Cases43 cases

6 to 10 of 43

1. IL-0063 VPN outage after recreating a Redis StatefulSet restarted the whole Redis cluster without readiness checks

   AdGuard VPN2025

   Symptoms

   The day began with an alert that several instances of the Redis cluster were running out of disk space. That part was fixed quickly. The outage came an hour later: right after the Redis StatefulSet was deleted and recreated at 8:10 UTC, Kubernetes restarted every Redis pod within about five minutes, and problems started to appear.

   Service outageOOMKilledProbe failuresHigh latency

   Trigger

   To make Git match the volumes that had been resized by hand, the Redis StatefulSet was deleted with orphan (non-cascading) deletion and recreated with the new storage template; Kubernetes reconciled the new spec by restarting the entire Redis cluster.

   Root cause

   Configuration change

   Component

   Persistent volumes and stateful setsIn-memory cacheContainer memory limit

   Contributing factor

   No health probesUnverified assumptionSingle point of failureRetry stormNo headroom

   Detection

   An alert about Redis instances running out of disk space started the work; the outage itself showed up as backend errors about Redis connections right after the StatefulSet was recreated.

   Remediation

   • made the Redis masters forget obsolete node identities, reset the affected replicas and rejoined them, then waited for replication to catch up
   • identified the component that was actually running out of memory, scaled the backend service and raised memory limits for that component
   • temporarily relaxed probes so backend instances took traffic immediately instead of flapping between ready and not ready
   • reverted the temporary workarounds at 10:00 UTC

   Action items

   • Add startup probes so a Redis pod is not marked ready until its dataset is loaded
   • Build an own control plane for Redis that handles restarted pods stuck to an old master address
   • Reduce dependence on the single Redis cluster: move some important data to other stores, such as a separate Redis cluster per server, and use Kafka for some backend components

   Lessons

   The disk resize was not the problem; the cleanup afterwards was. An immutable field pushes operators toward deleting and recreating the controller, and that can set off a full reconcile. Startup and readiness probes are what slow a rolling restart down to a safe pace; without them a distributed store restarts in the most fragile way. Recovery plans must expect a cluster to heal slowly, and dependent services need to degrade gracefully while it does.

   View case
2. IL-0188 Rolling updates on EKS returned 502 errors for about 30 seconds because Kubernetes deleted old pods as soon as new pods passed readiness, before the ALB had marked the new pod targets healthy

   Yonghyun Kim2025

   Symptoms

   Every rolling update of the team's microservices on Amazon EKS, behind an Application Load Balancer, was followed by intermittent 502 and 503 errors for users. A Slack message at 3 a.m. read "배포는 성공했는데, 서비스는 죽었다" (the deployment succeeded, but the service died). Every pod was Ready and Datadog was green while users saw errors.

   High error rateService outage

   Trigger

   A rolling update of a Deployment served through an ALB that registers pod IPs as targets, with no pod readiness gates.

   Root cause

   Faulty deployment

   Component

   Cloud load balancerIngress controllerPod termination

   Contributing factor

   Unverified assumptionUnreviewed defaultsMonitoring blind spot

   Detection

   A Slack message at 3 a.m. reported that the deployment had succeeded but the service was down; Datadog dashboards were green at the time.

   Remediation

   • labelled the namespace with elbv2.k8s.aws/pod-readiness-gate-inject=enabled so the AWS Load Balancer Controller injects pod readiness gates and Kubernetes waits for the ALB target to be healthy before deleting old pods

   Lessons

   A pod being Ready in Kubernetes does not mean the load balancer will send it traffic. Without a readiness gate, a rolling update can delete the old pods while the ALB still considers the new targets unhealthy, and dashboards that track pod status stay green during the errors. Making pods wait longer on shutdown reduced the errors but did not fix the start-up gap and made every deployment much slower; the readiness gate ties the pod's Ready state to the ALB's view of the target.

   View case
3. IL-0171 Karpenter node auto-repair replaced EKS GPU nodes about 10 minutes after they became Ready, in a loop, because the Bottlerocket NVIDIA AMI lacked libdcgm.so and set AcceleratedHardwareReady to False

   Inflab2025

   Symptoms

   GPU nodes for an AI video upscaling job were created normally and reached Ready, and then about 10 to 11 minutes later Karpenter treated each one as unhealthy and replaced it. Every new node went the same way, so the GPU pods kept restarting and the upscaling work could not make progress.

   Node recreation loopGPU unavailable

   Trigger

   Enabling Karpenter's node auto-repair feature gate (nodeRepair: true) and then running GPU pods on nodes from the Bottlerocket NVIDIA AMI.

   Root cause

   Software defect

   Component

   KarpenterNode operating system imageGPU driver

   Contributing factor

   Automation amplificationUnverified assumption

   Detection

   The team noticed GPU nodes being replaced repeatedly after running GPU pods, then found the DCGMError message in the node events.

   Remediation

   • Disabled Karpenter node auto-repair by setting the nodeRepair feature gate to false

   Lessons

   Automatic node repair acts on every unhealthy node condition, including AcceleratedHardwareReady, so an image defect that only breaks GPU health monitoring is enough to have every GPU node terminated and recreated in a loop. Before turning on auto-repair, check which conditions it watches and whether the node image reports them correctly. Disabling auto-repair stops the loop at once but is only a temporary measure; installing the missing library or changing the AMI address the cause but cost more on an immutable OS such as Bottlerocket.

   View case
4. IL-0006 VPC subnets ran out of IP addresses because the AWS VPC CNI kept IPs on full nodes, and WARM_IP_TARGET changes then blocked pod starts after every pod deletion

   Neon2025

   Symptoms

   On 16 May 2025, in Neon's Kubernetes cluster in AWS us-east-1, new compute pods were scheduled but could not start and failed with IP allocation errors. Each Neon compute is a Postgres database running in a lightweight virtual machine in its own pod, so without new pods, idle customer databases could not be woken up.

   IP address exhaustionPods Pending

   Trigger

   On 16 May the Neon control plane's periodic job that terminates idle computes started failing. Compute creations continued, and the number of active computes rose from about 5k to about 8.1k within minutes. On 19 May, a follow-up change that set WARM_IP_TARGET back to 1 started the second round of failures.

   Root cause

   Capacity exhaustion

   Component

   AWS VPC CNIVPC subnet or pod address rangeCluster autoscaler

   Contributing factor

   Unreviewed defaultsNo capacity planningUnverified assumptionHarmful manual actionIneffective rollbackRetry storm

   Detection

   As cluster-autoscaler added nodes, the team saw IP allocation errors on new pods that were scheduled but could not start. The rate of successful pod starts and VPC subnet metrics were then used to follow the problem.

   Remediation

   • Set WARM_IP_TARGET=1 to release IPs held by nodes without free CPU or memory
   • Fixed the control plane's periodic job so idle computes were terminated again
   • Restarted the aws-node DaemonSet several times, each with a temporary improvement
   • Set WARM_IP_TARGET=0, MINIMUM_IP_TARGET=0 and WARM_ENI_TARGET=0, which let nodes allocate enough IPs and stabilised the cluster
   • On 19 May, set WARM_IP_TARGET back to 0 after setting it to 1 failed again
   • Enlarged the pool of pre-created warm compute pods to absorb failed pod starts

   Action items

   • Open a pull request to AWS CNI to improve how ipamd releases IPs in this situation
   • Follow up with AWS Support about the roughly 100 subnet IPs that were never allocated, for reasons Neon could not explain
   • Change the architecture to support horizontal scaling

   Lessons

   With default settings AWS CNI can keep many IPs on nodes that cannot use them, so the share of subnet IPs assigned to pods does not tell how close the subnet is to running out. WARM_IP_TARGET and WARM_ENI_TARGET follow different logic, a warm IP target of zero switches to the ENI logic, and IPs in cooldown count toward the target, which behaves badly under high pod churn. Changing these settings during an incident without knowing that logic prolonged the outage, and reverting a setting did not bring back the earlier healthy state.

   View case
5. IL-0324 Three MediaWiki deployments lost Memcached after a PHP 8.1 rollout and overloaded the External Storage database cluster

   Wikimedia Foundation2025

   Symptoms

   On 2025-03-12 from 10:58 UTC, alerts for high backend response times came in, and the MediaWiki deployments mw-api-int, mw-parsoid and mw-jobrunner, which serve MediaWiki for non-user-facing purposes, showed high latency and database errors. The External Storage cluster, whose hosts store the compressed text of wiki page revisions, had a sharp rise in load and connections. The same happened again on 2025-03-13 from 11:15 UTC.

   Database overloadHigh latencyHigh error rateLog floodMonitoring outage

   Trigger

   At 10:44 UTC on 12 March a Scap deployment of the PHP 8.1 rollout for mw-api-int, mw-parsoid and mw-jobrunner was run with the flag -Dbuild_mw_container_image:False; the rollout was deployed with Scap again at 11:15 UTC on 13 March.

   Root cause

   Faulty deployment

   Component

   DatabaseIn-memory cacheConfigMaps and SecretsHelmCI system

   Contributing factor

   Unverified assumptionDependence on a warm cacheSignal buried in noise

   Detection

   Alerts for high backend response times started at 10:58 UTC on 12 March, 14 minutes after the Scap deployment; 2 on-call engineers were paged. The MediaWikiMemcachedHighErrorRate alert that pointed at the cause fired only briefly because its log pipeline was overloaded.

   Remediation

   • Deployed a patch to cirrus-streaming-updater that reduced SUP parallelism
   • Reduced the concurrency of the categoryMembershipChange job in changeprop-jobqueue
   • On 12 March, reverted 1126607 and 1126650 and re-ran Scap
   • On 13 March, rolling-restarted changeprop-jobqueue
   • Rolling-restarted mw-api-int
   • Disabled the categorymembership job in changeprop-jobqueue
   • Reverted 1127476 and 1127478
   • Redeployed mw-api-int to pick up the revert
   • Increased the number of mw-api-int replicas
   • Redeployed mw-jobrunner to pick up the revert
   • Redeployed mw-parsoid to pick up the revert

   Action items

   • Reduce the number of messages sent through channel:Memcached during failures
   • Add helm rollback functionality to Scap
   • Decide whether Scap should update helmfile-defaults when -Dbuild_mw_container_image:False is set
   • Consider removing envvars.inc from MediaWiki images
   • Consider "Why" dashboards that give an overview of a group of related components together with relevant links to follow
   • Coordinate incident responders better
   • Be able to work with MediaWiki developers during incidents

   Lessons

   The write-up puts the weight not on the technical root cause but on why the team could not see, even after the 12 March incident, that three MediaWiki deployments were running without Memcached. Alert dashboards showed what was happening, such as many busy PHP workers, but not why; dashboards linking components that affect each other, such as SQLBlobStore keys and External Storage, would have helped. The alert that pointed in the right direction did not stay firing and would probably have been lost among 78 others. The incident coordinator could not tell who was working on what or who had joined or left, the channels were noisy, the external impact was judged more critical than it was, and trace.wikimedia.org was not used although it might have helped.

   View case

By root cause10