Incidents/2020-04-07 Wikidata's wb items per site table dropped

From Wikitech

document status: in-review

Summary

Impact: For about 20 minutes, almost all content page parses were broken. They received, along with any other user interaction with the wb_items_per_site table in some way, a database error message instead of page content. This affected many (but not all or even most) page views for logged-in users, and many (possibly most, but not all) edits made to wikis. For 19 hours, all users looking at pages rendered since the incident start received some incorrect metadata on pageviews (infobox content, cross-language links, wikdata item relation). Also, Wikidata's checks for duplicate site links failed, leading to many hundreds of duplicate items being created.

Cause: Wikidata's wb_items_per_site secondary table, used for the rendering of every page connected to Wikidata, was dropped by a mis-configured weekly cron script which executed the update.php code path, which itself had been misconfigured for eight years to drop this table. This immediately led to a DBQueryError on content page loads (reads). The table was re-created as blank, at which point pages began to paint again (though wrongly).

Timeline

All times in UTC.

2020-04-06
  • 23:00 weekly cron from puppet cron/wikibase/dumpwikibaserdf.sh executes
  • 23:02 The cron calls cron/wikibase/wikibasedumps-shared.sh, which calls sql.php to run a minor ad hoc query.
    • Due to phab:T157651, noted in February 2017, this ran LoadExtensionSchemaUpdates, bypassing the checks in update.php to prevent this ever happening in production
    • Due to a bug in Wikibase's LoadExtensionSchemaUpdates code dating from eight years ago, this wrongly dropped the wb_items_per_site table. OUTAGE BEGINS
  • 23:02:13: first log message indicating the table is missing
  • 23:03:57: first user report of an issue: <NotASpy> just got an error - [Xou1IQpAIDEAAC74T5wAAAOQ] 2020-04-06 23:03:09: Fatal exception of type "Wikimedia\Rdbms\DBQueryError" (log)
  • 23:05:33: first automated report of the issue: <icinga-wm> PROBLEM - MediaWiki exceptions and fatals per minute on icinga1001 is CRITICAL
  • 23:26 Amir re-creates the table as blank. Users no longer receive error messages instead of content (but metadata might be inaccurate). OUTAGE ENDS, USER IMPACT CONTINUES
  • 23:31 Amir triggers rebuildItemsPerSite to begin
  • 23:59 Adam removes broken code from Wikibase's LoadExtensionSchemaUpdates in production
2020-04-07

TODO: Add epilogue/cleanup.

Mediawiki exceptions/fatals

Edge traffic error rate

Detection

  • There was not an alert that indicated this was happening for most parser cache renders, or that there was a high rate (>700/second) of Mediawiki serving 50X errors to the cache servers [1]. There should have been automated detection of such a condition, including paging SRE -- if this had happened at off-hours, it could have gone unnoticed by engineers for far too long.

Underlying problems

wb_terms, when introduced 7.5 years ago, was actually made by merging several tables into one. The code to update the database when you migrate from an old version checks if wb_terms exists and if not, it creates the wb_terms table and drops those old tables (like wb_aliases). It mistakenly had an important table in the list that we still use today: wb_items_per_site (here's the original patch). This bug went unnoticed for 7.5 years.

There's another aspect of the problem. The code that runs the database update (update.php) should never ever be run in production. After too many outages, update.php checks if you're in WMF production and it errors out if that's the case. But here comes another bug, this time in core: sql.php runs LoadExtensionSchemaUpdates which basically means one of the production scripts called sql.php (not update.php) that is being used to debug production database (an interactive shell so you can run db queries against production) mistakenly runs update.php for extensions behind the scene and as a bonus without the check that actually update.php does (to error if it's in WMF production). This is a quote from Tgr on the ticket three years ago: "This is a very, very ugly accident waiting to happen".

2020-04-07 around midnight CET, a cronjob for building the dumps called sql.php to run a query (instead of doing it directly), this has been being run weekly for years. Meaning update.php in production has been running every week on Monday 23:00 UTC for years without us noticing. At the same day, wb_terms table got renamed (as a step to drop the table phab:T208425). The update.php code path, thinking that our production database was an old version that didn't even have wb_terms, tried to create it and drop old tables and mistakenly dropped an important table with 70M rows.

Conclusions

What went well?

  • Outage cause was root-caused quickly
  • WMF Backup/restore system worked as intended, in terms of data recovery
  • Developers familiar with the relevant codebase were readily available and actively working on a fix/mitigation
  • DBAs were responsive despite late night/early morning time for them

What went poorly?

  • Options for restoring such a large table from backup late at night with limited coverage were not straightforward or appealing. More availability/coverage would have allowed us to proceed with this earlier.
    Comment (jynus): I disagree- I think there was a missunderstanding thinking the last backup was 1 week old. However, that is not true, we could have done a full recovery of the full wikidata database in under 2 hours to all hosts (requiring read only), last backup was less than 24 hours old, and we had also binlogs. A decision was taken to do a partial, point in time recovery, as only non-user data was affected while keeping the site read-write (which I agree with, but which of course, takes more time). Of course, someone available has to know how and what to restore.
    More comments on the Discussion page
  • No pages were sent
  • Unnecessary GRANTS (DROP) on wikiadmin user

Where did we get lucky?

  • This was "only" secondary data, and could be re-created by a maintenance script (albeit one that will take 1-2 weeks to run for 80 million items (rough estimate)).
  • We've been running this cron for months/years, effectively running update.php on wikidatawiki every week, and this was the first time it shot us in the foot.

How many people were involved in the remediation?

  • The initial response: 1 IC (SRE), 1 DBA, 1 SRE, 2 WMDE Developers, 3 WMF Developers, 2 volunteers
  • Aftermath fixing a few hours later: 2 WMDE Developers, 2 DBAs

Links to relevant documentation

Actionables

  • Remove ALTER and DROP permissions from the wikiadmin user (which is used by automated crons). Create a separate user for humans performing schema changes. https://phabricator.wikimedia.org/T249683
    • Dumps processes probably shouldn't run with a SQL user that can alter any data, let alone drop tables.
    • Potentially, permissions should be much more fine-grained: separate users for maintenance scripts/crons; for testing ro queries; one for dumps generation; core vs extensions maintenance, etc.
    • Maybe DROP TABLE should not granted to anyone except DBAs?
    • Maybe "Database::dropTable" should check for production environment (similar to update.php) and fail if something like that happens.
  • sql.php must re-implement update.php's not-in-production checks if we're going to restore it.
  • Having both mwscript sql.php and sql (aka mwscript mysql.php) which run very different code paths is confusing. Fix this.
  • Change Wikidata cronjob to use mwscript mysql.php instead of mwscript sql.php: gerrit:587218
  • sql.php must not run LoadExtensionSchemaUpdates hooks: phab:T157651
  • Pursue restructuring Wikibase repository's SQL directory: phab:T205094
  • Consider adding paging based on ats-be-perceived 50X error rate, which considering recent history looks to be a good signal. TODO make a task
  • Wikibase schema updaters must not modify database directly phab:T249598

TODO: Add the #Wikimedia-Incident Phabricator tag to these tasks and move them to the "Follow-up" column.