Cases4 cases

1 to 4 of 4

1. IL-0542 Twingate authentication and authorization requests failed at higher rates after traffic moved onto a smaller GKE cluster

   Twingate2025

   Symptoms

   On 21 March 2025, between 21:10 and 21:41 UTC, part of the authentication and authorization requests to Twingate services failed with elevated error rates.

   High error rate

   Trigger

   Twingate saw anomalies in one of its GKE clusters and scaled down the deployments in that cluster to protect the rest of the system, which moved its traffic to the other clusters.

   Root cause

   Cloud provider outage

   Component

   Cloud load balancerGKE control planeHorizontal pod autoscalerCluster autoscaler

   Contributing factor

   Retry stormNo headroomOverlapping failuresHarmful manual action

   Detection

   Twingate observed anomalies in one of its GKE clusters at the start of the incident.

   Remediation

   • Reintroduced capacity in the cluster that had been scaled down, so traffic was rebalanced across regions

   Action items

   • Keep working with the cloud provider to find the cause of the unexpected timeouts and of the cluster update's impact
   • Run deployments in all regions with the same number of replicas in the HPA configuration
   • Raise the maximum node counts of the cluster node autoscalers to leave more room to scale up
   • Configure the HPA to scale up faster

   View case
2. IL-0115 One of OpenAI's three production AKS clusters went offline through a human error, and rebuilding it took 10½ hours because an AKS Kubernetes upgrade broke the ingress and AKS cannot be downgraded

   OpenAI2022

   Symptoms

   At 18:56 PDT on 23 June 2022 (01:56 UTC on 24 June) several models went offline at once: some fine-tuned curie and davinci models, the Codex beta models code-davinci-001 and code-cushman-001, the legacy GPT models curie, babbage and ada, and all Embeddings models.

   Service outageHigh error rateIntermittent timeouts

   Trigger

   A destructive operation on one of the three production clusters, caused by human error that the existing safeguards did not catch, took the whole cluster offline.

   Root cause

   Operator error

   Component

   Ingress controllerModel hub

   Contributing factor

   No safeguard on destructive changeNo headroomDeferred upgradesOverlapping failures

   Remediation

   • re-allocated the models from the lost cluster to the remaining two clusters
   • rebuilt the VPN link with Azure support after losing the connection to AKS during cluster setup
   • built workarounds with Azure support after losing the connection to Azure Private Link
   • tore down the cluster and rebuilt it from scratch after the upgraded Kubernetes version broke the ingress
   • moved traffic back into the rebuilt cluster between 05:30 and 06:30 PDT

   Action items

   • Add checks and locks so that this and similar production-critical infrastructure cannot be modified by accident
   • Add better visibility into the operations being performed and make parts of the system more lexicographically distinct
   • Keep working with Azure on the AKS network connectivity issues so they do not disrupt new cluster operations
   • Update playbooks with the new cluster failure modes and change the setup steps for new clusters and infrastructure
   • Upgrade Kubernetes systematically and set aside time so the clusters do not fall behind again
   • Shard fine-tuned models across multiple clusters and work towards more automatic cross-cluster replication
   • Distribute and replicate Azure Blob Storage so fine-tuned models load faster and more reliably
   • Finish rolling out node-to-node caching to relieve Azure Blob Storage bottlenecks
   • Change how model weights are stored so they are more efficient and faster to load

   Lessons

   Losing one of three clusters exhausted the GPU headroom, and GPUs could not be added on short notice, so full recovery depended on bringing the same cluster back. Rebuilding it exposed a deferred Kubernetes upgrade: the newer version broke the ingress, and because AKS cannot downgrade there was no way back except a full rebuild. Recovering many large models at once is limited by the bandwidth of the storage they load from, and the caches that normally hide this do not help when everything moves to a different cluster.

   View case
3. IL-0193 A GCP persistent disk degradation took down a Cassandra StatefulSet whose OrderedReady policy blocked recovery, and Metrictank could not reconnect to the recreated pods, causing a 23-hour Graphite query outage

   Grafana Labs2019

   Symptoms

   An alert fired because some Metrictank instances could not drain their write queue, which meant they could not write to Cassandra. Google had not yet published an incident, but a team member's Discord server, also hosted on GCP, was down, which pointed at GCP; Google confirmed an SSD persistent disk incident about 90 minutes after it began.

   High error rateNode NotReadyPods PendingService outage

   Trigger

   A Google Cloud incident in which SSD persistent disks performed so poorly that many I/O operations failed for an extended period.

   Root cause

   Cloud provider outage

   Component

   Persistent volumes and stateful setsDatabaseNode root disk

   Contributing factor

   Overlapping failuresUnpatched known bugUnverified assumption

   Detection

   An alert that Metrictank write instances could not flush their write queue to Cassandra. Error rates alone rose too slowly to notice quickly, because most queries were served from memory.

   Remediation

   • Added compute nodes hoping new nodes could attach the disks, which did not work
   • Deleted and recreated the Cassandra StatefulSet with an updated pod management policy, so all pods but one started
   • Raised Kafka topic retention to avoid data loss during the long downtime
   • Temporarily added nodes so Metrictank restarts were not slowed by CPU starvation
   • Restarted all Metrictank instances, which took more than 20 hours, so they connected to the new Cassandra IPs

   Action items

   • Re-instantiate the gocql connection in Metrictank when it is detected dead, so it reconnects even when all Cassandra IPs change
   • Fix the Metrictank write-queue deadlock by updating an internal data structure with atomics instead of a lock

   Lessons

   An OrderedReady StatefulSet cannot recover past a single pod whose volume will not attach; when every pod is down, the ordering that protects normal start-up blocks the whole cluster. A client that learns peer addresses only from live peers needs a way to go back to the service name when all peers are replaced at once. A long-tolerated bug with a config workaround can surface at the worst time; this one turned out to have a simple fix. Writing incoming data to a Kafka write-ahead log first meant all data could still be ingested after recovery, and Kafka on non-SSD disks escaped the disk incident. With both Metrictank bugs fixed, the team expects a similar event to take less than 3 hours to recover from instead of 23.

   View case
4. IL-0003 linkerd lost Kubernetes service discovery after an etcd reconfiguration, and restarting every linkerd instance turned failed payments into a full platform outage

   Monzo2017

   Symptoms

   At 14:10 an engineer deployed a change to a service that Monzo's current account needs to process payments, and every request to that service started to fail. Current account customers saw payment failures; the prepaid card was not affected at first because it does not use that service. The change was rolled back at 14:12, but the errors continued.

   High error rateStale endpoint routingService outage

   Trigger

   At 14:10 an engineer deployed a change to a service needed to process current account payments. Two weeks earlier the etcd cluster had been upgraded and grown from three to nine nodes, and one day earlier another team had scaled a new service down to zero replicas while leaving its Kubernetes service in place.

   Root cause

   Software defect

   Component

   etcdKubernetes API serverLinkerdkubelet

   Contributing factor

   Ineffective rollbackHarmful manual actionOverlapping failuresMonitoring blind spotSingle point of failure

   Detection

   Payment failures began right after the 14:10 deploy. The team rolled back at 14:12 and declared an outage internally at 14:16. At 15:13, when no service could serve requests, practically all automated alerts fired.

   Remediation

   • Rolled back the payments service change (the errors continued)
   • Activated internal backup processes so most customers could still use their card
   • Restarted all linkerd instances (this made the outage worse)
   • Restarted the three Kubernetes API server processes so replacement linkerd instances could start
   • Began deploying a newer linkerd version with the fix for empty services
   • Deleted the Kubernetes service that had no endpoints, which let linkerd load service discovery again

   Action items

   • Fix the bug in Kubernetes that can trigger timeouts after a cluster reconfiguration
   • Roll out a new version of linkerd that fixes the parsing error
   • Create better health checks, dashboards and alerts for the affected components so they give clearer signals about what is wrong and prevent human error
   • Improve procedures to communicate outages internally and externally as clearly and quickly as possible

   Lessons

   A large failure in a distributed system can be very hard to understand, and well-meant human action can compound it, as restarting every linkerd instance did here by exposing a second incompatibility. Because Monzo deploys often and its network changes constantly, the deployment and request-routing systems have to be trustworthy. Monzo committed to publishing a similar post-mortem for any incident of this severity.

   View case

By root cause3