wp_get_post_revisions returns incorrect order

[kovshenin]

When a post has multiple revisions and each revision has the exact same timestamp, wp_get_post_revisions returns revisions in reversed order, causing wp_save_post_revision to ​make a bad choice for $last_revision (which is actually the first revision) and produce an extra revision, even if it has not been changed.

In a normal environment this doesn't happen, unless you run wp_update_post several times in a row in under one second, so this is an edge case. This edge case causes our ​revisions.php tests to fail randomly, depending on whether the updates happened with the same timestamp or not.

To reproduce, try running the revisions unit tests several times in a row.

[kovshenin]

If you're having trouble reproducing this, try running 23923.reproduce.diff​ with --group 23923.

An easy fix would be to order by date and ID instead of date only in wp_get_post_revisions, but we would need something that explicitly sets descending ordering on both columns. #17065 might have a solution.

[nacin]

Could we also just fetch them ASC then flip them?

[adamsilverstein]

thanks for identifying this bug! i had noticed the intermittent unit test failures in revisions, but never figured out what was causing it. your explanation makes perfect sense.

what about just ordering by ID?

[azaozz]

Yeah, thinking we should be ordering only by ID. This is the best field to determine the "real" order for revisions.

Edit: scratch that, revisions cannot be edited, so ordering by ID will always match ordering by date. However autosaves are edited/overwritten. @kovshenin's original suggestion to order by both date and ID would work properly.

[SergeyBiryukov]

#26042 has more traction, let's move the unit test there.