See below for an extract of the discussion on the recurring disappearance of interface messages recently. It was a mistake for the discussion to unfold on an internal list, but it happened quite by chance, starting with an incident report and developing from there.

---
Ori Livneh
ori@wikimedia.org


---------- Forwarded message ----------
From: Ori Livneh <ori@wikimedia.org>
Date: Thu, Apr 10, 2014 at 1:23 AM
Subject: Re: [Engineering] Localisation not working on MediaWiki.org
To: "Brad Jorsch (Anomie)" <bjorsch@wikimedia.org>
Cc: Bryan Davis <bd808@wikimedia.org>, Development and Operations Engineers <engineering@lists.wikimedia.org>


On Tue, Apr 8, 2014 at 6:56 AM, Brad Jorsch (Anomie) <bjorsch@wikimedia.org> wrote:
On Mon, Apr 7, 2014 at 9:37 PM, Bryan Davis <bd808@wikimedia.org> wrote:
The obvious change that caused this was that `mwversionsinuse
--withdb` changed from returning "1.23wmf21=testwiki" to
"1.23wmf21=test2wiki". This result is used within scap by the
mw-update-l10n script to run the maintenance script that builds the
ExtensionMessages file. In theory the exact wiki passed to `mwscript
mergeMessageFileList.php --wiki=<WIKIDB>` shouldn't matter, but
obviously there are now some circumstances where it does indeed
matter.

It looks to me like it has always mattered to an extent: the final result from maintenance/mergeMessageFileList.php is the combination of extensions loaded for the --wiki wiki (e.g. in CommonSettings.php) and the extensions loaded by the script itself from the passed list of extensions. Hopefully the latter is always a superset of the former so that turns out not to matter.

Interface messages went missing again on wikidata.org. l10nupdate ran updates on cawikibooks, where $wmgUseWikibaseClient is false. The theory that the exact wiki shouldn't make a difference is pretty shaky. You should expect to run on testwiki and fail loudly if you can't.

We should rethink our whole approach; I don't have any confidence in the architecture. What is especially damning is not so much the recurrence of failures as the way they were discovered (that is to say: by chance) and the hard time we have had reasoning about their cause and the state of localization on the cluster generally.