MediaWiki r103763 - Code Review

I guess there's an issue (I didn't really test it):

In WMF branch:

• zh-hans
  • msg => AA
• zh-cn
  • msg => B

In trunk:

• zh-hans
  • msg => A
• zh-cn
  • msg => B

So in l10ncache:

• zh-hans
  • msg => A
• zh-cn
  • (nothing)

onRecache for zh-cn, input:

• msg => B

after 1st round $codeSequence merge of zh-hans:

• msg => A

after 2nd round $codeSequence merge of zh-cn:

• msg => A

output:

• msg => A

expected:

• msg => B

Yes, that is true. I have at the moment no idea, why.

But now it works, when there is no "zh-cn" message in WMF branch and no "zh-cn" message in trunk. Before this commit, the fallback was never triggered. (msg was than AA, not A). For this case the order is not relevant.

Hmm, yeah, that's a problem. Good point.

The cache entry for zh-hans overrides the message, but it is hard to fix, because for the one usecase there should not a merge, but in the other usecase it needs a merge.

Some more notes for myself:

There are three cases where LU will store nothing for a message:

1. Message is not defined in the trunk language file. This is the case in which we want to use the fallback
2. Translation in the trunk language file is identical to the message in the WMF branch. Here we don't want to use the fallback
3. English version differs between trunk and WMF. In this case the message won't be stored in any language, so whether or not we use the fallback doesn't matter

For case #2, I think we'll be fine if we store the translation in the LU cache file even though it's not needed, but I'll have to investigate that in more detail.

I've done #2 in r104041, so this issue should now be fixed.

I meant to say I've done this (storing unchanged messages).

I guess it's still buggy:

WMF:

• en
  • msg => A
• de
  • msg => B

trunk:

• en
  • msg => A
• de
  • (nothing)

input onRecache de:

• msg => B

1st round fallback, en, skipping:

• msg => B

2nd round fallback, de, no msg cached:

• msg => B

output:

• msg => B

expected (fallback to en):

• msg => A

Can we cache English messages as well?

Status on this?

Seems partly fixed and Roan decided to deploy the partly fixed version live.

changing back to "NEW". If things still need fixing, then restore "FIXME" and give a list of what needs to be fixed still.

If a de message was in wmf-branch and isn't in trunk (and developers expect it to fall back to en), it cannot be updated by LU.

Some ways to fix:

• Iterate over all keys in English where English text is not changed when updating (= building cache), then calculate (resolve fallback now) its message text in every language and store it in cache. When recaching the message, simply merging cached data into main data.
• or, include English in l10ncache.

This is a bug, yes, but it's not a regression, because fallbacks weren't handled at all before. Can we file a bug about the missing message behavior and unfixme this revision?

Done. bug 34272.