Cases14 cases

1 to 5 of 14

1. IL-0022 Three overlapping infrastructure bugs (context-window misrouting, TPU output corruption and an XLA top-k miscompilation) degraded Claude's responses

   Anthropic2025

   Symptoms

   From early August 2025 a number of users reported that Claude's responses had become worse. At first the reports were hard to tell apart from the normal variation in user feedback. By late August they had become frequent and persistent enough for Anthropic to open an investigation.

   Response quality degradationMisrouted requests

   Trigger

   A bug from 5 August that sent some Sonnet 4 requests to servers configured for the upcoming 1M-token context window; a misconfiguration deployed to the Claude API TPU servers on 25 August; token-selection and sampling-code changes deployed on 25 and 26 August that removed a December 2024 workaround and exposed a latent XLA:TPU compiler bug; and a routine load-balancing change on 29 August that increased the misrouted traffic.

   Root cause

   Software defect

   Component

   Model request routingTPUXLA compiler

   Contributing factor

   Overlapping failuresMonitoring blind spotLimited debug accessIgnored early signal

   Detection

   User reports of degraded responses from early August, including reports online. They were hard to tell apart from normal feedback variation, and their rising frequency and persistence by late August led Anthropic to open an investigation; internal evaluations did not detect the degradation.

   Remediation

   • Fixed the routing logic so short- and long-context requests go to the correct server pools (deployed 4 September, rolled out to all platforms by 18 September)
   • Rolled back the TPU server change that corrupted token generation (2 September)
   • Rolled back the top-k sampling change for Haiku 3.5 (4 September), Opus 3 (12 September) and Sonnet 4
   • Switched from approximate to exact top-k and standardized some additional operations on fp32 precision

   Action items

   • Add detection tests for unexpected character outputs to the deployment process
   • Develop more sensitive evaluations that reliably tell working and broken implementations apart
   • Run quality evaluations continuously on true production systems
   • Build infrastructure and tooling to debug community-sourced feedback faster without sacrificing user privacy
   • Fix the approximate top-k compiler bug together with the XLA:TPU team

   Lessons

   Benchmarks, safety evaluations and canary deployments missed degradation that users noticed, partly because the model often recovers from isolated mistakes, and Anthropic relied too heavily on noisy evaluations, with no clear way to connect rising reports to each of its recent changes. Overlapping bugs with different symptoms on different platforms looked like random degradation, and privacy controls on user interactions kept engineers from examining the interactions needed to identify or reproduce the bugs. Anthropic concluded that it needs evaluations running continuously on production and a continuous signal from users, such as the /bug command in Claude Code and the thumbs-down button in the Claude apps.

   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-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
5. IL-0107 Kyverno stuck in a crash loop behind a webhook with failurePolicy fail blocked every change to Kubernetes Service objects for 88 minutes

   DFDS2022

   Symptoms

   At 11:38 UTC on 3 August 2022 an end-user team told the platform team that applying a Service resource to one of the Kubernetes clusters failed.

   Deploys failingContainer crash loop

   Trigger

   A certificate chain in a third-party application expired, through no action of DFDS, and the outdated Kyverno installation got stuck in a crash loop.

   Root cause

   Certificate or credential expiry

   Component

   Policy engineAdmission webhook

   Contributing factor

   Deferred upgradesUnreviewed defaultsMonitoring blind spotIgnored early signal

   Detection

   An end-user team alerted the platform team at 11:38 UTC that applying a Service resource failed; the platform team started looking into it at 11:50 UTC.

   Remediation

   • reproduced the issue in a sandbox cluster and deployed the fix there first
   • updated Kyverno to a newer Helm chart version in the affected production cluster

   Action items

   • Implement monitoring on Kyverno
   • Create internal documentation on the Kyverno setup
   • Investigate the exact certificate issue in the third-party application
   • Consider changing the failurePolicy on ClusterPolicy resources going forward

   Lessons

   Operational issues should be added to the 'Ops Day to Day' board from the start.

   View case

By root cause7