Incident documentation/20170323-wikibase

From Wikitech
Jump to: navigation, search

Summary

Commit https://gerrit.wikimedia.org/r/344564 in the Wikibase extension was meant to make a change to Special:RecentChanges. This change was tested and verified to not be broken.

The commit also moved around unrelated code for magic word {{noexternallanglinks}}. This part of the change contained a bad class reference (missing namespace declaration), which, when executed would result in a PHP Fatal Error.

  • This magic word did not have unit tests, hence Jenkins did not catch it.
  • The magic word is only rarely used on a page, which meant that cursory testing did not reveal the breakage immediately. It also meant that the canary traffic Scap waits for during the deployment did not trigger the error (or not often enough).
  • One such page using this magic word is en.wikipedia.org's Main Page. However, since it is only rarely edited, it is mostly served from parser cache. However, unlike most pages which have a parser cache age of 30 days, Main Page uses time-based functions and hence has a parser cache age of 1 hour. That hour was up about 20 minutes after the deployment, at which point it got re-parsed and revealed the Fatal Error.

Timeline

(Times on this timeline refer to UTC though the date in the page title refers to US Time. According to UTC this event happened on 24th of March, not 23th.).

  • 2017-03-24T01:02 Roan merges an update for Wikidata (Wikibase) (patch for RCFilters with a broken class autoloader) https://gerrit.wikimedia.org/r/#/c/344564/
  • 2017-03-24T01:15 !log catrope@tin Started scap: Wikidata cherry-picks (with i18n)
  • 2017-03-24T01:40 !log catrope@tin Finished scap: Wikidata cherry-picks (with i18n) (duration: 25m 03s)
  • 2017-03-24T01:46 Icinga says: <icinga-wm> PROBLEM - MediaWiki exceptions and fatals per minute on graphite1001 is CRITICAL: CRITICAL: 80.00% of data above the critical threshold [50.0] (but does not page anyone)
  • 2017-03-24T02:01 Eddiegp reports on IRC: <eddiegp> Uhmm, dewiki seems down for me <eddiegp> Request from 94.134.245.179 via cp3033 cp3033, Varnish XID 735457820 <eddiegp> Error: 503, Service Unavailable at Fri, 24 Mar 2017 02:00:09 GMT
  • 2017-03-24T02:05 icinga-wm starts flooding first alerts from app servers (but still does not page)
02:05 <icinga-wm> PROBLEM - HHVM rendering on mw1256 is CRITICAL: HTTP CRITICAL: HTTP/1.1 500 Internal Server Error - 50620 bytes in 7.848 second response time
02:05 <icinga-wm> PROBLEM - HHVM rendering on mw1237 is CRITICAL: HTTP CRITICAL: HTTP/1.1 500 Internal Server Error - 50620 bytes in 7.895 second response time
02:05 <icinga-wm> PROBLEM - HHVM rendering on mw1189 is CRITICAL: HTTP CRITICAL: HTTP/1.1 500 Internal Server Error - 50620 bytes in 7.998 second response time
02:05 <icinga-wm> PROBLEM - HHVM rendering on mw1245 is CRITICAL: HTTP CRITICAL: HTTP/1.1 500 Internal Server Error - 50620 bytes in 8.123 second response time
  • 2017-03-24T02:09 Daniel sees the above and starts looking at mw1185, calls Brandon while doing so
  • 2017-03-24T02:10 Brandon starts investigating (<bblack> the first and most-persistent of the spam is on rendering), <mutante> on a random one, mw1185, status of HHVM is "active (running)"..
  • 2017-03-24T02:10 <mutante> Fatal error: Class undefined: Wikibase\Client\Hooks\NoLangLinkHandler in /srv/mediawiki/php-1.29.0-wmf....n line 99. Meanwhile Icinga service alerts are flapping from RECOVERY back to CRIT and vice versa.
  • 2017-03-24T02:14 <Reedy> Fatal error: Class undefined: Wikibase\Client\Hooks\NoLangLinkHandler in /srv/mediawiki/php-1.29.0-wmf.17/extensions/Wikidata/extensions/Wikibase/client/includes/Hooks/MagicWordHookHandlers.php on line 99
  • 2017-03-24T02:14 <Krinkle> commit 1f32a896a6207bf78d9e97a9fdb49818256e0f3a
  • 2017-03-24T02:16 Krinkle starts reverting
  • 2017-03-24T02:17 <Reedy> RoanKattouw: Looks like wrong NS
  • 2017-03-24T02:24 l10nupdate slows down recovery: <MaxSem> !log Killed l10nupdate on tin, was blocking emergency pushes
  • 2017-03-24T02:24 <logmsgbot> !log krinkle@tin Synchronized php-1.29.0-wmf.17/extensions/Wikidata: revert (duration: 02m 34s)
  • 2017-03-24T02:25 <Krinkle> !log All apaches are back up
02:25 <icinga-wm> RECOVERY - HHVM rendering on mw1230 is OK: HTTP OK: HTTP/1.1 200 OK - 78707 bytes in 0.099 second response time
02:25 <icinga-wm> RECOVERY - HHVM rendering on mw1280 is OK: HTTP OK: HTTP/1.1 200 OK - 78707 bytes in 0.072 second response time
  • 2017-03-24T02:25 <eddiegp> works for me again


Conclusions

  • Magic words and hook handlers should be covered by unit tests, even if just for the sake of code coverage with no useful assertions.
  • The autonomous nature of LocalisationUpdate interfered with a critical deployment when least expected. If l10nupdate were done via SWAT or Train deployment by a human, it would not have started on its own while in the the middle of an outage.
  • Scap's monitoring of logstash with canary traffic is insufficient. High profile requests may not hit the canaries most of the time due to caching. Testing a small set of fixed urls on Tin's local Apache (or canaries, or even every server) would give a much stronger confidence in at least basic functionality without question.

Actionables