You are browsing a read-only backup copy of Wikitech. The primary site can be found at wikitech.wikimedia.org

Incident documentation/2018-10-15 eqsin-power

From Wikitech-static
Jump to navigation Jump to search

Summary

The eqsin PoP went offline for approximately 2h10, between approximately 14:15 and 16:25 UTC on 2018-10-14. Of those, 11 minutes were service impacting for the PoP's service region (APAC) for the majority of users, with the long tail of non-conforming resolvers being unable to reach our sites for longer periods of time.

Root cause was a combination of scheduled power maintenance by the facility and lapses in our processes.

Timeline

  • Sep 07 08:17 Facility notifies us of a "SERVICE IMPACTING Scheduled Customer Outage Level" with power feed A window.
  • Sep 07 08:20 Facility notifies us of a "SERVICE IMPACTING Scheduled Customer Outage Level" with power feed B window.
  • Sep 10 10:06 Facility notifies us of power feed B window being rescheduled.
  • Oct 12 13:01 Facility sends a "maintenance listed will commence in one hour" notification for power feed A.
  • Oct 12 14:16 Power feed A is lost in eqsin, and is noticed by monitoring.
  • Oct 12 14:26 Site gets depooled and task is filed.
  • Oct 13 00:00 Maintenance window is over, facility sends a "completed" notice.
  • Oct 13 03:07 Power feed A is (seemingly) back up, so the site is repooled. Unknown at the time was that multiple alerts were still outstanding at that point, and power had not been recovered in the top-half of rack 0603, which included all networking equipment.
  • Oct 14 13:00 Facility sends a "maintenance listed will commence in one hour" notification for power feed B.
  • Oct 14 14:15 Power feed B is lost in eqsin, and with it the PoP goes dark. Users in APAC are unable to access our sites. Monitoring notices and pages, multiple respondents react.
  • Oct 14 14:16 Site is depooled.
  • Oct 14 14:26 DNS TTLs expire for the majority of recursors, site back up for most of APAC users via ulsfo.
  • Oct 14 14:38 Facility's service deks is being contacted to report a double power failure.
  • Oct 14 16:02 Service desk responds with a work order; will "be contacted by a technician in 30 to 45 minutes".
  • Oct 14 16:25 Site is back up, facility technician informs us that they reset a feed A in-rack breaker that was tripped.
  • Oct 14 19:24 Power feed B is restored.
  • Oct 15 00:00 Maintenance window is over, facility sends a "completed" notice.
  • Oct 15 08:48 Site is repooled.

Conclusions

  • Manual triaging of maintenance alerts is error-prone.
  • Repool should always be preceded by a monitoring health-check where all alerts have been cleared or can be reasonably explained.
  • eqsin does not have router or rack redundancy, so a power feed + PDU or PSU failure can take the site down.
  • eqsin's PDUs are not monitored by neither the facility nor us.

Actionables

NOTE: Please add the #wikimedia-incident Phabricator project to these follow-up tasks and move them to the "follow-up/actionable" column.