You are browsing a read-only backup copy of Wikitech. The live site can be found at


From Wikitech-static
Jump to navigation Jump to search


Phabricator ( was unavailable for 1 hour due to incorrect puppet configuration. The configuration was applied automatically after the server was rebooted during a scheduled maintenance window.


  • 01:00: Scheduled phabricator maintenance
  • 01:00: Moritz reboots iridium (Phabricator host) for a scheduled kernel update.
  • 01:04: Phabricator is back up.
  • 01:05: Puppet run on iridium updates Phabricator source tree; it had hither to been prevented via a lock file in /var/run, which the reboot cleared. The versions of Phabricator and its dependency, libphutil, are now mismatched.
  • 01:05: Phabricator is down. Icinga alert from PyBal: could not depool server about iridium-vcs.eqiad.wmnet!
  • 01:08: Mukunda resets Phabricator source tree to desired state
  • 01:23: puppet run resets repositories to broken state again but this goes unnoticed because apache is still serving cached good state
  • ~01:30: Mukunda ensures the lock file is in /var/run as it should be, but without realizing a puppet run happened at 01:24.
  • 01:30: Mukunda starts writing patches for puppet to prevent further incidents
  • 01:41: Mukunda submits a patch to move the lock to a persistent location:
  • 01:44: Mukunda submits a patch to fiix a bug in the git:install puppet module (discovered while writing previous patch):
  • 01:45: Everything still appears to be running fine and patches submitted. Mukunda finished with maintenance.
    • expectation: Puppet won't touch the repos again unless it's rebooted before merges.
    • reality: puppet already reset the repos but opcode cache is serving cached-good-state.
  • 06:30: cron.daily logrotate triggers apache graceful restart which clears the opcode cache, reverting once again to the broken state.
  • 06:31: Phabricator down again; check_https_phabricator alerts.
  • 07:30: Chase checks out release/2015-11-18/1 tag in Phabricator source tree. Phabricator comes back up.
  • ...
  • Our custom Phabricator "Sprint" extension is not displaying sprint workboards still - task T125832
  • 16:52: Mukunda restarted apache2 on iridium so that phabricator recognizes sprint.phragile-uri, fixing the Sprint extension
  • 16:53: restarted phd to synchronize settings with phabricator


The root cause of this incident is a combination of multiple oversights, the most important are highlighted in bold below.

  1. Puppet is configured to update all phabricator related repositories to a specific hard-coded git tag in the absence of a lock file.
    1. The location chosen for the lock file was within /var/run
    2. /var/run is not on persistent storage, therefore, the lock file is lost when the server is rebooted.
  2. Phabricator deployments are being transitioned from a one-off puppet deployment (hard-coded tags in operations/puppet repository) to a scap3-based deployment. (See task T114363)
  3. The deployed state of phabricator has been maintained in the phabricator/deployment repository for some time, however, puppet hasn't been updated.
  4. Normally puppet doesn't update the deployed tags when a lock file is present, however, after a reboot the lock file is lost.
  5. Almost immediately after the reboot, the repositories were manually reset to the correct tags
  6. Some time elapsed before the lock file was checked.
  7. Puppet ran during this time without being noticed.
  8. apache graceful restart clears the opcode cache and apache begins serving the incorrect tags
    1. This is what triggered the outage.
    2. The lock file should have prevented this from occurring, however...
      1. Puppet ran at 01:23 ...Which was after the repo was fixed at 01:08 ...Unfortunately, that was before ~01:30 when the lock file was confirmed to have been created.
    3. The time it took for Mukunda to double-check the lock file, after repairing the repo, allowed puppet to revert phabricator once again, but without being noticed.


  • Completely refactor the phabricator module in puppet to remove the automatic git tag pinning behavior. task T125851
  • move /srv/phab/repos to /srv/repos then replace /srv/phab with a symlink to /srv/deployment/phabricator/deployment (Note: Need to schedule a maintenance window for this) task T125853
  • Use the puppet provider for scap to provision phabricator ( -- merged!
  • Further deployments of phabricator use scap3. task T114363 -- done but won't be closed til the above items are resolved