Cases46 cases

24 to 28 of 46

1. IL-0116 A CNI configuration change incompatible with a hardware variant that exists only in production cut network connectivity in an OpenAI GPU cluster and took text-davinci-003 fully down

   OpenAI2023

   Symptoms

   At 4:53 pm Pacific on 25 January 2023 one of OpenAI's GPU clusters had a major outage, and the text-davinci-003 model went fully down.

   Service outageThroughput degradation

   Trigger

   A configuration change to the CNI (Container Network Interface) plugins, which provide network connectivity to containers, was deployed to a GPU cluster that contained a hardware variant the staging environment did not have.

   Root cause

   Configuration change

   Component

   CNI plugin

   Contributing factor

   Unrepresentative stagingMonitoring blind spot

   Detection

   Engineers immediately identified the problem as network connectivity, but it took nearly an hour to identify CNI as the cause.

   Remediation

   • put a fix for the CNI configuration change in place, which restored 80% of traffic immediately
   • deployed models back onto GPUs over about three more hours until capacity was fully stable

   Action items

   • Make staging environments contain every type of hardware configuration so they better represent production
   • Monitor and alert on the health of the CNI system directly
   • Fix the network and storage bottlenecks that limit how quickly many replicas of a model can be deployed

   Lessons

   A change that had been deployed to other clusters over 24 hours and deemed safe still broke a cluster with different hardware, so staging needs every hardware configuration that production has.

   View case
2. IL-0513 PyPI backends on Kubernetes were unreachable over HTTPS for 20 minutes when the Ingress ELB lost its HTTPS listener

   Python Software Foundation2022

   Symptoms

   From 14:43 to 15:03 UTC on 24 October 2022, during regular maintenance of the Kubernetes cluster that hosts PyPI, PyPI's backends could not be reached over HTTPS. The CDN could not fetch pages, uploads to uploads.pypi.org were impossible, and other services such as the legacy file redirect service were interrupted.

   Service outage

   Trigger

   Regular maintenance did a rolling restart of every node in the cluster to roll out upgrades and new configuration; as the new Kubernetes API server hosts came up, Kubernetes validated and refreshed the Ingress configurations.

   Root cause

   Configuration change

   Component

   Ingress controllerCloud load balancer

   Contributing factor

   Configuration not in codeChange with a delayed effect

   Remediation

   • Pointed the Ingress configuration at the new ACM TLS certificate, and Kubernetes restored the HTTPS listener on the ELB

   Action items

   • Investigate managing the hostnames on the ACM TLS certificates for PyPI's Ingress through Kubernetes resources instead of the AWS console

   Lessons

   When every resource is managed through the Kubernetes API, drift between the desired state and reality is less likely to build up and then surface during later maintenance.

   View case
3. IL-0177 WebSocket connections to LINE's messaging-hub on Kubernetes dropped all at once whenever another service's change reloaded a shared HAProxy L7 load balancer, and the reconnect wave overloaded a token API and filled nf_conntrack on Nginx

   LINE Plus2022

   Symptoms

   The graph of WebSocket connections to messaging-hub, the messaging platform used by the Japanese food delivery service Demaecan, showed sudden deep drops even outside peak traffic. The drops kept coming back at irregular intervals. Messages still reached clients, because messaging-hub switches to Firebase Cloud Messaging when a WebSocket is gone.

   Connection resetsDatabase overloadIntermittent connection failuresHigh latency

   Trigger

   Another service that shared the same L7 load balancer deployed its own load balancer configuration in the Verda dashboard, which reloaded HAProxy; 30 minutes later, at the hard-stop-after limit, the old HAProxy process exited and closed every WebSocket connection it still held.

   Root cause

   Configuration change

   Component

   HAProxyCloud load balancerconntrack tableIngress controller

   Contributing factor

   Change with a delayed effectExcessive blast radiusRetry stormUnverified assumption

   Detection

   The developer noticed deep, sudden drops in the WebSocket connection graph that were not tied to peak traffic and kept recurring irregularly. The Nginx packet drops were found through network monitoring.

   Remediation

   • separated the Merchant system's token issuance API into its own deployment so reconnect traffic did not delay other Merchant APIs
   • scaled out Nginx with pod anti-affinity and a label selector so each pod ran on a different node, which stopped the nf_conntrack packet drops
   • raised Ingress Nginx proxy-read-timeout and proxy-send-timeout and the load balancer timeout to 3600s, which did not stop the drops
   • moved messaging-hub from the shared L7 load balancer to an L4 load balancer, which ended the WebSocket drops

   Lessons

   A reload of a shared load balancer by a different team can cut long-lived connections, and with hard-stop-after the cut comes 30 minutes later, far from the change that caused it. The author's lesson is not to narrow the scope of the search too early: trusting that the infrastructure was fine kept the investigation inside the team's own systems for days. A reconnect wave from many WebSocket clients can overload the systems that issue connection tokens and fill nf_conntrack on proxy nodes. Scaling Nginx out only helps for conntrack if the pods land on different nodes, and anti-affinity leaves extra replicas Pending once they outnumber the nodes. That messaging-hub problems can still load the connected systems remains an open problem.

   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-0009 A duplicated generated name made Argo CD replace the InfluxDB Cloud production Application and delete the whole production environment

   InfluxData2022

   Symptoms

   Right after a pull request to the infrastructure-as-code repository was merged, every API endpoint of one InfluxDB Cloud production cluster began returning 404 errors. Writes and reads failed alike: no one could collect data, scheduled tasks failed to run and external queries did not work.

   Service outageResources deletedHigh error rate

   Trigger

   A merged pull request whose rendered YAML added a new Argo CD Application and AppProject that an automation error had named app01-cluster01, the name of the existing production Application, instead of app02-cluster01; reviewers did not notice the app01 and app02 difference.

   Root cause

   Configuration change

   Component

   Argo CDPersistent volumes and stateful setsConfiguration templating

   Contributing factor

   Missed in reviewNo runbookNo safeguard on destructive changeMonitoring blind spotAutomation amplification

   Detection

   Monitoring and alerting did not detect the issue right away, so the team's first response was to try to roll the change back rather than plan a recovery. The write-up does not say how the team found out.

   Remediation

   • stopped the revert of the merged pull request
   • restored the etcd metadata cluster
   • recreated the missing Kubernetes objects so the retained PersistentVolumes became available, then recreated the StatefulSets
   • restored the Zookeeper volumes from an hourly Velero backup
   • rebuilt the Kafka StatefulSet one pod at a time with health and readiness checks disabled
   • accepted writes into the Kafka write-ahead log and scaled up the write path with more replicas and higher memory requests and limits
   • recreated the storage pods, which downloaded data from object storage, reindexed it and replayed the write-ahead log
   • re-enabled Argo CD, queries, dashboards and scheduled tasks, restored replica counts and ran final QA checks

   Action items

   • Write runbooks for restoring state when Kubernetes objects such as PersistentVolumeClaims are deleted but the data on the underlying volumes is preserved
   • Set every volume in every environment to retain data when its PersistentVolumeClaim is deleted
   • Improve the process for handling public-facing incidents
   • Store generated YAML as one object per file, so an overwrite of an existing object shows up as such in review
   • Detect duplicate objects when generating YAML and warn before review, treating objects with the same group, namespace and name as the same even when the apiVersion string differs
   • Add the Argo CD Prune=false annotation to all stateful resources and to Namespace objects managed by Argo CD
   • Set FailOnSharedResource=true on Argo CD Applications so Argo CD fails before changing an object another Application manages or managed

   Lessons

   The biggest gap was that monitoring did not notice the deletion quickly, which pushed the team toward a reflex rollback; in a stateful system a revert can create a new, empty instance instead of bringing the old one back. Retained volumes and clients that retry writes are what kept data loss at zero. Rendering many objects into one file let an overwrite of an existing object look like an addition, so generated configuration needs one object per file and duplicate detection that compares objects the way Kubernetes does. Stateful resources need protection from pruning, and that protection must cover their Namespace too, or deleting the namespace removes everything inside it.

   View case

By root cause1