2. IL-0054 All ingress traffic lost while draining a GKE node pool because the load balancer only targeted new nodes without ingress-nginx pods

   DevOps Hof2019

   Symptoms

   While the old nodes of a GKE cluster were being drained to move workloads to a newly created node pool, none of the services in the cluster could be reached through ingress-nginx any more.

   Service outageProbe failures

   Trigger

   Cordoning and draining the old node pool removed those nodes from the GCP load balancer backend while the ingress-nginx pods were still running on them.

   Root cause

   Operator error

   Component

   Ingress controllerkube-proxyNode pool

   Contributing factor

   Unreviewed defaultsUnverified assumption

   Detection

   The team noticed during the drain that no service in the cluster was reachable through ingress-nginx.

   Remediation

   • deleted the nginx-ingress-controller pods still running on the old nodes so they were rescheduled on the new, ready nodes

   Lessons

   With externalTrafficPolicy: Local, the load balancer can only reach nodes that run an ingress pod, and cordoned nodes leave the load balancer backend immediately; if the ingress pods stay behind on cordoned nodes, all ingress traffic stops. Before any change that might touch traffic, look twice into the documentation.

   View case
3. IL-0059 Single-master Kubernetes control plane kept dying because the kubelet restarted a healthy etcd on a failing kubeadm liveness probe

   NRE Labs2018

   Symptoms

   The daily NRE Labs statistics showed lesson activity collapsing, and kubectl could not reach the cluster: `The connection to the server 1.2.3.4:6443 was refused - did you specify the right host or port?`

   Control plane unresponsiveProbe failuresService outage

   Trigger

   Unknown. The liveness probe that kubeadm writes into the etcd static pod manifest (an `etcdctl get foo` against 127.0.0.1:2379, failureThreshold 8, timeoutSeconds 15) started failing. The author first suspected GCP maintenance the night before, but could not reproduce the probe failure by hand and does not know why it began.

   Root cause

   unknown

   Component

   etcdkubeletKubernetes API server

   Contributing factor

   Single point of failureUnreviewed defaultsHarmful manual actionIgnored early signalMonitoring blind spot

   Detection

   The author saw lesson activity drop in the daily NRE Labs statistics during his morning routine.

   Remediation

   • removed the liveness probe section from the kubeadm-generated etcd static pod manifest and applied it between etcd restarts

   Action items

   • Add monitoring, especially for the single-node master
   • Run a basic "down for maintenance" page in Google Cloud Functions that points to Twitter for updates, to redirect to during outages

   Lessons

   Because all of the infrastructure is managed as code, the cluster could be rebuilt quickly, but destroying it before finding the cause also destroyed the evidence. Monitoring matters, especially with a single-node master. Read the restart counts in kubectl output, not just the Running status. Remember the fundamentals instead of expecting the issue to be more complicated than it is, and do not let the urgency of an outage push you into hasty decisions. Put up a maintenance page instead of leaving the site failing halfway.

   View case
4. IL-0055 A failing CronJob with the default concurrency policy piled up thousands of restarting pods and took down a cluster

   Target2018

   Symptoms

   A Kubernetes cluster went down after one CronJob kept failing and its pods piled up. Because each failed run kept restarting instead of being replaced by the next one, the pods accumulated into the thousands. The speaker's arithmetic for a job scheduled every minute that fails for three days is 4,320 restarting pods.

   Service outage

   Trigger

   A CronJob kept failing.

   Root cause

   Configuration change

   Component

   Jobs and CronJobsNode pool

   Contributing factor

   Unreviewed defaultsAutomation amplification

   Lessons

   When deploying a CronJob, set its concurrency policy to Replace or Forbid; with the default, failing runs accumulate and restart together. An application should handle a missing dependency with probes and bounded back-off rather than crash in a loop, which also keeps it from hammering a dependency that is already struggling.

   View case
5. IL-0072 High etcd write latency made the API servers return stale data, and the production cluster flipped between two views of its state

   Nordstrom2017

   Symptoms

   A user reported that pods were starting up and then vanishing. When the cluster operators looked, kubectl get pods and kubectl get nodes returned two different sets of data that alternated back and forth, as if there were two opinions of the cluster's state.

   Stale or inconsistent API readsStale endpoint routingService outage

   Trigger

   etcd write latency rose far above the healthy single-digit milliseconds; the speakers say trouble starts once the database sync time crosses double digits or several hundred milliseconds.

   Root cause

   Software defect

   Component

   etcdKubernetes API serverIngress controllerPersistent volumes and stateful sets

   Contributing factor

   Unreviewed defaults

   Detection

   A user reported pods starting up and vanishing.

   Remediation

   • restored service on a replacement cluster
   • freed persistent volumes from the old cluster one by one so they could bind in the new cluster
   • updated DNS records that pointed at the old cluster's load balancer names

   Action items

   • Make the Kubernetes API server request quorum reads from etcd

   Lessons

   etcd is extremely sensitive to write latency; once the database sync time crosses double digits of milliseconds, stale reads or badly degraded performance follow. Without quorum reads, a consistent store can still hand out stale data. Documentation for a fast-moving system changes, so read it again regularly. Moving workloads to a new cluster is hard when they hold persistent volumes or when DNS records point at generated load balancer names.

   View case

By root cause8