Localization

Kitsune is localized with gettext. User-facing strings in the code or templates need to be marked for gettext localization.

We use Pontoon to provide an easy interface to localizing these files. Localizers are also free to download the PO files and use whatever tool they are comfortable with.

Making Strings Localizable

Making strings in templates localizable is exceptionally easy. Making strings in Python localizable is a little more complicated. The short answer, though, is just wrap the string in _().

Interpolation

A string is often a combination of a fixed string and something changing, for example, Welcome, James is a combination of the fixed part Welcome,, and the changing part James. The naive solution is to localize the first part and the follow it with the name:

_('Welcome, ') + username

This is wrong!

In some locales, the word order may be different. Use Python string formatting to interpolate the changing part into the string:

_('Welcome, {name}').format(name=username)

Python gives you a lot of ways to interpolate strings. The best way is to use Py3k formatting and kwargs. That’s the clearest for localizers.

The worst way is to use %(label)s, as localizers seem to have all manner of trouble with it. Options like %s and {0} are somewhere in the middle, and generally OK if it’s clear from context what they will be.

Localization Comments

Sometimes, it can help localizers to describe where a string comes from, particularly if it can be difficult to find in the interface, or is not very self-descriptive (e.g. very short strings). If you immediately precede the string with a comment that starts L10n:, the comment will be added to the PO file, and visible to localizers.

Example:

rev_data.append({
            'x': 1000 * int(time.mktime(rdate.timetuple())),
            # L10n: 'R' is the first letter of "Revision".
            'title': _('R', 'revision_heading'),
            'text': str(_('Revision %s')) % rev.created
            #'url': 'http://www.google.com/'  # Not supported yet
        })

Adding Context with msgctxt

Strings may be the same in English, but different in other languages. English, for example, has no grammatical gender, and sometimes the noun and verb forms of a word are identical.

To make it possible to localize these correctly, we can add “context” (known in gettext as “msgctxt”) to differentiate two otherwise identical strings.

For example, the string “Search” may be a noun or a verb in English. In a heading, it may be considered a noun, but on a button, it may be a verb. It’s appropriate to add a context (like “button”) to one of them.

Generally, we should only add context if we are sure the strings aren’t used in the same way, or if localizers ask us to.

Example:

from tower import ugettext as _

...

foo = _('Search', context='text for the search button on the form')

Plurals

“You have 1 new messages” grates on discerning ears. Fortunately, gettext gives us a way to fix that in English and other locales, the ngettext function:

ngettext('singular', 'plural', count)

A more realistic example might be:

ngettext('Found {count} result.',
         'Found {count} results',
         len(results)).format(count=len(results))

This method takes three arguments because English only needs three, i.e., zero is considered “plural” for English. Other locales may have different plural rules, and require different phrases for, say 0, 1, 2-3, 4-10, >10. That’s absolutely fine, and gettext makes it possible.

Strings in HTML Templates

When putting new text into a template, all you need to do is wrap it in a _() call:

<h1>{{ _('Heading') }}</h1>

Adding context is easy, too:

<h1>{{ _('Heading', 'context') }}</h1>

L10n comments need to be Jinja2 comments:

{# L10n: Describes this heading #}
<h1>{{ _('Heading') }}</h1>

Note that Jinja2 escapes all content output through {{ }} by default. To put HTML in a string, you’ll need to add the |safe filter:

<h1>{{ _('Firefox <span>Help</span>')|safe }}</h1>

To interpolate, you should use one of two Jinja2 filters: |f() or, in some cases, |fe(). |f() has exactly the same arguments as u''.format():

{{ _('Welcome, {name}!')|f(name=request.user.username) }}

The |fe() is exactly like the |f() filter, but escapes its arguments before interpolating, then returns a “safe” object. Use it when the localized string contains HTML:

{{ _('Found <strong>{0}</strong> results.')|fe(num_results) }}

Note that you do not need to use |safe with |fe(). Also note that while it may look similar, the following is not safe:

{{ _('Found <strong>{0}</strong> results.')|f(num_results)|safe }}

The ngettext function is also available:

{{ ngettext('Found {0} result.',
            'Found {0} results.',
            num_results)|f(num_results) }}

Using {% trans %} Blocks for Long Strings

When a string is very long, i.e. long enough to make Github scroll sideways, it should be line-broken and put in a {% trans %} block. {% trans %} blocks work like other block-level tags in Jinja2, except they cannot have other tags, except strings, inside them.

The only thing that should be inside a {% trans %} block is printing a string with {{ string }}. These are defined in the opening {% trans %} tag:

{% trans user=request.user.username %}
    Thanks for registering, {{ user }}! We're so...
    hope that you'll...
{% trans %}

You can also provide comments:

{# L10n: User is a username #}
{% trans user=request.user.username %}
    Thanks for registering, {{ user }}! We're so...
    hope that you'll...
{% trans %}

Strings in Python

Note

Whenever you are adding a string in Python, ask yourself if it really needs to be there, or if it should be in the template. Keep logic and presentation separate!

Strings in Python are more complex for two reasons:

  1. We need to make sure we’re always using Unicode strings and the Unicode-friendly versions of the functions.

  2. If you use the ugettext function in the wrong place, the string may end up in the wrong locale!

Here’s how you might localize a string in a view:

from tower import ugettext as _

def my_view(request):
    if request.user.is_superuser:
        msg = _(u'Oh hi, staff!')
    else:
        msg = _(u'You are not staff!')

Interpolation is done through normal Python string formatting:

msg = _(u'Oh, hi, {user}').format(user=request.user.username)

ugettext supports context, too:

msg = _('Search', 'context')

L10n comments are normal one-line Python comments:

# L10n: A message to users.
msg = _(u'Oh, hi there!')

If you need to use plurals, import the function ungettext from Tower:

from tower import ungettext, ugettext as _

n = len(results)
msg = ungettext('Found {0} result', 'Found {0} results', n).format(n)

Lazily Translated Strings

You can use ugettext or ungettext only in views or functions called from views. If the function will be evaluated when the module is loaded, then the string may end up in English or the locale of the last request! (We’re tracking down that issue.)

Examples include strings in module-level code, arguments to functions in class definitions, strings in functions called from outside the context of a view. To localize these strings, you need to use the _lazy versions of the above methods, ugettext_lazy and ungettext_lazy. The result doesn’t get translated until it is evaluated as a string, for example by being output or passed to str():

from tower import ugettext_lazy as _lazy

PAGE_TITLE = _lazy(u'Page Title')

ugettext_lazy also supports context.

It is very important to pass Unicode objects to the _lazy versions of these functions. Failure to do so results in significant issues when they are evaluated as strings.

If you need to work with a lazily-translated string, you’ll first need to convert it to a str object:

from tower import ugettext_lazy as _lazy

WELCOME = _lazy(u'Welcome, %s')

def my_view(request):
    # Fails:
    WELCOME % request.user.username

    # Works:
    str(WELCOME) % request.user.username

Strings in the Database

There is some user generated content that needs to be localizable. For example, karma titles can be created in the admin site and need to be localized when displayed to users. A django management command is used for this. The first step to making a model’s field localizable is adding it to DB_LOCALIZE in settings.py:

DB_LOCALIZE = {
    'karma': {
        'Title': {
            'attrs': ['name'],
            'comments': ['This is a karma title.'],
        }
    },
    'appname': {
        'ModelName': {
            'attrs': ['field_name'],
            'comments': ['Optional comments for localizers.'],
        }
    }
}

Then, all you need to do is run the extract_db management command:

$ python manage.py extract_db

Be sure to have a recent database from production when running the command.

By default, this will write all the strings to kitsune/sumo/db_strings.py and they will get picked up during the normal string extraction (see below).

Strings in Email Templates

Currently, email templates are text-based and not in HTML. Because of that you should use this style guide:

  1. The entire email should be wrapped in autoescape. e.g.

    1
2
3
4
5
6
7
8
    {% autoescape false %}
{% trans %}
The entire email should be wrapped in autoescape.
{% endtrans %}


...
{% endautoescape %}

  2. After an {% endtrans %}, you need two blank lines (three carriage returns). The first is eaten by the tag. The other two show up in the email. e.g.

    1
2
3
4
5
6
7
    {% trans %}
To confirm your subscription, stand up, put your hands on
your hips and do the hokey pokey.
{% endtrans %}


{{ _('Thanks!') }}

    Produces this:

    1
2
3
4
    To confirm your subscription, stand up, put your hands on
your hips and do the hokey pokey.

Thanks!

  3. Putting in line breaks in a trans block doesn’t have an effect since trans blocks get gettexted and whitespace is collapsed.

Testing localized strings

When we add strings that need to be localized, it can take a couple of weeks for us to get translations of those localized strings. This makes it difficult to find localization issues.

Enter Dennis.

Run:

$ ./scripts/test_locales.sh

It’ll extract all the strings, create a .pot file, then create a Pirate translation of all strings. The Pirate strings are available in the xx locale. After running the test_locales.sh script, you can access the xx locale with:

Strings in the Pirate translation have the following properties:

  1. they are longer than the English string: helps us find layout and wrapping issues

  2. they have at least one unicode character: helps us find unicode issues

  3. they are easily discernable from the English versions: helps us find strings that aren’t translated

Note

The xx locale is only available on your local machine. It is not available on -dev, -stage, or -prod.

Linting localized strings

You can lint localized strings for warnings and errors:

$ dennis-cmd lint locale/

Or just errors:

$ dennis-cmd lint --errorsonly locale/

You can see help text:

$ dennis-cmd

Getting the Localizations

Localizations are not stored in this repository, but are in a separate Git repo:

You don’t need the localization files for general development. However, if you need them for something, they’re pretty easy to get:

$ cd kitsune
$ git clone https://github.com/mozilla-l10n/sumo-l10n locale

Updating the Localizations

When strings are added or updated, we need to update the templates and PO files for localizers. Updating strings is pretty easy. Check out the localizations as above, then:

$ python manage.py extract
$ python manage.py merge

Congratulations! You’ve now updated the POT and PO files.

Sometimes this can leave a bunch of garbage files with .po~ extensions. You should delete these, never commit them:

$ find . -name "*.po~" -delete

Adding a New Locale

Say you wanted to add fa-IR:

$ mkdir -p locale/fa-IR/LC_MESSAGES
$ python manage.py merge

Then add ‘fa-IR’ to SUMO_LANGUAGES in settings.py and make sure there is an entry in lib/languages.json (if not, add it).

And finally, add a migration with:

INSERT INTO `wiki_locale` (`locale`) VALUES ('fa-IR');

Done!

Compiling MO Files

gettext is so fast for localization because it doesn’t parse text files, it reads a binary format. You can easily compile that binary file from the PO files in the repository.

We don’t store MO files in the repository because they need to change every time the corresponding PO file changes, so it’s silly and not worth it. They are ignored by .gitignore, but please make sure you don’t forcibly add them to the repository.

There is a shell script to compile the MO files for you:

$ ./locale/compile-mo.sh locale

Done!

Why aren’t localized strings getting updated on prod?

We use Dennis to lint .po files for errors that cause HTTP 500 errors in production. Things like malformed variables, variables in the translated string that aren’t in the original and that sort of thing.

For example, this would cause the site to break:

#: kitsune/questions/templates/questions/includes/answer.html:19
msgid "{num} answers"
msgstr "{0} antwoorden"

In this example, the {0} is wrong.

Reporting errors in .po files

When we do a deployment to production, we dump all the Dennis output into:

https://support.mozilla.org/static/postatus.txt

We need to check that periodically and report the errors.

If there are errors in those files, we need to open up a bug in Mozilla Localizations -> locale code with the specifics.

Product:

Mozilla Localizations

Component:

The locale code for the language in question

Bug summary:

Use the error line

Bug description template:

We found errors in the translated strings for Mozilla Support
<https://support.mozilla.org/>. The errors are as follows:


<paste errors here>


Until these errors are fixed, we can't deploy updates to the
strings for this locale to production.

Mozilla Support strings can be fixed in the Support Mozilla project
in Pontoon <https://pontoon.mozilla.org/projects/sumo/>.

If you have any questions, let us know.