Cases62 cases

37 to 41 of 62

1. 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
2. IL-0140 The GKE API in us-central1 returned errors and high latency for 3 hours 19 minutes after an unexpected rise in backend traffic that client retries made worse

   Google Cloud2022

   Symptoms

   From 03:03 US/Pacific on 14 June 2022, calls to the GKE API in us-central1 through gcloud and the Google Cloud Console were slow and returned errors. Retrying sometimes succeeded.

   High error rateHigh latencyControl plane lag

   Trigger

   An unexpected increase in backend traffic to the GKE API in us-central1.

   Root cause

   Capacity exhaustion

   Component

   GKE control plane

   Contributing factor

   Retry storm

   Remediation

   • scaled up the GKE control plane to handle the increased traffic

   Lessons

   Client retries against an overloaded management API add to the overload. When the cloud provider's control plane is degraded, running workloads may be fine while cluster operations fail, and the status page offered no workaround beyond advising customers to use another region if they could.

   View case
3. IL-0206 Consul agents in GitLab.com's GKE pods were repeatedly OOM-killed, so the Rails pods lost database replica discovery and sent every query to the primary

   GitLab2022

   Symptoms

   On 2022-04-25 the Sidekiq Apdex (the share of background jobs that finish within a target time) dropped, and an engineer declared an incident in Slack at 15:13 UTC. Users saw intermittent slowness across GitLab.com, and asynchronous work such as pipeline creation and import and export took unusually long.

   High latencyDatabase overloadOOMKilledHigh error rate

   Trigger

   The Consul agent in GitLab.com's GKE pods had memory demand spikes that did not fit in the pods' 500MB memory limit, and the agent was killed again and again.

   Root cause

   Capacity exhaustion

   Component

   Container memory limitService discoveryDatabase

   Contributing factor

   –

   Detection

   An engineer declared the incident in Slack at 15:13 UTC after the Sidekiq Apdex dropped; the second event was found when monitoring detected an elevated error rate at 01:21 UTC on 2022-04-26.

   Remediation

   • planned at 16:13 to raise the per-user connection limit for the gitlab database user and the pgbouncer backend pools for the main database; changes were applied at 17:26 and confirmed at 17:47
   • raised the pod memory limits from 500MB to 1GB so the Consul agents survive memory spikes

   Action items

   • Monitor the Consul agents on GKE pods to see whether the 1GB memory limit keeps them healthy
   • Investigate the Consul memory demand spikes and what can be done beyond raising the hard memory limit

   Lessons

   When the local Consul agent that Rails uses for database service discovery keeps dying, Rails does not fail; it sends every query to the last known primary. What engineers saw was database saturation and slow background jobs, which pointed the investigation first at a recent pgbouncer pool change. The memory limit of the Consul agent in every pod therefore decided the load on the primary database.

   View case
4. IL-0461 Every rollout of Datadog's metrics query service on Kubernetes raised errors that the logs blamed on DNS

   Datadog2022

   Symptoms

   From September 2021, each rollout of an update to Datadog's metrics query service, which serves real-time and historical metric data to the web frontend and to monitor evaluation, came with a rise in errors. Clients retried failed queries, so few errors reached users, but the retries slowed dashboards and monitors. The service's logs showed DNS errors stopping it from connecting to the data stores it depends on.

   High error rateDNS resolution failuresHigh latencyOOMKilledPacket lossStale endpoint routing

   Trigger

   Rollouts of the metrics query service, which delete the old pods. The client behaviour behind the failures came from an August 2021 rollout of the metrics query clients that applied round_robin as the new default gRPC load balancing policy.

   Root cause

   Configuration change

   Component

   conntrack tableCiliumLinux kernelCoreDNSClient-side DNS resolver and cacheexternal-dnsHTTP and RPC client library

   Contributing factor

   Change with a delayed effectRetry stormUnknown limitMonitoring blind spot

   Detection

   Errors rose during rollouts of the metrics query service, and the service's logs showed DNS errors when connecting to its data stores.

   Remediation

   • Raised the memory of the node-local-dns pods from 64 MB to 256 MB, which stopped the OOM kills but not the DNS errors
   • Tested larger instance types and confirmed that scaling up the service's infrastructure resolved the errors
   • Removed the custom gRPC reconnect parameters from the clients to restore the defaults

   Action items

   • Contribute a change to Cilium that adds an unreachable route for a deleted pod's IP so clients get an ICMP error

   Lessons

   The errors looked like DNS but were not. ENA metrics and VPC Flow Logs gave the visibility into VPC connection tracking that host metrics did not, and Cilium, Kubernetes and AWS networking interact in ways that can make reverse path filtering drop packets it should allow. DNS propagation time depends on many settings. The causes were changes Datadog had made itself, to work around a dependency bug, to resolve earlier incidents and to override default reconnect parameters, which looked safe at the time and showed their effects much later, when the link was hard to see.

   View case
5. IL-0208 A GitLab chart bump that renamed Kubernetes Service and Deployment ports made the GCP load balancer reconfigure, and GitLab.com slowed down and returned intermittent 50x errors

   GitLab2022

   Symptoms

   At 07:53 UTC on 2022-03-31 the rate of 50x errors on GitLab.com went up, and at 07:54 PagerDuty paged the engineer on call (EOC) about the rising error rate. Users saw slow pages and intermittent 50x error pages. The incident lists the Web, API, Git and Pages services as affected.

   High error rateHigh latencyIntermittent timeouts

   Trigger

   A k8s-workloads/gitlab-com pipeline deployed merge request !1643, a GitLab chart bump that changed the port names of Kubernetes Services and Deployments (related to the Prometheus service and pod monitors), starting with the canary environment gprd-cny at 07:49 UTC.

   Root cause

   Configuration change

   Component

   HelmCloud load balancerHAProxy

   Contributing factor

   Automation amplification

   Detection

   A PagerDuty alert on the rising error rate paged the EOC at 07:54 UTC, one minute after the 50x errors went up. The engineer deploying the change noticed the errors and linked them to it.

   Remediation

   • drained the canary environment and re-enabled it after twenty minutes
   • cancelled the deployment jobs for clusters c and d before they started
   • drained gprd-us-east1-d in HAProxy and re-enabled it five minutes later
   • raised the minimum replica count above 100 (merge request !1650)
   • drained gprd-us-east1-c in HAProxy, deployed the change to it, and re-enabled it
   • reverted the minimum replica increase (merge request !1652)

   Action items

   • Implement graceful draining of a whole regional cluster
   • Create a Chatops command that puts PagerDuty into maintenance mode

   Lessons

   Renaming ports on Kubernetes Services and Deployments behind a GCP load balancer made the load balancer reconfigure, and requests were slow or timed out until the new configuration had fully applied, which took minutes. GitLab notes that alerting worked well. The follow-up work is a way to drain a whole regional cluster gracefully and a Chatops command to put PagerDuty into maintenance mode.

   View case

By root cause10