Cases16 cases

1 to 5 of 16

1. IL-0236 After an Istio 1.15 to 1.24 upgrade, half of Wikimedia's ML model servers returned 5XX for about a day

   Wikimedia Foundation2026

   Symptoms

   From 17:30 UTC on 23 February 2026, right after the ml-serve-codfw Kubernetes cluster was upgraded, 5XX responses appeared in the metrics of a subset of Wikimedia's machine learning model servers and alerts started to fire. About half of the running model servers were affected and stayed unreachable for roughly 24 hours.

   High error rate

   Trigger

   The two production Machine Learning Kubernetes clusters were upgraded to version 1.31, which included upgrading Istio from 1.15 to 1.24; ml-serve-codfw was done at 17:21 UTC on 23 February and ml-serve-eqiad between 08:00 and 11:00 UTC on 24 February.

   Root cause

   Breaking change in an upgrade

   Component

   IstioModel inference server

   Contributing factor

   Ignored early signal

   Detection

   Users of the affected model servers first reported the issue in the talk-to-machine-learning Slack channel. Alerts for HTTP 5XX had fired in the Machine Learning IRC channel right after the codfw upgrade, but nobody on the team noticed them.

   Remediation

   • Deployed an Istio hot patch that restored the broken functionality and fixed most of the HTTP 500 errors
   • Deployed the permanent fix to all affected clusters after verifying it on ml-staging-codfw

   Action items

   • Deliver alerts to the relevant Slack channels and watch IRC or alerts.wikimedia.org
   • Add more SLOs and fix the existing ones, whose Istio metrics changed

   Lessons

   The problem could have been caught on the staging cluster first, or by ad-hoc testing within the team after each cluster upgrade, which had been requested when the upgrades were announced. Not everyone involved knew that httpbb tests existed, and nobody on the Machine Learning team saw the alerts that fired right after the codfw upgrade. Recovery was fairly quick once users reported the problem, helped by having the right person online to find the Istio hot patch.

   View case
2. IL-0138 A failed Nvidia driver build left broken kernel packages on the only Kubernetes master node, which hit a kernel panic on reboot after a power outage and took the homelab cluster down for six days

   Barush Mendez2025

   Symptoms

   Every service in the homelab Kubernetes cluster stopped answering: the V-Rising server, Grafana, Jellyfin and Linkding were all unreachable, and the Kubernetes API was down, so the cluster could not be managed at all. The outage was found by hand, about 3 hours after it began, when someone tried to use the services.

   Control plane unresponsiveService outageMonitoring outage

   Trigger

   A power outage forced the Kubernetes master node to reboot.

   Root cause

   Software defect

   Component

   GPU driverLinux kernelKubernetes API server

   Contributing factor

   Ignored early signalChange with a delayed effectSingle point of failureMonitoring blind spot

   Detection

   Manually, when someone tried to access the services; there were no automated alerts.

   Remediation

   • Connected a monitor to the master node and read the boot logs, which showed the kernel panic
   • Booted the previous kernel version, which brought the system back
   • Removed the broken Nvidia packages with sudo apt-get remove --purge '^nvidia-.*' and ran sudo dpkg --configure -a
   • Found remaining broken packages with dpkg -l | grep -v ^ii and removed them by hand
   • Ran sudo apt-get install -f and reinstalled linux-headers-$(uname -r)
   • Rebooted and confirmed the kernel panic was gone

   Action items

   • Make sure GPU driver installations fully succeed, or revert immediately when they fail
   • Move the cluster to a high-availability control plane
   • Expose the Kubernetes API through a load balancer in the HA setup
   • Set up a ticketing system to record incidents and their resolution
   • Write predefined troubleshooting workflows to shorten response time
   • Add infrastructure monitoring such as Prometheus and Grafana
   • Add blackbox HTTP and TCP checks from outside the cluster for critical services

   Lessons

   A driver installation that fails half-way can leave kernel packages broken without any visible effect until the next reboot, which may come weeks later and at the worst time. A failed DKMS build should be fixed or reverted at once. With a single master node and no monitoring from outside the cluster, one unbootable machine meant a total outage that nobody was told about; the fix itself was simple once the boot logs were seen.

   View case
3. IL-0083 Intermittent 5xx errors in a multi-tenant Kubernetes cluster came from internal DNS: the node-local dnsmasq hit its 150 concurrent query limit, did not cache cluster.local answers, and ndots search lookups flooded CoreDNS

   Adevinta2023

   Symptoms

   On a Friday afternoon a tenant of SCHIP, Adevinta's multi-tenant Kubernetes runtime, reported a spike of HTTP 5xx errors in some of its services. The drops in quality of service came back periodically through the following days, each one short, and were reported mostly around midnight.

   High error rateDNS resolution failuresHigh latency

   Root cause

   DNS or network fault

   Component

   CoreDNSClient-side DNS resolver and cacheTelemetry agent

   Contributing factor

   Unreviewed defaultsUnknown limitIgnored early signalUnverified assumption

   Detection

   A tenant reported a spike of HTTP 5xx errors in some of its services; the DNS Latency SLI had already been degraded for some time without being acted on.

   Remediation

   • raised the dnsmasq concurrent query limit from 150 to 1500
   • added a CoreDNS header block that sets the ra and aa flags and clears rd, so dnsmasq caches cluster.local answers
   • configured dnsmasq on each node to answer NXDOMAIN for unresolvable search-domain names instead of forwarding them to CoreDNS
   • changed the fluent-bit log router hostname to a fully qualified name
   • asked the affected tenants to use fully qualified names for internal calls

   Lessons

   The DNS Latency SLI had shown the problem long before anyone looked at it. An SLO only protects you when its SLIs are trusted and a drop is investigated. The investigation took a month partly because several on-call groups worked on it separately, without one place for observations and hypotheses, and sometimes chased hypotheses already discarded. Disprove one hypothesis before starting the next, and collect data the metrics system lacks: DNStop showed the garbage lookups that no dashboard did. With ndots:5, short service names multiply DNS traffic; a node-local cache only helps if it can cache the cluster DNS answers at all.

   View case
4. IL-0231 Deleting EKS managed node groups also deleted Karpenter's instance profile, and every Stytch worker node shut down

   Stytch2023

   Symptoms

   On 23 February 2023 Stytch's EKS worker nodes went Ready:Unknown one after another, Karpenter failed to launch replacement nodes, and with fewer and fewer pods scheduled the Live Stytch API went down for 14 minutes, returning 5XX on every call.

   Node NotReadyScale-up failurePod evictionPods PendingService outage

   Trigger

   At 14:21 UTC Stytch landed a change to remove unused EKS managed node groups, because it had decided to run its data plane applications fully on nodes provisioned by Karpenter. Because Crossplane resources use an orphan deletion policy, engineers deleted the node groups in the AWS console, first in staging and then in production.

   Root cause

   Configuration change

   Component

   KarpenterNode poolConfiguration templating

   Contributing factor

   Unverified assumptionShared credentialChange with a delayed effectIgnored early signal

   Detection

   Low-urgency alerts that Karpenter could not find the instance profile fired in staging at 16:06 UTC and in production at 17:48 UTC but were not treated as an incident. The incident process started at 20:26 UTC when node launch failures spiked the Karpenter error rate, and Kubernetes monitoring alerts in production followed at 20:44 UTC. The support team flagged the outage at 21:14 UTC.

   Remediation

   • Reverted the node group deletion, which recreated the managed node groups
   • Manually scaled up one of the Live API node groups with more and much larger nodes and removed its taints
   • Removed node selectors from the applications one after another so their pods could be scheduled

   Action items

   • Treat Karpenter errors as high-urgency alerts that page the on-call engineer
   • Tune cluster alerting thresholds to send a high-urgency page earlier
   • Add alerts for nodes that stay NotReady or Unknown for long periods
   • Raise the thresholds of other deployment alerts, such as unschedulable pods and replicas not available
   • Give Karpenter its own IAM instance profile, separate from the one used by the node groups
   • Separate other AWS resources that were shared between uses, specifically IAM and security groups
   • Rework the EKS and Karpenter configuration so system instance profiles are unique, static per cluster and generated by Karpenter
   • Introduce a cloud visualization tool for engineers making infrastructure changes
   • Add live system and product metrics to the status page
   • Ask AWS to confirm the undocumented actions and side effects of deleting a node group
   • Correct the shared use of IAM roles and security groups across the infrastructure

   Lessons

   The IAM role for Karpenter's self-managed nodes should never have been the same as the one for the managed node groups; sharing it is what let the node group deletion break Karpenter, and the roles had been left combined for speed and simplicity when Karpenter was first set up. The hours-long delay between the configuration change and its effect made the incident hard to assess, and AWS confirmed that this delay is expected, so Stytch will allow for it in future IAM instance profile changes.

   View case
5. 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

By root cause8