Dennis Retrospective (2013)

Note: This is an old post in a blog with a lot of posts. The world has changed, technologies have changed, and I've changed. It's likely this is out of date and not representative. Let me know if you think this is something that needs updating.



3 months

  • fixed l10n-related HTTP 500 errors for SUMO and Input

  • paved road for MDN, AMO and other Mozilla sites to use the same strategy

Problem statement

When we deploy (SUMO) and Input [1], it fetches the most recent localized strings .po files from SVN, compiles them to .mo files, and ships them with the site. Because SUMO supports many languages and the deploy pulls down the most recent changes, there's no way to effectively test the entire site for all languages before deploying to production. Because of that, users experience HTTP 500 server errors on pages that have bad strings.

When there are server errors, we get notified, write up a bug, and then have to go fix it and push the fix out as soon as possible. Fixing issues is difficult since we don't know most of the languages the site is translated in and our l10n community spans many timezones, so getting help can take many hours.

That was covered roughly in [bug 841412] for SUMO and [bug 875313] for Input.

On top of that, SUMO sends emails to SUMO users and if the localized strings are bad, then emails don't get sent. That's covered in [bug 850215].

Why are there problems in localizations? Judging by the strings we were seeing, we think we had a few issues:

  1. The localizer changes the formatting token.

    We're using gettext and Python and several different token formats. For example, %s, %(name)s, and {name}. If the localizer changes the structure of the token, that will cause a server error.

  2. The localizer translates the token.

    For example, the token is {name}, but the translated token is {nombre}.

  3. The localizer copies a string from somewhere else with different tokens.

    For example, the string has {url}, but the translated string has {helpurl}.

Thus we're faced with:

  1. a series of problems with strings that happen semi-frequently

  2. server errors preventing users from seeing certain parts of SUMO

  3. we can't suss out these errors by testing the whole site for every locale before every prod deploy

  4. each server error is a priority 1 interrupt

SUMO and Input weren't the only sites that has this problem---it's a problem for MDN, AMO, and other sites, too.


I wondered if we could effectively lint .po files during deploy and only compile the .po files that didn't have problems. If we did that, then the problem would stop.

I wrote a script in SUMO [2] that let me experiment with writing a linter and tying it to the deploy process.

That went well and I wanted to use the same system for two different projects plus I suspected others would want to use it, too. Further, we had a script that let us debug layout problems related to translated strings and I wanted to merge both of these into a new library.

I created Dennis that let you lint and fake-translate strings.

On July 29th, I released Dennis v0.3.3 which I felt was good enough for us and other people to use. We switched SUMO and Input to use it.

Our server errors from localization issues mostly ended. Periodically, we'd encounter a new kind of issue and would improve Dennis to catch it.

We made the deploy logs viewable so linting errors could be seen, bugs could be written, and errors could get fixed.

I talked about how we use it on Input and SUMO along with the bash script we used that compiled .po files that linted successfully to .mo files. MDN, AMO, and other sites switched to using Dennis thus eliminating the errors for them, too.


This eliminated a frequent cause of HTTP 500 errors which caused downtime for the site, prevented users from getting Firefox support, and created frequent interruptions for site engineers.

Building it as a library allowed other Mozilla sites to use it eliminating the problem for them, too.



The potools project that had a linter, but it:

  1. didn't return a non-zero exit code when it encountered linting errors

  2. performed a quality-of-string kind of linting and didn't really know about variables

We couldn't go with this as is and it looked too difficult to improve to meet our needs.

Want to comment? Send an email to willkg at bluesock dot org. Include the url for the blog entry in your comment so I have some context as to what you're talking about.