Cases4 cases

1 to 4 of 4

1. 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
2. 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+2022

   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
3. IL-0035 Worker nodes went NotReady after a new ElastAlert rule without memory limits used up node memory

   NU.nl2019

   Symptoms

   The NU.nl development team got reports that several services on one of its Kubernetes clusters behaved erratically: occasional error pages, slow responses and timeouts. The affected services were mostly internal tools and dashboards.

   Node NotReadyHigh error rateHigh latencyIntermittent timeoutsMonitoring outageOOMKilled

   Trigger

   A new ElastAlert rule, deployed shortly before from the command line, that fired if the Cloudfront-* Elasticsearch indexes stopped receiving documents; those indexes take a few million requests per hour, and the rule made ElastAlert's memory grow without bound.

   Root cause

   Configuration change

   Component

   Container memory limitkubeletPrometheus

   Contributing factor

   No resource requests or limitsUnreviewed defaultsNo alert on limitMonitoring blind spotExcessive blast radius

   Detection

   Users reported error pages, slow responses and timeouts from several services; no alert fired.

   Remediation

   • terminated the NotReady nodes one by one so the auto-scaling group replaced them (the new nodes also went NotReady, but Prometheus and Grafana landed on a stable node)
   • found the ElastAlert pods' memory growth on the Grafana pod memory dashboard
   • fixed the ElastAlert problem that caused the memory growth

   Action items

   • Define alerts on resource usage such as CPU, memory and disk space
   • Keep researching an alerting strategy that combines Elasticsearch, Prometheus Alertmanager and New Relic
   • Consider exporting metrics out of the cluster and running Grafana outside it, on EC2 or a separate Kubernetes cluster
   • Either visualise the important metrics in the ELK stack or make Prometheus and Grafana more available, and improve the metrics strategy
   • Set resource limits for ElastAlert through the Helm values
   • Configure namespaces with default memory requests and limits
   • Isolate customer-facing workloads from build, logging and monitoring workloads (a dedicated node group, or later a separate tools cluster)
   • Deploy everything through automation such as Jenkins pipelines, and consider a GitOps approach
   • Deploy through a pipeline that includes a smoke test step
   • Make Kubernetes and cloud infrastructure work part of the team's sprints, picked up by all team members while pairing with more experienced members, and hold workshops that go deep into specific topics

   Lessons

   Fixing the root cause covers only a small part of what an incident can teach; the chain of smaller failures around it is where most of the learning is. A pod without a memory limit can take whole nodes down, so limits need defaults at the namespace level. Monitoring that runs inside the cluster it watches goes blind with it. Changes installed by hand from the command line leave the rest of the team without a record of what changed.

   View case
4. IL-0073 Intermittent Kafka connectivity overloaded Docker daemons, set off a node rescheduling cascade and poisoned the Consul gossip mesh

   Target2019sev1

   Symptoms

   An upgrade of the network subsystem under Target's datacenter OpenStack, meant to cause only a brief disruption, ran into problems and left network connectivity disrupted for several hours. Kafka, which runs on that OpenStack, became intermittently available.

   Node NotReadyLoad imbalanceHigh latencyCredentials unavailable

   Trigger

   A network subsystem upgrade in the datacenter OpenStack infrastructure disrupted connectivity for several hours and made Kafka intermittently available.

   Root cause

   Dependency failure

   Component

   Container runtimeService discoveryTelemetry agentCredential broker for pods

   Contributing factor

   Excessive blast radiusAutomation amplificationSingle point of failureUnpatched known bug

   Remediation

   • enabled gossip encryption on Consul so messages from the poisoned mesh were rejected
   • stabilized the deployment engine and redeployed the development environment with encryption, terminating instances without it
   • upgraded Consul in place to 1.2.3 and then to 1.3.0

   Action items

   • Reduce how much of an outage a single Docker daemon can cause, together with running smaller and more numerous clusters

   Lessons

   Smaller and more numerous clusters limited the damage: the smaller development clusters and production were not hit the same way as the one large cluster, and dev and prod on one cluster would have been far worse. A Docker daemon shared by many workloads is a brittle failure point. Per-workload logging and metrics sidecars were judged the right design; a shared cluster-level agent would almost certainly have made the problem worse and harder to fix.

   View case

By root cause2