path: root/doc/plugins/contrib/po.mdwn

I've been working on a plugin called "po", that adds support for multi-lingual wikis,
translated with gettext, using po4a.
More information:

It can be found in my "po" branch:
git clone git://gaffer.ptitcanardnoir.org/ikiwiki.git
It is self-contained, i.e. it does not modify ikiwiki core at all.
It is documented (including TODO and plans for next work steps) in
doc/plugins/po.mdwn, which can be found in the same branch.
No public demo site is available so far, I'm working on this.

My plan is to get this plugin clean enough to be included in ikiwiki.
The current version is a proof-of-concept, mature enough for me to dare submitting it here,
but I'm prepared to hear various helpful remarks, and to rewrite parts of it as needed.
Any thoughts on this?

Well, I think it's pretty stunning what you've done here. Seems very
complete and well thought out. I have not read the code in great detail
yet.
Just using po files is an approach I've never seen tried with a wiki. I
suspect it will work better for some wikis than others. For wikis that
just want translations that match the master language as closely as
possible and don't wander off and diverge, it seems perfect. (But what happens
if someone edits the Discussion page of a translated page?)
Please keep me posted, when you get closer to having all issues solved
and ready for merging I can do a review and hopefully help with the
security items you listed. --[[Joey]]



Thanks a lot for your quick review, it's reassuring to hear such nice words
from you. I did not want to design and write a full translation system, when
tools such as gettext/po4a already have all the needed functionality, for cases
where the master/slave languages paradigm fits.
Integrating these tools into ikiwiki plugin system was a pleasure.
I'll tell you when I'm ready for merging, but in the meantime,
I'd like you to review the changes I did to the core (3 added hooks).
Can you please do this? If not, I'll go on and hope I'm not going to far in
the wrong direction.

Sure.. I'm not completly happy with any of the hooks since they're very
special purpose, and also since run_hooks is not the best interface
for a hook that modifies a variable, where only the last hook run will
actually do anything. It might be better to just wrap
targetpage, bestlink, and beautify_urlpath. But, I noticed
the other day that such wrappers around exported functions are only visible by
plugins loaded after the plugin that defines them.
Update: Take a look at the new "Function overriding" section of
[[plugins/write]]. I think you can just inject wrappers about a few ikiwiki
functions, rather than adding hooks. The inject function is pretty
insane^Wlow level, but seems to work great. --[[Joey]]

Thanks a lot, it seems to be a nice interface for what I was trying to achieve.
I may be forced to wait two long weeks before I have a chance to confirm
this. Stay tuned. --[[intrigeri]]

I've updated the plugin to use inject. It is now fully self-contained,
and does not modify the core anymore. --[[intrigeri]]



The Discussion pages issue is something I am not sure about yet. But I will
probably decide that "slave" pages, being only translations, don't deserve
a discussion page: the discussion should happen in the language in which the
pages are written for real, which is the "master" one. --[[intrigeri]]
I think that's a good decision, you don't want to translate discussion,
and if the discussion page turns out multilingual, well, se la vi. ;-)
Relatedly, what happens if a translated page has a broken link, and you
click on it to edit it? Seems you'd first have to create a master page
and could only then translate it, right? I wonder if this will be clear
though to the user.

Right: a broken link points to the URL that allows to create
a page that can either be a new master page or a non-translatable
page, depending on po_translatable_pages value. The best
solution I can thing of is to use [[plugins/edittemplate]] to
insert something like "Warning: this is a master page, that must
be written in $MASTER_LANGUAGE" into newly created master pages,
and maybe another warning message on newly created
non-translatable pages. It seems quite doable to me, but in order
to avoid breaking existing functionality, it implies to hack a bit
[[plugins/edittemplate]] so that multiple templates can be
inserted at page creation time. [[--intrigeri]]

I implemented such a warning using the formbuilder_setup hook.
--[[intrigeri]]


And also, is there any way to start a translation of a page into a new
lanauge using the web interface?

When a new language is added to po_slave_languages, a rebuild is
triggered, and all missing PO files are created and checked into
VCS. An unpriviledged wiki user can not add a new language to
po_slave_languages, though. One could think of adding the needed
interface to translate a page into a yet-unsupported slave
language, and this would automagically add this new language to
po_slave_languages. It would probably be useful in some
usecases, but I'm not comfortable with letting unpriviledged wiki
users change the wiki configuration as a side effect of their
actions; if this were to be implemented, special care would be
needed. [[--intrigeri]]

Actually I meant into any of the currently supported languages.
I guess that if the template modification is made, it will list those
languages on the page, and if a translation to a language is missing,
the link will allow creating it?

Any translation page always exist for every supported slave
language, even if no string at all have been translated yet.
This implies the po plugin is especially friendly to people who
prefer reading in their native language if available, but don't
mind reading in English else.
While I'm at it, there is a remaining issue that needs to be
sorted out: how painful it could be for non-English speakers
(assuming the master language is English) to be perfectly able
to navigate between translation pages supposed to be written in
their own language, when their translation level is most
often low.
(It is currently easy to display this status on the translation
page itself, but then it's too late, and how frustrating to load
a page just to realize it's actually not translated enough for
you. The "other languages" loop also allows displaying this
information, but it is generally not the primary
navigation tool.)
IMHO, this is actually a social problem (i.e. it's no use adding
a language to the supported slave ones if you don't have the
manpower to actually do the translations), that can't be fully
solved by technical solutions, but I can think of some hacks
that would limit the negative impact: a given translation's
status (currently = percent translated) could be displayed next
to the link that leads to it; a color code could as well be used
("just" a matter of adding a CSS id or class to the links,
depending on this variable). As there is already work to be done
to have the links text generation more customizable through
plugins, I could do both at the same time if we consider this
matter to be important enough. --[[intrigeri]]

The translation status in links is now implemented in my
pobranch. It requires my meta branch changes to
work, though. I consider the latter to be mature enough to
be merged. --[[intrigeri]]








FWIW, I'm tracking your po branch in ikiwiki master git in the po
branch. One thing I'd like to try in there is setting up a translated
basewiki, which seems like it should be pretty easy to do, and would be
a great demo! --[[Joey]]

I have a complete translation of basewiki into danish, and am working with
others on preparing one in german.  For a complete translated user
experience, however, you will also need templates translated (there are a few
translatable strings there too).  My not-yet-merged po4a Markdown improvements
(see bug#530574) correctly handles multiple
files in a single PO which might be relevant for template translation handling.
--[[JonasSmedegaard]]


I've merged your changes into my own branch, and made great
progress on the various todo items. Please note my repository
location has changed a few days ago, my user page was updated
accordingly, but I forgot to update this page at the same time.
Hoping it's not too complicated to relocated an existing remote...
(never done that, I'm a Git beginner as well as a Perl
newbie) --[[intrigeri]]

Just a matter of editing .git/config, thanks for the heads up.

Joey, please have a look at my branch, your help would be really
welcome for the security research, as I'm almost done with what
I am able to do myself in this area. --[[intrigeri]]

I came up with a patch for the WrapI18N issue --[[Joey]]






I've set this plugin development aside for a while. I will be back and
finish it at some point in the first quarter of 2009. --[[intrigeri]]

Abstract: Joey, please have a look at my po and meta branches.
Detailed progress report:

it seems the po branch in your repository has not been tracking my
own po branch for two months. any config issue?
all the plugin's todo items have been completed, robustness tests
done
I've finished the detailed security audit, and the fix for po4a
bugs has entered upstream CVS last week
I've merged your new checkcontent hook with the cansave hook
I previously introduced in my own branch; blogspam plugin updated
accordingly
the rename hook changes we discussed elsewhere are also part of my
branch
I've introduced two new hooks (canremove and canrename), not
a big deal; IMHO, they extend quite logically the plugin interface
as highlighted on [[bugs/pagetitle_function_does_not_respect_meta_titles]],
my meta branch contains a new feature that is really useful in a
translatable wiki

As a conclusion, I'm feeling that my branches are ready to be
merged; only thing missing, I guess, are a bit of discussion and
subsequent adjustments.
--[[intrigeri]]


I've looked it over and updated my branch with some (untested)
changes.

I've merged your changes into my branch. Only one was buggy.

Sorry, I'd forgotten about your cansave hook.. sorry for the duplicate
work there.
Reviewing the changes, mostly outside of po.pm, I have
the following issues.

renamepage to renamelink change would break the ikiwiki
3.x API, which I've promised not to do, so needs to be avoided
somehow. (Sorry, I guess I dropped the ball on not getting this
API change in before cutting 3.0..)


Fixed, see [[todo/need_global_renamepage_hook]].


I don't understand the parentlinks code change and need to figure it
out. Can you explain what is going on there?


I'm calling bestlink there so that po's injected bestlink is
run. This way, the parent links of a page link to the parent page
version in the proper language, depending on the
po_link_to=current and po_link_to=negotiated settings.
Moreover, when using my meta branch enhancements plus meta title to
make pages titles translatable, this small patch is needed to get
the translated titles into parentlinks.


canrename's mix of positional and named parameters is way too
ugly to get into an ikiwiki API. Use named parameters
entirely. Also probably should just use named parameters
for canremove.
skeleton.pm.example's canrename needs fixing to use either
the current or my suggested parameters.


Done.


I don't like the exporting of %backlinks and $backlinks_calculated
(the latter is exported but not used).


The commit message for 85f865b5d98e0122934d11e3f3eb6703e4f4c620
contains the rationale for this change. I guess I don't understand
the subtleties of our use, and perldoc does not help me a lot.
IIRC, I actually did not use our to "export" these variables, but
rather to have them shared between Render.pm uses.

My wording was unclear, I meant exposing. --[[Joey]]

I guess I still don't know Perl's our enough to understand clearly.
No matter whether these variables are declared with my or our,
any plugin can use IkiWiki::Render and then access
$IkiWiki::backlinks, as already does e.g. the pagestat plugin.
So I guess your problem is not with letting plugins use these
variables, but with them being visible for every piece of
(possibly external) code called from Render.pm. Am I right?
If I understand clearly, using a brace block to lexically enclose
these two our declarations, alongside with the calculate_backlinks
and backlinks subs definitions, would be a proper solution, wouldn't
it? --[[intrigeri]]

No, %backlinks and the backlinks() function are not the same thing.
The variable is lexically scoped; only accessible from inside
Render.pm --[[Joey]]





What is this IkiWiki::nicepagetitle and why are you
injecting it into that namespace when only your module uses it?
Actually, I can't even find a caller of it in your module.


I guess you should have a look to my meta branch and to
[[bugs/pagetitle_function_does_not_respect_meta_titles]] in order
to understand this :)

It would probably be good if I could merge this branch without
having to worry about also immediatly merging that one. --[[Joey]]

I removed all dependencies on my meta branch from the po one.
This implied removing the po_translation_status_in_links and
po_strictly_refresh_backlinks features, and every link text is now
displayed in the master language. I believe the removed features really
enhance user experience of a translatable wiki, that's why I was
initially supposing the meta branch would be merged first.
IMHO, we'll need to come back to this quite soon after po is merged.
--[[intrigeri]]
Maybe you should keep those features in a meta-po branch?
I did a cursory review of your meta last night, have some issues with it,
but this page isn't the place for a detailed review. --[[Joey]]

Done. --[[intrigeri]]





I'm very fearful of the add_depends in postscan.
Does this make every page depend on every page that links
to it? Won't this absurdly bloat the dependency pagespecs
and slow everything down? And since nicepagetitle is given
as the reason for doing it, and nicepagetitle isn't used,
why do it?


As explained in the 85f865b5d98e0122934d11e3f3eb6703e4f4c620 log:
this feature hits performance a bit. Its cost was quite small in my
real-world use-cases (a few percents bigger refresh time), but
could be bigger in worst cases. When using the po plugin with my
meta branch changes (i.e. the nicepagetitle thing), and having
enabled the option to display translation status in links, this
maintains the translation status up-to-date in backlinks. Same when
using meta title to make the pages titles translatable. It does
help having a nice and consistent translated wiki, but as it can
also involve problems, I just turned it into an option.

This has been completely removed for now due to the removal of
the dependency on my meta branch. --[[intrigeri]]



The po4a Suggests should be versioned to the first version
that can be used safely, and that version documented in
plugins/po.mdwn.


Done.
--[[intrigeri]]

--[[Joey]]

I reverted the %backlinks and $backlinks_calculated exposing.
The issue they were solving probably will arise again when I'll work
on my meta branch again (i.e. when the simplified po one is merged),
but the po thing is supposed to work without these ugly our.
Seems like it was the last unaddressed item from Joey's review, so I'm
daring a timid "please pull"... or rather, please review again :)
--[[intrigeri]]

Ok, I've reviewed and merged into my own po branch. It's looking very
mergeable.

Is it worth trying to fix compatability with indexpages?


Supporting usedirs being enabled or disabled was already quite
hard IIRC, so supporting all four combinations of usedirs and
indexpages settings will probably be painful. I propose we forget
about it until someone reports he/she badly needs it, and then
we'll see what can be done.


Would it make sense to go ahead and modify page.tmpl to use
OTHERLANGUAGES and PERCENTTRANSLATED, instead of documenting how to modify it?


Done in my branch.


Would it be better to disable po support for pages that use unsupported
or poorly-supported markup languages?


I prefer keeping it enabled, as:

most wiki markups "almost work"
when someone needs one of these to be fully supported, it's not
that hard to add dedicated support for it to po4a; if it were
disabled, I fear the ones who could do this would maybe think
it's blandly impossible and give up.





What's the reasoning behind checking that the link plugin
is enabled? AFAICS, the same code in the scan hook should
also work when other link plugins like camelcase are used.


That's right, fixed.


In pagetemplate there is a comment that claims the code
relies on genpage, but I don't see how it does; it seems
to always add a discussion link?


It relies on IkiWiki::Render's genpage as this function sets the
discussionlink template param iff it considers a discussion link
should appear on the current page. That's why I'm testing
$template->param('discussionlink').

Maybe I was really wondering why it says it could lead to a broken
link if the cgiurl is disabled. I think I see why now: Discussionlink
will be set to a link to an existing disucssion page, even if cgi is
disabled -- but there's no guarantee of a translated discussion page
existing in that case. However, htmllink actually checks
for this case, and will avoid generating a broken link so AFAICS, the
comment is actually innacurate.. what will really happen in this case
is discussionlink will be set to a non-link translation of
"discussion". Also, I consider $config{cgi} and %links (etc)
documented parts of the plugin interface, which won't change; po could
rely on them to avoid this minor problem. --[[Joey]]



Is there any real reason not to allow removing a translation?
I'm imagining a spammy translation, which an admin might not
be able to fix, but could remove.


On the other hand, allowing one to "remove" a translation would
probably lead to misunderstandings, as such a "removed" translation
page would appear back as soon as it is "removed" (with no strings
translated, though). I think an admin would be in a position to
delete the spammy .po file by hand using whatever VCS is in use.
Not that I'd really care, but I am slightly in favour of the way
it currently works.

That would definitly be confusing. It sounds to me like if we end up
needing to allow web-based deletion of spammy translations, it will
need improvements to the deletion UI to de-confuse that. It's fine to
put that off until needed --[[Joey]]




Re the meta title escaping issue worked around by change.
I suppose this does not only affect meta, but other things
at scan time too. Also, handling it only on rebuild feels
suspicious -- a refresh could involve changes to multiple
pages and trigger the same problem, I think. Also, exposing
this rebuild to the user seems really ugly, not confidence inducing.
So I wonder if there's a better way. Such as making po, at scan time,
re-run the scan hooks, passing them modified content (either converted
from po to mdwn or with the escaped stuff cheaply de-escaped). (Of
course the scan hook would need to avoid calling itself!)
(This doesn't need to block the merge, but I hope it can be addressed
eventually..)


--[[Joey]]

I'll think about it soon.
--[[intrigeri]]

Did you get a chance to? --[[Joey]]





As discussed at [[todo/l10n]] the templates needs to be translatable too.  They
should be treated properly by po4a using the markdown option - at least with my
later patches in bug#530574) applied.


It seems to me that the po plugin (and possibly other parts of ikiwiki) wrongly
uses gettext.  As I understand it, gettext (as used currently in ikiwiki) always
lookup a single language, That might make sense for a single-language site, but
multilingual sites should emit all strings targeted at the web output in each own
language.
So generally the system language (used for e.g. compile warnings) should be separated
from both master language and slave languages.
Preferrably the gettext subroutine could be extended to pass locale as optional
secondary parameter overriding the default locale (for messages like "N/A" as
percentage in po plugin).  Alternatively (with above mentioned template support)
all such strings could be externalized as templates that can then be localized.