Ouch, only 3 are current. It may be difficult to recover the rest cleanly. :(

mysql> select page_id,page_latest from page,revision where page_latest=rev_id and rev_page=0;
+---------+-------------+

+---------+-------------+

+---------+-------------+
3 rows in set (0.13 sec)

On sourceswiki, there are now 245 revisions with rev_page=0; the extra one has a date stamp of 20081627
On enws there are two revisions with rev_page=0, both with datestamp of 20031218
For comparison, on enwp there are 709 revisions with rev_page=0

There are 318 orphaned revisions on mediawikiwiki. This was the cause of bug 43625.

Looks like new revs are still getting assigned to page_id 0 somehow.

Bumping severity per comment 4

Within the past two years, sorted by latest timestamp:

max(rev_timestamp) - count(*) - dbname
20150314151620 - 1300 - knwiki
20150123102603 - 89 - orwiki
20150109170449 - 9 - enwikivoyage
20141109142220 - 36 - fawikivoyage
20141016100102 - 502 - maiwiki
20140923170000 - 5 - idwiktionary
20140609215946 - 403 - metawiki
20140430194021 - 290 - hewikivoyage
20140421205055 - 420 - itwikisource
20131222123349 - 32 - zhwikivoyage
20131124123227 - 99 - tewiki
20131030141235 - 412 - nowiki
20131013172657 - 2104 - dewiki
20130729220133 - 2470 - mlwiki
20130723065815 - 6 - enwiktionary
20130718033244 - 26 - viwikivoyage
20130506140648 - 128 - simplewiki
20130329211512 - 26 - testwiki
20130329012222 - 4 - glwiktionary

Found while looking into T111605: many of those revisions have rev_user = 0 and a rev_user_text which is not an IP address, hence they were likely imported.

MariaDB [dewiki_p]> select count(rev_id) from revision where rev_page = 0;
+---------------+
| count(rev_id) |
+---------------+
|          2104 |
+---------------+
1 row in set (0.00 sec)

MariaDB [dewiki_p]> select count(rev_id) from revision where rev_page = 0 and rev_user = 0;
+---------------+
| count(rev_id) |
+---------------+
|           908 |
+---------------+
1 row in set (0.01 sec)

MariaDB [dewiki_p]> select count(rev_id) from revision where rev_page = 0 and rev_user = 0 AND rev_user_text RLIKE "[[:alpha:]]{3}";
+---------------+
| count(rev_id) |
+---------------+
|           440 |
+---------------+
1 row in set (0.01 sec)

Many of the revisions on knwiki with rev_page = 0 occurred when a user ran a simultaneous transwiki and upload import of the same pages. The transwiki import was of all revisions, while the upload import contained only the top revision. Several revisions during this import got imported with rev_page = 0. See the log. Obviously running two simultaneous imports of the same pages is not recommended.

On enwikivoyage, one of the affected pages was imported triply. See the log.

Indeed, I just reproduced this on testwiki (revisions 259876 and 259877) by clicking the submit button on Special:Import 3 times in a row.

A slightly different thing happened on orwiki. See the log. In this case, the import was working fine, until all of a sudden the pages started to be imported twice. I don't think we can blame the user for this.

I should note that the timestamps that @Krenair listed out are misleading. Since these edits are imported, the timestamp reflects the time the edit was made on the source wiki, not the time that they were imported with rev_page = 0.

As for fixing the problem on WMF wikis, I would favour just deleting these rows from the DB. There's no point trying to reconnect them up with the pages they belong to.

Change 262078 had a related patch set uploaded (by TTO):
Import: Try to stop revisions getting created with rev_page = 0

https://gerrit.wikimedia.org/r/262078

Change 262078 merged by jenkins-bot:
Import: Try to stop revisions getting created with rev_page = 0

https://gerrit.wikimedia.org/r/262078

As for fixing the problem on WMF wikis, I would favour just deleting these rows from the DB. There's no point trying to reconnect them up with the pages they belong to.

How many are they? Do we need a maintenance script and if so is there one already?

Change 263604 had a related patch set uploaded (by TTO):
Prevent revisions with rev_page = 0 from being inserted into the DB

https://gerrit.wikimedia.org/r/263604

In T18674#1915951, @Nemo_bis wrote:

As for fixing the problem on WMF wikis, I would favour just deleting these rows from the DB. There's no point trying to reconnect them up with the pages they belong to.

How many are they? Do we need a maintenance script and if so is there one already?

We can use runBatchedQuery.php to delete the rows.

Please make sure I am near before running a script that deletes rows so we have proper backups/reversion process. 0:-)

Change 263604 merged by jenkins-bot:
Prevent revisions with rev_page = 0 from being inserted into the DB

https://gerrit.wikimedia.org/r/263604