Cases5 cases

1 to 5 of 5

1. 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
2. IL-0110 A Grafana Agent pod's high memory use and crash-looping Hangfire pods drove a Kubernetes node into system OOM and NotReady, twice in one day

   DFDS2024

   Symptoms

   At 10:01 UTC on 2024-03-13 a Kubernetes node entered NotReady state, with CPU requests at 78% and memory requests at 82%. An alert fired in Slack at 10:03 and was picked up at 10:05. An attempt to SSH into the node at 10:13 failed because it was unresponsive; it started responding again at 10:19.

   Node NotReadyContainer crash loopOOMKilledMonitoring outage

   Trigger

   Pods of a dev container that used Hangfire to trigger jobs failed with OOM and crash-looped on a node where a Grafana Agent pod was already using a lot of memory.

   Root cause

   Capacity exhaustion

   Component

   Telemetry agentContainer memory limitApplication database

   Contributing factor

   Excessive blast radiusNo resource requests or limitsHarmful manual actionMonitoring blind spot

   Detection

   A Slack alert fired two minutes after the node went NotReady and was picked up two minutes later; the second node's NotReady at 15:03 also raised a Slack alert.

   Remediation

   • marked the EC2 instance unhealthy so a new node started
   • cordoned and force-drained the NotReady node and terminated its EC2 instance
   • removed the new LimitRange from the Grafana Agent namespace and restarted the Grafana Agent pod
   • scaled the dev container to 0 replicas, after which EC2 network traffic returned to normal
   • after the second NotReady, scaled the dev container to 0 replicas again and added a new node
   • emailed the team that owns the dev container, which confirmed it was working on a fix

   Action items

   • Move grafana-agent to the monitoring node group
   • Add alerts for when bandwidth across the cluster reaches a certain threshold
   • Investigate whether memory limits can be set on the grafana-agent Helm chart
   • Investigate grafana-agent memory usage
   • Ask Grafana support whether grafana-agent should use a PVC
   • Consider pushing node logs to Loki with limited retention
   • Create a wiki page with common metrics and log queries to check when investigating issues with Kubernetes nodes

   Lessons

   Dev workloads mixed with production let one crash-looping dev container take down nodes that production also used, and a monitoring agent outside its dedicated node group added to the memory pressure. Scaling the suspect container back up before its owners had fixed it brought the failure back on the replacement node. The team also noted that action items from earlier post-mortems need follow-up, and that knowledge of useful CloudWatch metrics had not been shared, which made the diagnosis harder.

   View case
3. IL-0043 Application logs missing in Kibana after a newer Fluent Bit version changed its chunk drop strategy

   UK Ministry of Justice2023

   Symptoms

   At 09:14 on 4 August 2023, users of the Cloud Platform reported in the #ask-cloud-platform Slack channel that Kibana was missing their application logs for long periods.

   Monitoring outage

   Trigger

   The newer Fluent Bit version in use had changed its chunk drop strategy; the fix that followed handled memory buffer overflow.

   Root cause

   Capacity exhaustion

   Component

   Telemetry agent

   Contributing factor

   No alert on limitMonitoring blind spot

   Detection

   Users reported long periods of missing logs in Kibana in the #ask-cloud-platform Slack channel at 09:14; the incident was declared at 10:09.

   Remediation

   • configured Fluent Bit to handle memory buffer overflow by writing buffered logs to the file system
   • made Fluent Bit flush logs in smaller chunks

   Action items

   • Push notifications from the logging clusters to #lower-priority-alerts (#4704)
   • Add an integration test that checks logs are being sent to the logging cluster

   View case
4. IL-0041 fluent-bit silently stopped shipping application logs from one node to Elasticsearch

   PrometheusKube2020

   Symptoms

   On a Tuesday morning a developer reported that some production logs were missing. Only the logs from a single node were gone; the other nodes kept shipping normally through fluent-bit, the log shipper the team had run for more than three years.

   Monitoring outage

   Root cause

   Software defect

   Component

   Telemetry agentElasticsearch or OpenSearch

   Contributing factor

   Monitoring blind spotUnpatched known bug

   Detection

   A developer reported missing production logs; no alert fired because fluent-bit's metrics showed no errors.

   Remediation

   • added CPU to the Elasticsearch virtual machines, which cleared the write queue rejections
   • rolled out Fluentd one node at a time, using node affinity to place Fluentd and anti-affinity to remove fluent-bit
   • lowered Fluentd's flush_thread_count to 1 and set flush_interval to 5 seconds so the backlog of old logs did not overflow the buffer
   • enabled Fluentd's Position DB so already-ingested data was skipped

   Action items

   • Build monitoring for Fluentd on its Prometheus integration, including a throughput alert with separate day and night thresholds or a black-box log check

   Lessons

   A log shipper can fail without an error message and with metrics that simply stop moving, so error-based alerts never fire; throughput alerts or end-to-end checks can catch it. Open source software can break in ways nobody expects, and sometimes the way out is to switch to a different shipper fast, node by node. A new shipper that starts from scratch replays every old log on the node and needs to be slowed down.

   View case
5. IL-0075 Kubernetes API servers pegged on CPU when the cluster grew to about 290 nodes, traced to Datadog agents polling the API

   Civis Analytics2019

   Symptoms

   After Civis released its new script infrastructure on Kubernetes internally in September 2018, developers reported that kubectl commands were slow, and users intermittently could not start notebooks and services.

   Control plane lagControl plane unresponsiveHigh latencyHigh error rateDNS resolution failures

   Trigger

   Migrating clients' Python, R and container scripts from the in-house Bocce system to Kubernetes (the project called Croquet) pushed the cluster to about 290 EC2 instances under a large script load.

   Root cause

   Capacity exhaustion

   Component

   Kubernetes API serverTelemetry agentetcd

   Contributing factor

   Unknown limitUnverified assumptionMonitoring blind spotNo capacity planning

   Detection

   Developer reports of slow kubectl commands and user reports of notebooks and services failing to start, then high CPU on the master nodes.

   Remediation

   • moved internal users back to the old Bocce infrastructure
   • resized the three master nodes from m4.large to m4.4xlarge
   • changed the Datadog setup so per-node agents no longer polled the API server for service metadata
   • raised max-requests-inflight for non-mutating requests from 400 to 1200
   • grew the etcd EBS volumes from 20 to 100 GB (100 to 300 IOPS)
   • switched client scripts from the ClusterFirst DNS policy to Default

   Action items

   • Upgrade etcd from 2.2.1 to 3.x once Kops supports a safe migration
   • Solve the in-cluster DNS timeouts so client scripts can talk to other workloads in the cluster

   Lessons

   Kubernetes can reach thousands of nodes only after tuning, and third-party DaemonSets are a common cause of control plane load: an agent on every node that polls or watches the API server multiplies with the cluster size. Datadog, kube2iam and the fluentd metadata filter all showed this. A small configuration change often fixes it, but finding it took load testing and the API server access logs.

   View case

By root cause3