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

Difference between revisions of "Incident documentation/2021-07-26 ruwikinews DynamicPageList"

From Wikitech-static
Jump to navigation Jump to search
imported>Legoktm
(→‎Summary: +graphs from Google Doc)
imported>Legoktm
Line 1: Line 1:
{{irdoc|status=draft}} <!--
{{irdoc|status=review}} <!--
The status field should be one of:
The status field should be one of:
* {{irdoc|status=draft}} - Initial status. When you're happy with the state of your draft, change it to status=review.
* {{irdoc|status=draft}} - Initial status. When you're happy with the state of your draft, change it to status=review.
Line 9: Line 9:
[[File:Errors during 2021-07-26 DPL outage.png|thumb|Spikes in 503 errors returned to users]]
[[File:Errors during 2021-07-26 DPL outage.png|thumb|Spikes in 503 errors returned to users]]
[[File:Overall traffic during 2021-07-26 DPL outage.png|thumb|Corresponding graphs of traffic during the outage, showing approximately 15% of requests receiving errors.]]
[[File:Overall traffic during 2021-07-26 DPL outage.png|thumb|Corresponding graphs of traffic during the outage, showing approximately 15% of requests receiving errors.]]
Following the bot import of 200,000 pages to the Russian Wikinews in the span of 3 days, slow queries originating from ruwikinews's usage of the [[mw:Extension:DynamicPageList (Wikimedia)|DynamicPageList extension]]  (also known as "[[git:mediawiki/extensions/intersection|intersection]]") overloaded the s3 cluster of databases, causing php-fpm processes to hang/stall, eventually taking down all wikis with it. The outage was resolved by disabling the DynamicPageList extension on ruwikinews and aggressively killing queries on s3 replicas. In worst case usage, DPL's database queries roughly scale to the size of the largest category being checked, which on ruwikinews is orders of magnitude higher than other wikis. One slow query seen during the outage took more than 3 minutes to finish on an idle replica (see [[phab:P16896|EXPLAIN]]).
Following a large bot import to the Russian Wikinews, expanding the size of that project to 13 million pages, slow queries originating from ruwikinews's usage of the [[mw:Extension:DynamicPageList (Wikimedia)|DynamicPageList extension]]  (also known as "[[git:mediawiki/extensions/intersection|intersection]]") overloaded the s3 cluster of databases, causing php-fpm processes to hang/stall, eventually taking down all wikis with it. The outage was resolved by disabling the DynamicPageList extension on ruwikinews and aggressively killing queries on s3 replicas. In worst case usage, DPL's database queries roughly scale to the size of the largest category being checked, which on ruwikinews is orders of magnitude higher than other wikis. One slow query seen during the outage took more than 3 minutes to finish on an idle replica.
 
The [[phab:P16896|EXPLAIN]] indicates that MariaDB is using the PRIMARY index on (cl_from, cl_to) instead of the expected index on (cl_to, cl_timestamp) because the DPL query specified ordermethod=created (aka page_id). This might cause MariaDB to look at 44,435,648 rows, instead of the only 180,231 rows for В_мире, which may have made the query perform even worse.
 
It appears the triggering event was the creation of [[n:ru:Category:В мире]] which was linked to on about 180,000 pages that had a DPL query on them. This page creation triggered a large number of parse jobs (so links could change from red to blue) making the same DPL query, which began to overload the s3 DB, which then snowballed. The caching mitigation introduced during the previous incident did not work properly as the query often did not complete in time to be cached, even before the incident.


'''Impact''': There was 30 minutes of outage with high latencies, failing to load pages and errors around all wikis due to unavailable PHP-FPM workers. Based on traffic graphs the outage impacted to approximately a 15% of all incoming HTTP requests for wikis, those being either lost, suffering high latencies or 5XX error codes. The main impact was uncached requests, suffering a 0% availability during several moments of the outage, on all wikis.
'''Impact''': There was 30 minutes of outage with high latencies, failing to load pages and errors around all wikis due to unavailable PHP-FPM workers. Based on traffic graphs the outage impacted to approximately a 15% of all incoming HTTP requests for wikis, those being either lost, suffering high latencies or 5XX error codes. The main impact was uncached requests, suffering a 0% availability during several moments of the outage, on all wikis.
Line 19: Line 23:
'''Prologue'''
'''Prologue'''
* 2020-09-07 and 2020-09-08: Following rapid bot imports (~100k pages in 1 day), DynamicPageList queries from ruwikinews caused problems on s3, though it did not lead to a sitewide outage. A summary of that incident is available at [[Phab:T262240#6449531|T262240#6449531]] (TODO: create proper incident report).
* 2020-09-07 and 2020-09-08: Following rapid bot imports (~100k pages in 1 day), DynamicPageList queries from ruwikinews caused problems on s3, though it did not lead to a sitewide outage. A summary of that incident is available at [[Phab:T262240#6449531|T262240#6449531]] (TODO: create proper incident report).
* 2021-07-14 through 2021-07-17: [[n:ru:User:NewsBots|NewsBots]] imports/creates 200,000 pages to the Russian Wikinews in the span of 3 days (per [[m:Wikimedia News|Wikimedia News]]).
* 2021-07-12 through 2021-07-26: [[n:ru:User:NewsBots|NewsBots]] imports/creates 3,182,955 pages to the Russian Wikinews.
* 2021-07-26: Something happens, triggering cache invalidation of pages using the <code><nowiki><DynamicPageList></nowiki></code> tag.
* 2021-07-24 through 2021-07-26 - ruwikinews starts logging "Async refresh failed for ruwikinews:DPLQuery:4ea73c61c64e57c48f89a3da9caee058cbc5888bc11d2597bcfa957dd542f4e4". This will spike during the incident. The hash corresponds to the DPL query used in the infobox included on pages in [[n:ru:Категория:В_мире]]. This happens when the query takes longer than 60 seconds, and indicates that the caching solution from the previous incident meant to reduce risk of cache stampedes is no longer working.
* 2021-07-26: In the lead up to the incident, DPL queries are often taking up to 20 seconds to complete


'''Main outage, all times in UTC.'''
'''Main outage, all times in UTC.'''
 
* 10:29: [[n:ru:Категория:В_мире]] is created by a logged out user [https://ru.wikinews.org/w/index.php?title=%D0%A1%D0%BB%D1%83%D0%B6%D0%B5%D0%B1%D0%BD%D0%B0%D1%8F:%D0%96%D1%83%D1%80%D0%BD%D0%B0%D0%BB%D1%8B&logid=19137921]. This category is linked to on 179,385 pages, triggering a re-parse job for each page. Most of these pages have an infobox containing the problematic DPL query in question. During the incident much of the log entries have the request ID ''b0324fd-a93e-4aba-ab50-e58be6f2c38d'' from this edit.
* 10:30: Database overload starts '''OUTAGE BEGINS'''
* 10:30: Database overload starts '''OUTAGE BEGINS'''
* 10:30: Lots of the Parsoid jobs triggered by the edit to the category are reporting as failed.
* 10:33: Page fire for both appserver and api_appserver clusters: <code>Not enough idle PHP-FPM workers for Mediawiki</code>
* 10:33: Page fire for both appserver and api_appserver clusters: <code>Not enough idle PHP-FPM workers for Mediawiki</code>
* 10:34: Significant IRC alert spam ensues, comms move to #wikimedia-sre and #mediawiki_security
* 10:34: Significant IRC alert spam ensues, comms move to #wikimedia-sre and #mediawiki_security
Line 86: Line 92:
* {{stalled}} [[phab:T263220|T263220]] Limit concurrency of DPL queries
* {{stalled}} [[phab:T263220|T263220]] Limit concurrency of DPL queries
*[[phab:T287983|T287983]] Raw "upstream connect error or disconnect/reset before headers. reset reason: overflow" error message shown to users during outage
*[[phab:T287983|T287983]] Raw "upstream connect error or disconnect/reset before headers. reset reason: overflow" error message shown to users during outage
*[[phab:T288180|T288180]] Investigate [[mw:Extension:GoogleNewsSitemap|Extension:GoogleNewsSitemap]] query performance
* Maybe it would be cool if job runners slow down if a DB overload is detected
*Consider removing ordermethod=created as an unnecessary footgun
* ...
* ...

Revision as of 00:16, 5 August 2021

document status: in-review

Summary

File:Errors during 2021-07-26 DPL outage.png
Spikes in 503 errors returned to users
File:Overall traffic during 2021-07-26 DPL outage.png
Corresponding graphs of traffic during the outage, showing approximately 15% of requests receiving errors.

Following a large bot import to the Russian Wikinews, expanding the size of that project to 13 million pages, slow queries originating from ruwikinews's usage of the DynamicPageList extension (also known as "intersection") overloaded the s3 cluster of databases, causing php-fpm processes to hang/stall, eventually taking down all wikis with it. The outage was resolved by disabling the DynamicPageList extension on ruwikinews and aggressively killing queries on s3 replicas. In worst case usage, DPL's database queries roughly scale to the size of the largest category being checked, which on ruwikinews is orders of magnitude higher than other wikis. One slow query seen during the outage took more than 3 minutes to finish on an idle replica.

The EXPLAIN indicates that MariaDB is using the PRIMARY index on (cl_from, cl_to) instead of the expected index on (cl_to, cl_timestamp) because the DPL query specified ordermethod=created (aka page_id). This might cause MariaDB to look at 44,435,648 rows, instead of the only 180,231 rows for В_мире, which may have made the query perform even worse.

It appears the triggering event was the creation of n:ru:Category:В мире which was linked to on about 180,000 pages that had a DPL query on them. This page creation triggered a large number of parse jobs (so links could change from red to blue) making the same DPL query, which began to overload the s3 DB, which then snowballed. The caching mitigation introduced during the previous incident did not work properly as the query often did not complete in time to be cached, even before the incident.

Impact: There was 30 minutes of outage with high latencies, failing to load pages and errors around all wikis due to unavailable PHP-FPM workers. Based on traffic graphs the outage impacted to approximately a 15% of all incoming HTTP requests for wikis, those being either lost, suffering high latencies or 5XX error codes. The main impact was uncached requests, suffering a 0% availability during several moments of the outage, on all wikis.


Timeline

Prologue

  • 2020-09-07 and 2020-09-08: Following rapid bot imports (~100k pages in 1 day), DynamicPageList queries from ruwikinews caused problems on s3, though it did not lead to a sitewide outage. A summary of that incident is available at T262240#6449531 (TODO: create proper incident report).
  • 2021-07-12 through 2021-07-26: NewsBots imports/creates 3,182,955 pages to the Russian Wikinews.
  • 2021-07-24 through 2021-07-26 - ruwikinews starts logging "Async refresh failed for ruwikinews:DPLQuery:4ea73c61c64e57c48f89a3da9caee058cbc5888bc11d2597bcfa957dd542f4e4". This will spike during the incident. The hash corresponds to the DPL query used in the infobox included on pages in n:ru:Категория:В_мире. This happens when the query takes longer than 60 seconds, and indicates that the caching solution from the previous incident meant to reduce risk of cache stampedes is no longer working.
  • 2021-07-26: In the lead up to the incident, DPL queries are often taking up to 20 seconds to complete

Main outage, all times in UTC.

  • 10:29: n:ru:Категория:В_мире is created by a logged out user [1]. This category is linked to on 179,385 pages, triggering a re-parse job for each page. Most of these pages have an infobox containing the problematic DPL query in question. During the incident much of the log entries have the request ID b0324fd-a93e-4aba-ab50-e58be6f2c38d from this edit.
  • 10:30: Database overload starts OUTAGE BEGINS
  • 10:30: Lots of the Parsoid jobs triggered by the edit to the category are reporting as failed.
  • 10:33: Page fire for both appserver and api_appserver clusters: Not enough idle PHP-FPM workers for Mediawiki
  • 10:34: Significant IRC alert spam ensues, comms move to #wikimedia-sre and #mediawiki_security
  • 10:35: "upstream connect error or disconnect/reset before headers. reset reason: overflow" on enwiki
  • 10:38: Manuel depools db2149, that seems the most affected DB (SAL entry)
  • 10:39: T287362 filed by users unable to access arwiki,
  • 10:40: after a brief apparent recovery the load shifts to another DB
  • 10:42: Slow query identified as coming from DynamicPageListHooks::processQuery
  • 10:42: Link to previous incident from 2020-09 established (T262240), people involved in that ticket pinged on IRC
  • 10:46-10:49: Manuel slowly repools db2149
  • 10:48: Recommendation made to disable DynamicPageList on ruwikinews instead of increasing cache TTL
  • 10:50: Incident opened (private Google Doc).
File:MariaDB traffic during 2021-07-26 DPL outage.png
MariaDB traffic during the outage showing a dramatic increase in rows read and number of open connections.
  • 10:51: Jaime sets the query killer on S3 replicas to 10 seconds for the MediaWiki user (SAL entry)
  • 10:55: Amir disables DPL on ruwikinews (SAL entry)
  • 10:56: Icinga recoveries starts to be fired
  • 10:59: Database throughput back to normal levels OUTAGE ENDS
  • 11:01: Last Icinga recovery

Detection

Icinga sent two pages at 10:33 for Not enough idle PHP-FPM workers for Mediawiki on the appserver and api_appserver clusters.

The first user report on IRC appears to have been at 10:35 in #wikimedia-sre: <RhinosF1> Meta is down.

Because this was a full outage, every host was individually alerting and so were services that depend upon MediaWiki. Each appserver triggered two alerts, like:

<icinga-wm> PROBLEM - Apache HTTP on mw2316 is CRITICAL: CRITICAL - Socket timeout after 10 seconds https://wikitech.wikimedia.org/wiki/Application_servers
<icinga-wm> PROBLEM - PHP7 rendering on mw2316 is CRITICAL: CRITICAL - Socket timeout after 10 seconds https://wikitech.wikimedia.org/wiki/Application_servers/Runbook%23PHP7_rendering

icinga-wm sent 145 messages to #wikimedia-operations between 10:34 and 10:36 before being kicked off the Libera Chat network for flooding. That IRC channel was unusable and discussion was moved to #wikimedia-sre and then #mediawiki_security.

Conclusions

What went well?

  • Automated monitoring detected the incident before humans did
  • Once the problematic query was identified, it was immediately linked to the previous incident
  • Once the extension was disabled + query killer running on s3, everything came back up

What went poorly?

  • DPL had already been identified as problematic and the primary mitigation to limit concurrency of DPL queries (T263220) had never actually been deployed.
  • Users did not see the standard "Wikimedia error" screen, instead got a cryptic plaintext error message
  • icinga-wm spammed #wikimedia-operations to unusability and then flooded off
  • A formal incident was only opened two-thirds of the way through the outage
  • Harassment of sysadmins who were involved in incident response aftewards

Where did we get lucky?

  • We did not get lucky.

How many people were involved in the remediation?

  • ~7 SREs and 1 software engineer

Links to relevant documentation

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

High-level discussion and brainstorming is happening in T287380: Decide on the future of DPL. Out of that, some specific actionables have been identified:

  • Yes Done 708390 Send queries to "vslow" database group
  • Underway In progress T287916 Disable DPL on wikis that aren't using it
  •  Stalled T263220 Limit concurrency of DPL queries
  • T287983 Raw "upstream connect error or disconnect/reset before headers. reset reason: overflow" error message shown to users during outage
  • T288180 Investigate Extension:GoogleNewsSitemap query performance
  • Maybe it would be cool if job runners slow down if a DB overload is detected
  • Consider removing ordermethod=created as an unnecessary footgun
  • ...