Incidents/2021-09-13 cirrussearch restart

From Wikitech

document status: final

Summary

Applying a change to our cirrus elasticsearch systemd unit triggered puppet to automatically restart the associated elasticsearch services. Sufficient number of hosts (6) were having puppet-agent ran simultaneously to drop the cluster into yellow and then red status; the large bandwidth/resources required to reschedule and assign shards when the respective elasticsearch services came back up on each host likely contributed to general cluster instability leading to a cascading failure scenario.

Elasticsearch itself did its job in coming back and immediately recovering shards, but given the volume of hosts restarted, it took about 40 minutes for the impact to users to become largely invisible, and 2 hours and 5 minutes for a full recovery of the backend (i.e. green cluster status).

  • 16:18 Puppet change touching cirrus elasticsearch systemd units merged: https://gerrit.wikimedia.org/r/c/operations/puppet/+/720667
  • 16:23 operator runs puppet manually on elastic2052, confirmed elasticsearch service came back up properly
  • 16:25 operator runs puppet on rest of fleet, 6 hosts at a time: sudo cumin -b 6 'P{elastic*}' 'sudo run-puppet-agent'
  • 16:29 PROBLEM - PyBal backends health check on lvs1015 is CRITICAL: PYBAL CRITICAL - CRITICAL - search-psi-https_9643: Servers elastic1059.eqiad.wmnet, elastic1049.eqiad.wmnet, elastic1044.eqiad.wmnet, elastic1048.eqiad.wmnet, elastic1052.eqiad.wmnet, elastic1047.eqiad.wmnet, elastic1067.eqiad.wmnet, elastic1035.eqiad.wmnet, elastic1045.eqiad.wmnet are marked down but pooled https://wikitech.wikimedia.org/wiki/PyBal
  • 16:31 ElasticSearch health check for shards on 9243 on search.svc.codfw.wmnet is CRITICAL. At the same time, production-search-codfw has dropped into red cluster status
  • 17:09 enwiki searches begin to work again (presumably due to recovery of the associated enwiki shards). User-visible impact decreases.
  • 18:xx codfw enters green cluster status (full recovery)
  • 18:xx eqiad enters green cluster status (full recovery)
  • 18:30 MediaWiki api_appserver latency returns to normal levels
  • 18:34 PoolCounter stops rejecting requests, user-visible impact ends.
The api_appserver cluster saw higher latency during the outage.

Impact: For about 2 hours (from 16:29 until 18:34) search requests on en.wikipedia.org (and likely other wikis) failed with "An error has occurred while searching: Search is currently too busy. Please try again later.". Search suggestions (from API opensearch) were absent or delayed. During the incident, the api_appserver cluster saw higher average latency overall due to the proportion of search queries.

Documentation:

Actionables

  • task T290902 Iron out procedure to roll out cirrus/elasticsearch changes involving [implicitly via puppet or explicitly] service restarts