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

Incident documentation/2019-08-23 network codfw

From Wikitech-static
Jump to navigation Jump to search

document status: final

Summary

A provider outage on our primary transport link between eqiad and codfw caused it to be in a constant flapping (going down and up) state.

This flapping caused routing re-convergence churn and packet loss between the two sites.

On the application level, this translated to elevated 5xx/s from Varnish from ulsfo, eqsin, and codfw from 21:20 to 21:55 UTC. Varnish reported "No backend" for many of the requests. Host checks in Icinga were flapping "TTL exceeded" and service checks flapping "No route to host."

Impact

Surfaced a bit more than 52,000 5xx responses.

https://grafana.wikimedia.org/d/000000479/frontend-traffic?panelId=2&fullscreen&from=1566594998796&to=1566597283616&var-status_type=5

Detection

Monitoring caught and reported the issue via SmokePing and Icinga.

Timeline

All times in UTC.

  • 2019-08-23 21:20 OUTAGE BEGINS
  • 21:25 Investigation begins
  • 21:33 Zayo (the link's provider) reports issue with service (email unnoticed)
  • Lots of errors and recoveries - flapping
  • 21:41 Arzhel starts investigating
  • 21:46 Brandon called
  • 21:47 Decided to depool codfw (ended up not needing it)
  • 21:48 Arzhel promotes backup link to primary
  • 21:55 OUTAGE ENDS
  • 2019-08-25 01:37 Link stops flapping

Conclusions

What went well?

  • The root cause was quickly worked-around once the cause (network link) was identified.

What went poorly?

  • Due to the frequency of the flapping Icinga checks for link status, OSPF and BFD didn't trigger, causing SREs to think of an application layer issue
  • The work-around (failing over to the backup link) is not documented and requires Netops to be done.
  • Nothing paged even though it had user facing impact

Where did we get lucky?

  • Giuseppe, and Filippo responded outside of their office hours.

How many people were involved in the remediation?

  • No Incident coordinator appointed - 5 SREs

Links to relevant documentation

Actionables

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

  • Those two will help mitigate the consequences of an overly flapping link:
    • Configure interface damping on primary links - T196432
    • ospf link-protection - T167306
  • This one will make it easier (down the road) to a non-netops to failover a link if the need arises:
    • Configuration management for network operations - T228388
  • This one is about having better monitoring and alerting by replacing Smokeping by something Prometheus based
    • Investigate/setup prometheus blackbox_exporter - T169860