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

Incident documentation/2018-04-10 Deleting a page on enwiki

From Wikitech-static
Jump to navigation Jump to search


Summary

An alter table running on the archive table on the English Wikipedia master database server caused deletion of pages to fail for en.wikipedia.org for ~40 minutes.

It was originally reported at: task T191875 According to logtash 85 errors happened during that time

Timeline (IN UTC)

  • 5:20 an ALTER table on enwiki master (db1052) on the externallinks table was started
  • 8:20 an ALTER table on enwiki master (db1052) on the archive table was started

Example of a failure:

Read timeout is reached (10.64.16.77) INSERT INTO archive (ar_namespace,ar_title,ar_timestamp,ar_minor_edit,ar_rev_id,ar_parent_id,ar_text_id,ar_text,ar_flags,ar_len,ar_page_id,ar_deleted,ar_sha1,ar_comment,ar_comment_id,ar_user,ar_user_text,ar_content_model,ar_content_format) VALUES ('14','Articles_needing_sections_from_August_2015','xx','0','xx','0','xx',,,'xx','xx','0','xx','Creating monthly dated maintenance category for current month','xx','xx','xx',NULL,NULL)

  • 9:19: Alter table finished and errors are gone

A total of 85 errors happened between 08:20 and 09:19

Conclusions

The following ALTER table caused issues on db1052 (enwiki master) when deleting (and probably when moving) a page but not on the other 6 masters it was executed before: SET SESSION innodb_lock_wait_timeout=1; SET SESSION lock_wait_timeout=30; ALTER TABLE archive MODIFY COLUMN ar_text mediumblob NULL, MODIFY COLUMN ar_flags tinyblob NULL;

There were a total of 85 errors during the time of the incident.

It was also noticed that the master had ongoing queries (task T191875#4119715) for the archive table, that might have also contributed to this issue. Whether it is the cause or the consequence is still unknown.

As described at: task T191875#4119820 this ALTER is fully online (and has not caused issues anywhere else)

Actionables

We are still investigating why this has been caused, but so far everything looks like a race condition.

The following task has also been created to try to check if the new queries may be creating additional contention (although it is not yet clear this is the root/only cause of the incident)

  • Reduce locking contention on deletion of pages (phab:T191892)