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

Incidents/2022-12-18 World Cup

From Wikitech-static
Jump to navigation Jump to search

document status: draft

Summary

Incident metadata (see Incident Scorecard)
Incident ID 2022-12-18 World Cup Start 2022-12-18 17:58:00
Task T325477 End 2022-12-18 18:22:00
People paged 34 Responder count 6 (initial page, later others took over final corrections)
Coordinators Jcrespo Affected metrics/SLOs ?
Impact All API users experienced 5XX errors (11-12 thousand errors per second) or unreasonable latencies for 24 minutes. After that, there was degraded performance in the form of increased -but more reasonable- latency for around 3 hours. Edit rate got low during the 24 minute hard outage period, but recovered quickly after it (to a higher level than usual).

For approximately 24 minutes, uncached calls to the API on the eqiad datacenter overloaded the application servers, running out of threads (all busy) creating unreasonable latency or failing to respond to requests. This caused sending errors to some clients using the action API and Parsoid in the primary datacenter. Elevated latencies persisted for the following 3 hours, when traffic load organically went down.

This was caused by the DiscussionTools MediaWiki extension adding a ResourceLoader module on almost all page views -even non-discussion pages- which created an API call, that, combined with a significant 50% increase in overall traffic, led to an overload and increased latencies on the API cluster. codfw app server cluster was mostly unaffected due to not receiving POST uncached traffic at the time (it is read-only).

Timeline

File:World cup traffic increase.png
Increased (around 50%) traffic on both text and eqiad cluster, trigger of the subsequent issues
File:Discussiontools load world cup.png
Excesive load created by the Discussiontools module, taking more execution time than even query and parse, the underlying issue
File:API appserver eqiad busy workers and status codes world cup.png
Direct cause of the outage: running out of available API workers to execute user's requests, leading to errors and latencies
File:API appserver eqiad latencies world cup.png
Increased latences during the outage, and also after the saturation got resolved
File:Edit rate during World Cup.png
Edit rate during the World Cup final: it had elevated values except during the hard outage, probably due to editing bots being affected by the API POST and Parsoid/JS API calls being down
File:Parsercache hit ratio.png
Metrics show a dramatic increase in parsing requests after the 8th deploy and a recovery after the 19th fix

All times in UTC.

8 December 2022

18 December 2022

  • 17:55 Argentina scores the penalty that awards their national team the World Cup. A 50% traffic increase starts (we had larger spikes in the past, though, causing no load issues)
  • 17:58 API app servers get overload with requests OUTAGE STARTS HERE
  • 17:59 4 pages are sent: PHPFPMTooBusy (api_appserver & parsoid) / FrontendUnavailable (HAProxy & varnish)
  • 18:XX Thinking it is load related, a restart is considered not the right move at the moment
  • 18:20 App servers are attempted to be restarted in a desperate attempt to fix something, failing END (FAIL) - Cookbook sre.mediawiki.restart-appservers (exit_code=99)
  • 18:22 Load goes down enough for worker threads to not be 100% busy OUTAGE ENDS (but degraded state with increased latencies persist)
  • 18:55 A rolling restart is finally run to try to improve the latency issues: sudo cumin -b 1 -s 30 'A:mw-api and A:eqiad' 'restart-php7.4-fpm'
  • 19:31 Rolling restart finishes, latencies still high
  • 19:42 DiscussionTools is suspected to be the culprit based on utilization metrics, but there is no one around to debug
  • 19:46 Latencies -due to organic traffic reduction, not to any action taken- decrease to a degraded, but not outage-like state. It is decided to set the ongoing outage as "resolved" and ask for debugging help the following day.

19 December 2022

(any one of these three patches would have resolved the issue)

Detection

Monitoring and paging worked as intended, paging everybody (it was a weekend) as soon as the issue become major at 17:56:

  • 17:59 Service: [FIRING:1] ProbeDown (10.2.2.22 ip4 api-https:443 probes/service http_api-https_ip4 ops page eqiad prometheus sre)
  • 18:00 Service: [FIRING:1] PHPFPMTooBusy api_appserver (ops php7.4-fpm.service page eqiad prometheus sre)
  • 18:00 Service: [FIRING:1] FrontendUnavailable (varnish-text page thanos sre)
  • 18:01 Service: [FIRING:1] FrontendUnavailable cache_text (page thanos sre)

A task was created also by a community member at 19:03, when the team was already analyzing the issue.

However, there were reports that "php fpm busy has been flapping all weekend, just not enough to page".

Conclusions

OPTIONAL: General conclusions (bullet points or narrative)

What went well?

  • Multiple people reacted to page although it was during World Cup finals
  • Good insights with monitoring helped in pinpointing DiscussionTools as potential incident cause and World Cup traffic as the modulator

OPTIONAL: (Use bullet points) for example: automated monitoring detected the incident, outage was root-caused quickly, etc

What went poorly?

  • Traffic spike due to World Cup was quite a surprise although the event was scheduled beforehand for month
  • Knowledge and documentation of sre.mediawiki.restart-appservers cookbook (SERVICE name unknown). It took a while to find the cumin + "restart-php7.4-fpm" approach
  • Increased load following the 8 December deployment was not noticed or not connected to the deployment
  • Outage was caused by a known problem documented in a TODO comment T325477#8476954

OPTIONAL: (Use bullet points) for example: documentation on the affected service was unhelpful, communication difficulties, etc

Where did we get lucky?

  • Traffic declined naturally and incident resolved automatically
  • DoS between 18:17 and 18:36 had no big impact on performance(?), rate limiting worked(?)

OPTIONAL: (Use bullet points) for example: user's error report was exceptionally detailed, incident occurred when the most people were online to assist, etc

Links to relevant documentation

Service restarts#Application servers (also image/video scalers and job runners and parsoid)

Add links to information that someone responding to this alert should have (runbook, plus supporting docs). If that documentation does not exist, add an action item to create it.

Actionables

Scorecard

Incident Engagement ScoreCard
Question Answer

(yes/no)

Notes
People Were the people responding to this incident sufficiently different than the previous five incidents? yes
Were the people who responded prepared enough to respond effectively no SRE had no insight into existence of the DiscussionTools ongoing A/B Test
Were fewer than five people paged? no
Were pages routed to the correct sub-team(s)? no
Were pages routed to online (business hours) engineers?  Answer “no” if engineers were paged after business hours. no
Process Was the "Incident status" section atop the Google Doc kept up-to-date during the incident? yes
Was a public wikimediastatus.net entry created? yes https://www.wikimediastatus.net/incidents/2w4ygdj4vc20
Is there a phabricator task for the incident? yes T325477
Are the documented action items assigned? yes see above also completed so this incident doesn't repeat
Is this incident sufficiently different from earlier incidents so as not to be a repeat occurrence? no
Tooling To the best of your knowledge was the open task queue free of any tasks that would have prevented this incident? Answer “no” if there are

open tasks that would prevent this incident or make mitigation easier if implemented.

no Outage was caused by a known problem documented in a TODO comment T325477#8476954
Were the people responding able to communicate effectively during the incident with the existing tooling? yes
Did existing monitoring notify the initial responders? yes
Were the engineering tools that were to be used during the incident, available and in service? yes
Were the steps taken to mitigate guided by an existing runbook? no runbook failed
Total score (count of all “yes” answers above) 8