Cases2 cases

1 to 2 of 2

1. IL-0285 Clever Cloud's Pulsar brokers stalled in a minor upgrade, then hypervisors overloaded and crashed with kernel panics

   Clever Cloud2024

   Symptoms

   On 2 August 2024, two minutes after Clever Cloud's data team started a minor rolling upgrade of its Apache Pulsar messaging cluster to release 3.3.1, monitoring showed connection problems with the messaging platform. Producing data into the event bus failed, the backlog grew while no new producers or consumers succeeded, and agents on hypervisors, log forwarders on load balancers and deployments were affected. At 10:20 CEST the Pulsar cluster was declared defective.

   Deploys failingSustained high node CPUMemory pressureNode kernel panicOOMKilledAutoscaler misbehaviourMonitoring outage

   Trigger

   At 10:06 CEST the data team began a minor rolling upgrade of Apache Pulsar to release 3.3.1 that also switched on a new balancer for bundle placement: the extensible load manager (loadManagerClassName=org.apache.pulsar.broker.loadbalance.extensions.ExtensibleLoadManagerImpl) with the TransferShedder load shedding strategy instead of the native OverloadShedder.

   Root cause

   Configuration change

   Component

   Message brokerTelemetry agentLinux kernelAutomated capacity manager

   Contributing factor

   No configuration testsMonitoring blind spotIneffective rollbackUnreviewed defaultsAutomation amplificationRetry stormCircular dependency

   Detection

   Monitoring showed connection issues to the messaging platform at 10:08 CEST, two minutes after the upgrade began, and engineers confirmed at 10:09 that producing data into the event bus failed. Monitoring did not catch the first broker's failure to create the system topic, because that broker stayed up.

   Remediation

   • Rolled back the new unload strategy while keeping Pulsar 3.3.1
   • Shut down the client proxies to release the brokers
   • Switched deployment notification events to RabbitMQ as a workaround
   • Stopped all brokers and started only one, which loaded the conflicting system topic so it could be deleted, then started all brokers
   • Restarted the messaging proxies, after which messages flowed again
   • Stopped the watchdog so it would not reload every failed component, and started components one by one
   • Disabled Metrics API ingestion to protect its storage cluster
   • Stopped the long-term storage for observability metrics (the Warp10/FoundationDB cluster) to free resources
   • Put all monitoring instances in all regions into maintenance so they stopped sending events to the orchestrator
   • Killed the overloaded small VMs one by one
   • Flagged 12 hypervisors to drain workload and shut down some load balancers to free CPU and memory
   • Cleaned zombie VMs to speed up recovery
   • Cleared the monitoring alert queue so queued events would not flood the orchestrator
   • Added parked hypervisors to EU-FR-1 to relieve the overloaded ones
   • Removed an overloaded RabbitMQ node from the cluster used by the orchestrator
   • Rolling-restarted every regional L7 load balancer and every L4 load balancer
   • Restarted all HV agents and the add-on provider APIs
   • Destroyed all dangling VMs stuck in the stopping state
   • Re-enabled Metrics API ingestion with a ramp-up QoS policy

   Action items

   • Track KPIs for Pulsar broker and cluster health, hypervisor load, I/O and kernel panic reboots, zombie VMs, RabbitMQ broker and cluster health and deployment queue consumption, each by time of event, detection, decision and resolution
   • Deal with too many monitoring alerts
   • Look at a backup bastion, because a regional bastion is vulnerable
   • Rework the control plane recovery process
   • Rework the correlation between monitoring and orchestration
   • Revisit fault domain separation against multi-tenancy
   • Rethink runtime images and the role of the hypervisor
   • Isolate waved release upgrades of Pulsar
   • Redesign runtime images with better hypervisor integration
   • Regionalise observability for redundancy
   • Set Vector's buffer.max_events and buffer.max_size for small VMs
   • Switch Vector's buffer.type from memory, the default, to disk
   • Change Vector's buffer.when_full from block, the default, to drop_newest
   • Tune journald verbosity and SystemMaxUse on small VMs
   • Add cgroups restrictions for better process control
   • Add I/O restrictions for the VM manager
   • Detect zombie VMs better

   Lessons

   The simulation process that validates non-minor releases was not applied to minor upgrades or to small configuration changes, and this one went straight to production. Running Clever Cloud on Clever Cloud turned into a chicken-and-egg situation: recovery needed the scheduler to carry out large actions while the control plane was itself slowed and the infrastructure had no free resources. Measuring each key indicator by time of event, detection, decision and resolution showed where the team had lacked visibility or reaction and challenged established rules.

   View case
2. IL-0407 Azure OpenAI returned 5xx errors in 14 of 28 regions after an automated cleanup deleted its GPU VMs and endpoints

   Microsoft Azure2024

   Symptoms

   From 00:00 UTC on 13 July 2024 a subset of Azure OpenAI (AOAI) customers got errors, such as 5xx responses, when calling their Azure OpenAI endpoints in 14 of the 28 regions that offer the service. Service monitoring saw failure rates cross their thresholds at 00:05 and alerted engineers.

   High error rateResources deleted

   Trigger

   A change to the global configuration file of the Azure OpenAI resource automation removed 14 resource groups that were thought to be unused, and the automation then deleted every resource in them.

   Root cause

   Configuration change

   Component

   Model inference serverAutomated capacity manager

   Contributing factor

   No staged rolloutNo safeguard on destructive changeUnverified assumption

   Detection

   Service monitoring detected failure rates above the defined thresholds at 00:05 UTC and alerted engineers, five minutes after impact began.

   Remediation

   • Stopped the automation that was running the cleanup
   • Added resource locks and blocked delete requests at the ARM and Automation levels
   • Recreated the deleted resources, rebuilding the service in parallel across the 14 regions
   • Manually reconfigured Azure Redis and Private Link for GPT-4o with the Azure Redis and Azure Networking teams
   • Manually restored DALL-E, adding the managed identity between its frontend and backend to an allowlist
   • Turned the automation back on on 15 July, which restored fine-tuning deployments

   Action items

   • Make the configuration policy regional and follow Safe Deployment Practices for all later configuration updates
   • Remove incorrect metadata tags on resources the automation must not manage and exclude them in the automation configuration
   • Concentrate the regional makeup of workloads in fewer regions
   • Improve the automation that engages incident response personnel
   • Investigate additional change management controls so that all production changes go through a gated safe deployment process (estimated completion August 2024)
   • Update the communications tooling with Azure OpenAI data so affected customers can be scoped and notified faster
   • Investigate additional test coverage and verification procedures to de-risk future configuration updates (estimated completion September 2024)
   • Add active monitoring for each model family that does not depend on customer traffic
   • Build out model pools with more parallelism to shorten recovery

   Lessons

   Microsoft names the missing safe deployment step as the main gap in its change management: a configuration change delivered as a global content update could not be rolled out region by region, so a single change caused a multi-region incident. Resource groups that mixed resources the automation managed with resources it did not know had let the automation's configuration drift from what actually existed in the production subscription.

   View case

By root cause1