Cases18 cases

3 to 7 of 18

1. IL-0204 An etcd memory spike overloaded the control plane of a Render Frankfurt cluster, and DNS then failed because CoreDNS ran on the rebooted control plane nodes and could not start without the API server

   Render2024

   Symptoms

   In January 2022 an etcd memory spike overloaded the Kubernetes control plane of one of Render's clusters in Frankfurt. At first only operations that need the control plane failed: new deploys and cron job runs.

   Control plane unresponsiveDeploys failingDNS resolution failuresHigh error rate

   Trigger

   A memory spike in etcd overloaded the cluster's control plane.

   Root cause

   Capacity exhaustion

   Component

   etcdCoreDNSKubernetes API server

   Contributing factor

   Unreviewed defaultsHarmful manual action

   Action items

   • Run CoreDNS on data plane nodes instead of control plane nodes
   • Run etcd on dedicated nodes

   Lessons

   CoreDNS can keep answering from its in-memory cache while the API server is down, but it cannot initialise without it. With CoreDNS on the control plane nodes, a reboot or restart during a control plane outage turned a failure of deploys and cron jobs into a failure of requests to services that were already running. The core problem was that the data plane was not isolated enough from the control plane: even without the reboot, the etcd memory spike alone could have forced CoreDNS to restart.

   View case
2. IL-0005 A new telemetry service overloaded the Kubernetes API servers of large clusters and broke DNS-based service discovery for all OpenAI services

   OpenAI2024

   Symptoms

   Alerts fired at 3:13 PM PST on 11 December 2024, and from 3:16 PM every OpenAI service was significantly degraded or completely unavailable, including ChatGPT, the API and Sora. The status page reported API calls returning errors and difficulties logging in to platform.openai.com and ChatGPT. Customer impact peaked at 3:40 PM.

   Service outageHigh error rateControl plane unresponsiveDNS resolution failures

   Trigger

   At 2:23 PM PST a change introducing a new telemetry service, which collects detailed Kubernetes control plane metrics, was merged and the deployment pipeline started. Between 2:51 PM and 3:20 PM the change was applied to all clusters.

   Root cause

   Configuration change

   Component

   Kubernetes API serverClient-side DNS resolver and cacheTelemetry agent

   Contributing factor

   Unrepresentative stagingNo staged rolloutCache-masked failureMonitoring blind spotLimited debug access

   Detection

   OpenAI's detection tools worked: alerts fired at 3:13 PM PST, a few minutes before customer impact began at 3:16 PM, and the root cause was identified at 3:16 PM.

   Remediation

   • Scaled clusters down to reduce the total Kubernetes API load
   • Blocked network access to the Kubernetes admin APIs so no new expensive requests arrived
   • Scaled up the Kubernetes API servers
   • Removed the telemetry service once enough control was back
   • Moved traffic from impacted clusters to healthy ones
   • Intervened manually in clusters where simultaneous resource downloads saturated resource limits

   Action items

   • Put every infrastructure configuration change through a phased rollout with continuous monitoring of both the service workloads and the clusters, including the Kubernetes control plane
   • Run fault injection tests in which the data plane runs without the control plane for longer, and tests that roll out bad changes on purpose to check detection and rollback
   • Build break-glass mechanisms so engineers can reach the Kubernetes API servers in any circumstances
   • Decouple the Kubernetes data plane from the control plane so that service discovery for critical workloads does not depend on the control plane
   • Speed up recovery with better caching and dynamic rate limiters for resources needed at cluster startup, and regular drills that rapidly replace an entire cluster

   Lessons

   Several systems and processes failed at once and interacted in unexpected ways: tests did not catch the change's effect on the Kubernetes control plane, DNS caching put a delay between the change and the failures, and being locked out of the control plane made remediation very slow. Because service discovery used Kubernetes DNS, the control plane was carrying load for product traffic even though the data plane was meant to be independent of it.

   View case
3. IL-0118 A global change to Kubernetes namespace labels made the networking layer recompute metadata and overwhelmed the control planes of three of OpenAI's largest GPU clusters

   OpenAI2024

   Symptoms

   From 10:15 am PT on 25 November 2024 a large share of both API and ChatGPT traffic failed with timeouts and 503 errors.

   High error rateHigh latencyControl plane unresponsive

   Trigger

   A global change to Kubernetes namespace labels caused the networking layer to recompute metadata. The change had rolled out to the small clusters without problems.

   Root cause

   Configuration change

   Component

   Kubernetes API server

   Contributing factor

   Excessive blast radiusLimited debug access

   Remediation

   • moved workloads out of the impacted clusters by hand
   • with the cloud provider, scaled up the number of control plane nodes while scaling down the number of data plane nodes, which let the control planes recover

   Action items

   • Lock namespace changes so this incident cannot recur
   • Freeze certain types of infrastructure deploys
   • Allowlist identities to prevent a recurrence of this class of incidents
   • Improve the GPU load balancing logic so traffic is re-routed automatically and more effectively when possible
   • Improve change management for fleet-wide infrastructure changes

   View case
4. IL-0154 Line cards on an SFO3 core router rebooted together and halved backbone capacity, and DOKS nodes whose guest networking did not recover stayed NotReady with degraded API servers until all affected nodes in the SFO3 DOKS fleet were recycled

   DigitalOcean2024

   Symptoms

   From 22:25 to 22:50 UTC every DigitalOcean service in the SFO3 region had network connectivity problems, as traffic between the datacenter and the backbone dropped. Traffic through the router was back by 22:30 and SFO3's control plane systems had reconnected by 22:50, and most services recovered then.

   Node NotReadyControl plane unresponsiveIntermittent connection failures

   Trigger

   Most of the line cards on one SFO3 core router rebooted at the same time at 22:21 UTC because of hardware errors on the device.

   Root cause

   DNS or network fault

   Component

   Kubernetes API serverkubelet

   Contributing factor

   –

   Remediation

   • The crashed line cards came back online and traffic on the core router was restored
   • Manually recycled all affected nodes across the SFO3 DOKS fleet

   Action items

   • Work with the router vendor's support team to find the root cause of the line card crash
   • Upgrade software on the core routers in SFO3
   • Raise thresholds for automated remediation so affected nodes are restored without manual action

   Lessons

   A short network fault can leave Kubernetes nodes broken long after the network itself is healthy, when the guest networking on the machines does not recover by itself. Checking node health and the API servers after the network returns, and having automated remediation that is allowed to replace many nodes at once, shortens that tail.

   View case
5. IL-0077 EKS cluster went read-only for over seven hours after Kyverno ephemeralreports filled the etcd database past its size limit

   Marcin Cuber2024sev1

   Symptoms

   A production EKS cluster on Kubernetes 1.30 went into read-only mode: the API server rejected every request that would change an object, such as creating pods or scaling deployments, and deleting objects to free space did not work either.

   Control plane unresponsiveControl plane lag

   Trigger

   Kyverno 1.12 kept creating ephemeralreports objects that piled up in etcd until the database passed EKS's size limit.

   Root cause

   Software defect

   Component

   etcdKubernetes API serverPolicy engine

   Contributing factor

   No alert on limitLimited debug access

   Detection

   The cluster rejected writes; the audit logs carried the "etcdserver: mvcc: database space exceeded" error.

   Remediation

   • AWS raised the etcd storage limit from 8/10GB to 12GB to unlock writes
   • an AWS engineer deleted the leaked Kyverno ephemeralreports objects directly from etcd

   Action items

   • Monitor the etcd database size with apiserver_storage_size_bytes and never let it reach 8GB
   • Upgrade Kyverno 1.12 to 1.12.4, which fixes the ephemeralreports pile-up, or disable reporting for admission events and tune --aggregationWorkers

   Lessons

   On EKS, once etcd hits its size limit the cluster becomes read-only and only AWS can clean it up, so cleanup scripts must run before the limit is reached. Watch the etcd database size and object counts per resource, especially for controllers that create report objects.

   View case

By root cause6