Cases19 cases

14 to 18 of 19

1. IL-0222 Google Cloud US regions had heavy packet loss and some went unreachable after the network control plane was descheduled

   Google Cloud2019

   Symptoms

   On Sunday 2 June 2019, with user impact beginning between 11:47 and 11:49 US/Pacific, Google Cloud projects in several US regions saw elevated packet loss from network congestion, and some regions became unreachable. Customers saw higher latency, intermittent errors and lost connectivity to instances in us-central1, us-east1, us-east4, us-west2, northamerica-northeast1 and southamerica-east1, while us-west1, Europe and Asia had no regional congestion.

   Packet lossNetwork unreachableHigh latencyHigh error rateMonitoring outage

   Trigger

   At 11:45 US/Pacific a relatively rare type of datacenter maintenance event started in one physical location, and the automation that runs it descheduled logical clusters that also held network control plane jobs in other physical locations.

   Root cause

   Software defect

   Component

   Network control plane (SDN)BGP routingBackbone network (WAN)

   Contributing factor

   Excessive blast radiusCircular dependencyLimited debug accessIneffective rollback

   Detection

   Google engineers were alerted to the failure two minutes after it began and started the incident process used for the most serious production incidents. The first customer notice appeared on the status page at 12:53 US/Pacific.

   Remediation

   • Lowered the priority of less critical network traffic classes to reduce congestion
   • Halted the datacenter automation software responsible for the maintenance event
   • Rescheduled the network control plane and its supporting infrastructure in each location
   • Rebuilt and redistributed the lost network control plane configuration, rolling it out from 14:03
   • Directed service traffic away from the affected regions
   • Moved the Cloud VPN control plane in three regions out of its fail-safe state

   Action items

   • Keep the maintenance automation off until safeguards stop it descheduling jobs in several physical locations at once
   • Harden the cluster management software to reject such requests whatever their origin
   • Reconfigure the network control plane to reject maintenance requests of the type involved
   • Make the network control plane in each location persist its configuration
   • Let the network stay in fail static mode longer when the control plane is lost
   • Review, update and test emergency response tooling and procedures, including customer communication tooling, for such network failures
   • Extend continuous disaster recovery testing to this and similarly catastrophic failures

   View case
2. IL-0492 Hosted Graphite's self-hosted load balancers hit their connection limit in an AWS outage, timing out its site and APIs

   Hosted Graphite2018

   Symptoms

   On a Friday night at 23:28 UTC, Hosted Graphite's monitoring showed its website intermittently becoming unavailable: periodic checks of the website and of the graph rendering API were timing out. Traffic dropped across all ingestion protocols at the same time.

   Intermittent timeoutsIntermittent connection failures

   Trigger

   An AWS connectivity problem in us-east1 and us-west2 slowed down connections from AWS hosts in those regions to Hosted Graphite's load balancers without breaking them.

   Root cause

   Capacity exhaustion

   Component

   On-premises load balancer

   Contributing factor

   Circular dependency

   Detection

   Monitoring reported the website intermittently unavailable at 23:28 UTC, when periodic checks of the website and the graph rendering API started timing out.

   Action items

   • Enforce stricter client timeouts
   • Raise the load balancers' connection limits
   • Use the new visualizations built from load balancer logs when troubleshooting the load balancing layer

   Lessons

   An outage at a provider that did not host the service's website still made the service unavailable, here because slow clients in the affected regions used up the connection limit of the load balancers that served everyone.

   View case
3. IL-0534 Amazon S3 in US-EAST-1 could not serve GET, LIST, PUT or DELETE requests for hours while its index subsystem restarted

   Amazon Web Services2017

   Symptoms

   From 9:37AM PST on February 28th, 2017, Amazon S3 in the Northern Virginia (US-EAST-1) Region could not serve requests. AWS services in the region that store data in S3 failed with it, including the S3 console, new Amazon EC2 instance launches, Amazon EBS volumes that needed data from an S3 snapshot, and AWS Lambda.

   Service outage

   Trigger

   While debugging an S3 billing system that was progressing more slowly than expected, an authorized S3 team member followed an established playbook to remove a small number of servers from one S3 subsystem used by the billing process, and entered one input of the command incorrectly.

   Root cause

   Operator error

   Component

   Object storageCloud block storageServerless function platform

   Contributing factor

   No safeguard on destructive changeExcessive blast radiusCircular dependency

   Remediation

   • Restarted the index subsystem, which served GET, LIST and DELETE again from 12:26PM PST
   • Restarted the placement subsystem after the index subsystem worked, restoring PUT by 1:54PM PST

   Action items

   • Change the capacity removal tool to remove capacity more slowly
   • Add safeguards that stop capacity removal when it would take a subsystem below its minimum required capacity
   • Audit other operational tools for similar safety checks
   • Improve the recovery time of key S3 subsystems and start the planned further partitioning of the index subsystem into cells now
   • Run the SHD administration console across multiple AWS regions

   View case
4. IL-0333 A downsized RDS PostgreSQL database maxed out its CPU at US peak and failing ELB health checks took buildkite.com down

   Buildkite2016

   Symptoms

   On 22 August 2016 at 17:21 UTC buildkite.com went offline: nobody could log in, view build logs or read the documentation, although builds kept running and GitHub and Bitbucket pull request statuses kept updating. The Buildkite team in Melbourne was asleep and woke at 21:00 UTC, almost 4 hours later, to emails, tweets and Slack messages saying Buildkite was down.

   Service outageDatabase overloadProbe failuresCloud API rate limit exhausted

   Trigger

   Over a weekend maintenance window, Buildkite downgraded its m4.10xlarge Multi-AZ RDS PostgreSQL database to an r3.2xlarge Multi-AZ instance to cut its AWS bill, and moved to a centralised pgbouncer to pool connections. The first US weekday peak after the change pushed the new instance's CPU to its maximum at about 16:00 UTC.

   Root cause

   Capacity exhaustion

   Component

   DatabaseCloud load balancerLiveness and readiness probesNode launch template

   Contributing factor

   Unverified assumptionHealth check tied to a dependencyAutomation amplificationCircular dependencyPages not deliveredOverlapping failures

   Detection

   New Relic, Pingdom, Datadog, Calibre and CloudWatch raised PagerDuty alerts, but the on-call phone call never came and only silent push notifications arrived. The team noticed at 21:00 UTC when they woke to emails, tweets and Slack messages from users.

   Remediation

   • Moved servers from the Agent API to the dashboard ELB to bring buildkite.com back online
   • Switched the Auto Scaling Groups to EC2 health checks and locked in desired counts
   • Upgraded the database to an instance type that could handle the load, using the aws-cli
   • Built a new AMI that did not depend on buildkite.com/_secret/version and rolled out servers that could self-deploy
   • Removed the centralised pgbouncer and went back to pgbouncer on each host

   Action items

   • Plan load testing after significant infrastructure changes without affecting other users of the system
   • Make the ELB health check return OK when the app answers, without checking database connectivity
   • Consider a separate health check endpoint for Pingdom that checks the database and other dependencies
   • Store the latest deployed revision as a file on S3 for newly launched servers to self-deploy from
   • Test by hand that PagerDuty calls reach every on-call team member
   • Investigate regular company-wide tests of everyone's paging setup
   • Ask PagerDuty about calling members when they go on call or requiring an immediate phone rule

   Lessons

   Buildkite listed six lessons. Watch AWS credits, because billing reports show $0 while credits are spent, so the real cost is hidden and big changes end up rushed. Load test after significant infrastructure changes. Think about what a health check should do and what happens when every server behind an ELB fails it. Do not make new servers look up the deployed revision from the service they are replacing. Make sure every on-call member has the right PagerDuty and phone settings before going on call. When servers churn in AWS, stop health checks and autoscaling, for example with the Auto Scaling feature to suspend and resume processes, so the EC2 request limit is not used up.

   View case
5. IL-0273 A power disruption took GitHub's Redis clusters offline, the app could not boot, and GitHub served 503s for 2 hours

   GitHub2016

   Symptoms

   At 00:23 UTC on 28 January 2016 GitHub became unavailable and returned HTTP 503 responses with the unicorn error page. On-call engineers received alerts for unreachable servers and a spike in exceptions related to Redis connectivity, after slightly over 25% of the servers and several network devices in the primary data center rebooted.

   Service outage

   Trigger

   At 00:23 UTC a brief disruption in the systems that supply power to the servers and equipment of GitHub's primary data center rebooted slightly over 25% of the servers and several network devices.

   Root cause

   Dependency failure

   Component

   In-memory cacheData center power and cooling

   Contributing factor

   Unpatched known bugSingle point of failureGap in automated testsCircular dependency

   Detection

   The reboots generated alerts to several on-call engineers, starting with notifications for unreachable servers and a spike in exceptions related to Redis connectivity.

   Remediation

   • Worked with on-site facilities technicians to drain flea power from the failed servers and start them cold so their disks were visible
   • Rebuilt the affected Redis clusters on alternate hardware
   • Restored the Redis cluster data onto standby equipment and brought the Redis-server processes back online
   • Removed the maintenance page after the HAProxy load balancers returned the recovered application servers to the pool

   Action items

   • Update the server firmware with the known issue across the fleet
   • Make tooling open issues automatically when new firmware updates are available, so their changelogs are reviewed
   • Add tests that make sure application processes start even when certain external systems are unavailable
   • Improve circuit breakers so functionality degrades gracefully when backend services are down
   • Review the availability requirements of internal systems such as server provisioning so they are on par with user-facing systems
   • Strengthen communication between teams during incidents
   • Predefine escalation strategies for all-hands incidents
   • Improve status messaging to users during incidents

   Lessons

   Dependencies between components fail in new ways unless they are tested, and a power loss cannot be fully prevented, so the goal is fast and reliable recovery. The application should start without non-essential backends, and internal systems needed to recover from an outage, such as server provisioning, must be as reliable as the system being recovered.

   View case

By root cause8