Cases32 cases

15 to 19 of 32

1. IL-0364 A quarter of Cloud Platform ingress traffic was slow because an AWS NLB stopped health-checking targets in one zone

   UK Ministry of Justice2022

   Symptoms

   On 11 July 2022 users reported in the #ask-cloud-platform Slack channel that their applications on the Ministry of Justice Cloud Platform were slow some of the time. About 25% of ingress traffic was affected: after an AWS incident the day before, one of the three network interfaces of the 'default' ingress controllers performed slowly, and clients timed out when they tried to connect to one of the Elastic IPs of the Network Load Balancer (NLB) in front of them.

   High latencyIntermittent timeouts

   Trigger

   Some instances of the AWS NLB health checking subsystem were restarted, which set off a latent software race condition in that subsystem.

   Root cause

   Cloud provider outage

   Component

   Cloud load balancerIngress controller

   Contributing factor

   –

   Detection

   Users reported slow performance of their applications in the #ask-cloud-platform Slack channel.

   Remediation

   • AWS restarted the NLB health checking subsystem, which refreshed its list of targets and recovered the NLB in the affected Availability Zone

   Action items

   • Raised mitigation tickets after a post-incident review, filed as post-aws-incident issues in the ministryofjustice/cloud-platform repository

   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-0522 An ingress-nginx chart upgrade deleted the jenkins.io Ingress and Fastly served HTTP/503 errors for hours

   Jenkins project2021

   Symptoms

   At 11:32 UTC on 18 November 2021, a user reported in the jenkinsci/jenkins Gitter channel that https://www.jenkins.io/doc/pipeline/steps/ was answering HTTP/503 errors with the message "Guru meditation" from Fastly's varnish.

   Service outageResources deleted

   Trigger

   The ingress-nginx Helm chart was upgraded from 4.0.6 to 4.0.8 at 10:42 UTC, and the new chart added an admission webhook rule that forbids Nginx configuration blocks in Ingress annotations.

   Root cause

   Breaking change in an upgrade

   Component

   Ingress controllerAdmission webhookHelmCDN and edge network

   Contributing factor

   Unverified assumptionCache-masked failureMonitoring blind spotHarmful manual action

   Detection

   A user reported in the jenkinsci/jenkins Gitter channel at 11:32 UTC that a jenkins.io page answered HTTP/503 errors, and status.jenkins.io reported the incident at 11:49 UTC.

   Remediation

   • Disabled the automated systems, rolled the ingress-nginx chart back to 4.0.6 and deleted the admission webhook by hand
   • Removed the forbidden annotations from the jenkins.io configuration
   • Upgraded the jenkins.io Helm chart to the last version known to work with ingress-nginx 4.0.8
   • Redeployed the jenkins.io Helm chart through the automated system and told Fastly to drop all cached pages

   Action items

   • Always read the changelog in depth when updating the ingress controllers, with two people pairing in real time
   • Add custom Datadog monitoring for www.origin.jenkins.io so that such an outage is detected
   • Make the monitor send the Host header www.jenkins.io so it is not redirected to Fastly
   • Check a set of pages, such as the pipeline steps page, not only the front page
   • Consider an ephemeral k3s end-to-end test before deploying to production

   Lessons

   Trusting a patch version of a Helm chart for the Ingress controller without reading its changelog in depth let a new admission rule remove the jenkins.io Ingress in production. Without monitoring on origin.jenkins.io, a user found the outage. A complete staging cluster with the same DNS setup would have caught the problem before production.

   View case
4. IL-0442 Older pods on LIFULL's KEEL Kubernetes cluster were preempted all at once and the external error rate spiked

   LIFULL2021

   Symptoms

   Shortly after the KEEL team applied new PriorityClasses to KEEL, LIFULL's Kubernetes cluster, and saw pods with the low class being preempted as intended, the team received a flood of alerts from external monitoring: the error rate of requests coming from outside the cluster had risen sharply.

   Pod evictionHigh error rate

   Trigger

   The KEEL team deployed three PriorityClasses to the running cluster: low with value: 100, medium with globalDefault: true and value: 1000, and high with value: 10000.

   Root cause

   Configuration change

   Component

   Pod priority and preemptionIngress controller

   Contributing factor

   Unverified assumption

   Detection

   External monitoring sent the team a large number of alerts about a sharp rise in the error rate of requests from outside the cluster.

   Action items

   • Redefine the PriorityClasses as low (-10), medium (0, globalDefault) and high (10), each with a non-preempting variant

   Lessons

   Adding PriorityClasses to a running cluster needs care because every existing pod has priority 0, and a globalDefault class above 0 makes all pods created before it candidates for preemption. PodDisruptionBudgets do not reliably protect pods from preemption. In a multi-tenant cluster, the permission to deploy PriorityClasses should be restricted, and the PriorityClasses tenants may use should be controlled with ResourceQuota or OPA Gatekeeper.

   View case
5. IL-0106 A malformed Traefik IngressRoute triggered a Traefik v2 bug once a node rollover restarted every Traefik pod, and all Kubernetes services returned 404

   DFDS2021

   Symptoms

   At 08:00 UTC on 2021-09-06 every service in DFDS's Kubernetes cluster answered with 404. Investigation started five minutes later.

   Service outageHigh error rateProbe failures

   Trigger

   The automated Kubernetes node rollover on 2021-09-06 restarted all the Traefik v2 pods, five days after a malformed IngressRoute manifest had been applied.

   Root cause

   Software defect

   Component

   Ingress controllerCloud load balancer

   Contributing factor

   Change with a delayed effectNo configuration testsExcessive blast radius

   Detection

   All services returned 404 from 08:00 UTC and root cause investigation started at 08:05. Time to detection was 5 minutes.

   Remediation

   • changed the ALB health check path, which made the targets healthy but did not stop the 404s
   • rolled back by hand to Traefik v1 as the main ingress controller
   • deleted the malformed IngressRoute by hand and confirmed the Traefik v2 bug
   • applied the changes upstream

   Action items

   • Upgrade Traefik to v2.5, where the bug is fixed, and verify that it removes the bug
   • Change the Traefik v2 health check path to something more robust
   • Set up an external liveness probe on the ingress controller
   • Consider workload segmentation to limit the blast zone
   • Look into monitoring the Traefik v2 API for rapid changes
   • Look into smoke testing during and after node rollover

   Lessons

   A malformed IngressRoute did nothing visible for five days and then took down all ingress when every Traefik v2 pod had restarted during a node rollover. Because Traefik v2 failed gracefully, the first signals pointed at the ALB health checks instead of at Traefik. On incident handling, the team concluded that appointing an incident manager could have helped, that the war room meeting should be recorded and a log book kept, that people not involved should only make suggestions in writing, and that a short first round of sharing thoughts and gut feelings is useful.

   View case

By root cause8