Cases6 cases

3 to 6 of 6

1. IL-0401 Prow jobs on the k8s-prow-build GKE cluster failed and merges stalled as nodes hit IOPS throttling and kernel panics

   Kubernetes project2019

   Symptoms

   On 30 May 2019, in the run-up to the Kubernetes 1.15 code freeze, Prow test jobs in the k8s-prow-build cluster failed in surprising ways or were aborted partway through. Jobs were reported to GitHub as failed while Deck still showed them as running, because pods exited without uploading `finished.json`, the file most Prow tooling uses to tell a job's state.

   Node NotReadyNode kernel panicPods stuck in TerminatingThroughput degradation

   Trigger

   A spike in test load before the Kubernetes 1.15 code freeze pushed disk I/O on the build nodes far beyond the IOPS that GCP allows for 250 GB disks.

   Root cause

   Capacity exhaustion

   Component

   Cloud block storageNode root diskkubeletCI system

   Contributing factor

   Monitoring blind spotIgnored early signal

   Detection

   Several users reported jobs that GitHub showed as failed but Deck showed as still running, and the first report came on Slack at 13:26. Alerting for these failures did not exist or did not work.

   Remediation

   • Requested and received additional disk quota
   • Brought up 55 nodes with 500 GB disks in a new node pool
   • Migrated all nodes from the 250 GB disk pool to the 500 GB disk pool by draining old nodes, finishing at 23:55

   Action items

   • Create or fix monitoring so the oncall is paged when tide stops doing things
   • Set a process and guidelines for adding new members to the oncall rotation, with preparation and extra support

   Lessons

   Prow metrics were not recorded before the incident, which made it hard to understand, and alerting for these failures did not exist or did not work. After the cause was found and a fix devised, nothing happened for another two hours. The team counts it as luck that someone who recognised IOPS as the problem happened to be around to look.

   View case
2. IL-0518 gitlab.freedesktop.org on GKE ran short of volume space and a planned 30-minute storage upgrade became a 4-hour outage

   freedesktop.org2018

   Symptoms

   Shortly after 9am UTC on 17 July 2018, the disk space of the GitLab cluster behind gitlab.freedesktop.org was critical, and some API requests were failing because of the shortage. The administrator first learned of the failure when people mentioned it on IRC.

   Persistent volume fullHigh error rateService outage

   Trigger

   Many recently created forks of Mesa, each about 2.1GB, filled the disk space of the GitLab cluster.

   Root cause

   Capacity exhaustion

   Component

   Persistent volumes and stateful setsCloud block storageHelm

   Contributing factor

   Unverified assumptionUnrepresentative stagingMonitoring blind spotChange during peak hours

   Detection

   People mentioned the failure on IRC. Monitoring and alerting on disk usage and API error rate had not been set up yet.

   Remediation

   • Dropped all inbound traffic to GitLab at the firewall except from the administrator's IP
   • Took a manual backup by streaming tar over SSH to an encrypted partition on a local machine
   • Snapshotted all GCE disks behind the PersistentVolumes
   • Copied the data from the old disk snapshots to new enlarged volumes through a temporary GCE VM
   • Re-did the copy into new empty volumes with the correct directory structure and checked repositories, users and workflows
   • Restored public access to GitLab

   Action items

   • Set up monitoring, logging and alerts
   • Find ways to communicate outages much better; this one was announced only on IRC, and the lists were not told when it dragged on
   • Use the Cloud console clients instead of the slow Cloud Console web UI
   • Add more GitLab administrators, ideally someone familiar with Kubernetes deployments
   • Move from the single Omnibus container to GitLab's cloud native deployment with one container per service
   • Move storage to resizeable volumes or to Google Cloud Storage

   Lessons

   The backup step alone used up the outage window; backups had grown along with the disk problem, and a test run before announcing the window could have measured it. The changes had been tested on a scratch cluster, but not exactly as run, and tests of steps such as disk snapshots could have been more realistic. A single GitLab administrator who was new to Kubernetes was a single point of failure. Keeping at least three copies of the data from the first destructive step meant no data was lost.

   View case
3. IL-0468 New JupyterHub student pods stuck Pending on MaxVolumeCount after a memory limit cut with node size unchanged

   UC Berkeley2017

   Symptoms

   On 20 March 2017, from some time early in the day until about 13:00, some new student servers on UC Berkeley's JupyterHub stayed in Pending forever and the students got 500 errors that kept them from logging in.

   Pods PendingHigh error rate

   Trigger

   On 18 March at 16:30 the RAM per student was reduced from 2G to 1G to save resources while the nodes kept the same size of 26G RAM, and as a result the nodes hit the Google Cloud the limit on the number of disks per node.

   Root cause

   Capacity exhaustion

   Component

   Container memory limitCloud block storageNode pool

   Contributing factor

   Unknown limit

   Detection

   New student servers reported a 500 error that kept students from logging in. At 12:44 this was judged widespread, and at 12:53 `kubectl describe pod` showed the pods Pending on MaxVolumeCount.

   Remediation

   • Copied the Google Container Engine instance template and changed it to spawn smaller n1-highmem-2 nodes
   • Switched the managed instance group to the new template
   • Grew the instance group by 30 new nodes
   • Ran populate.bash so student pods would start on time on the new nodes
   • Stopped the simple autoscaler
   • Stopped and started affected user pods from the control panel
   • Restarted the hub and proxy pods

   Action items

   • Follow the status of kubernetes/kubernetes issue 24317
   • Find and fix the cause of the 'too many redirects' error in JupyterHub
   • Delete the older, larger nodes that are no longer in use
   • Alert when any pods stay Pending for a non-negligible time

   View case
4. IL-0537 Sakura Cloud host servers repeatedly lost their storage connection, about five minutes per port switchover

   Sakura Internet2012

   Symptoms

   From mid-December 2011, after Sakura Cloud launched, traffic between the host servers and the storage device grew, and the network connection between the hosts that carry users' servers and the storage device was cut again and again. Each cut caused by a port switchover lasted about five minutes.

   Storage disconnected or stale mountsPacket loss

   Trigger

   Traffic between the host servers and the storage device grew from mid-December 2011, after the service launched on 15 November 2011.

   Root cause

   Capacity exhaustion

   Component

   Cloud block storageInfiniBand fabric

   Contributing factor

   Automation amplificationUnrepresentative stagingLimited debug access

   Remediation

   • Switched the network interface monitoring from ICMP to a LINK state check, so the device fails over only on a physical link down
   • Changed the InfiniBand Subnet Manager's check from normal mode to heavy sweep mode
   • Updated the storage device firmware on 19 March, which stopped the packet loss
   • Added a second storage device on 12 March and migrated servers to it to reduce and spread the load

   Action items

   • Replace the current storage with an entirely different new storage device
   • Test the new storage in an environment equal to its maximum capacity under real load, with a free beta before paid service
   • Use network interface drivers and similar parts equivalent to those already proven on the host servers
   • Develop the new storage with in-house engineers so Sakura Internet can manage and operate it itself

   Lessons

   Sakura Internet wrote that it could not test the storage device at its performance limit, because it could not prepare as many servers as the design capacity and used load generated from predictions, so it could not respond quickly and accurately to problems in real operation. Packet loss on the InfiniBand port kept the device below its expected performance and took three months to fix by firmware, and its engineers could not grasp the whole of the device's specification and behaviour. Sakura Internet concluded that its process for choosing and validating the storage device had serious problems.

   View case

By root cause1