Languages refers to the different translations we support on docs.github.com. Currently, our supported languages include Spanish, Japanese, Portuguese, Chinese, Russian, French, Korean, and German.

1. Content is authored in English.
2. Periodically, translators grab newly authored English content and translate it.
3. Periodically, translators push newly translated content to language-specific private repos.
4. Every time we deploy to production, we checkout translations repos into a translations/ directory.
5. When the app starts up, we create identical trees for each language we support. The trees are made by following an index page's children frontmatter.
6. When a translated page is requested, we check for key rendering errors (e.g. corrupt Liquid syntax) and can fall back to the English equivalent if needed.

This document expands on the TLDR above in the following sections:

• Creation of translated content
• Deployment of translated content
• Serving translated content
• Management of translation-related assets
• Working with translated content locally

GitHub Docs uses a mix of machine and human translators to source the content.

Periodically, translators read the content/** and data/** directories from github/docs-internal and compare it with what's already translated. Any newly introduced English content that needs to be translated is then added to its translation repo. Files are never deleted from the translation repos.

During the build step of our deployment, we checkout every translation repo into the translations/ directory.

The enabled languages and their source directories are interpreted in src/languages/lib/languages.ts, which ensures English and translated content are in the same Docker image we deploy.

When the app starts up, we:

1. Create a tree of possible pages in English by following the children: property in each index.md file.
2. Create a translated tree by looping over the English tree and, for each file, we attempt to read the translated source using the page's relativePath, swapping the file containing the contents with the translated target. At this point, each node in the tree is a Page instance.
3. Transpose the data to req.context.pages which is a mapping object that has a key for every Page's permalinks to the Page instance.

Some notes about this approach:

• We do not read the children: frontmatter property in the translated content.
• The only frontmatter properties we use in translated files are: title, shortTitle, intro.
• We do not need to read translated content that doesn't have an English counterpart.

When a user requests a translated page, it's possible that the content contains errors, like invalid Liquid.

In these situations, we fall back to English content for the part that generated an error, not the whole page. For example, a Japanese translation might have a perfectly good title, shortTitle, and Markdown body, but the Liquid within the intro frontmatter property could be broken. In that case, you'll get a Japanese title, English intro, and Japanese body.

Details of how the error handling and logic for fallback works, see src/languages/lib/render-with-fallback.ts.

Translations are always delayed and, at best, are updated once a day. So we need the ability to serve content that's been newly introduced but has not yet been translated.

Suppose an author introduces a new English page (e.g. /en/my/new/page). As soon as that's in production, we should be able to load all translations (e.g., /es/my/new/page or /ru/my/new/page) but that page won't be translated yet.

Similarly, suppose an author deletes /en/some/old/page and sets it to redirect to /en/some/other/page. The translated pages must also redirect. To handle redirects, we record all redirect_from frontmatter entries from the English content, and map that in the translated tree, independent of the language. When a user requests a translated page, the handleRedirects middleware removes the language code from the URL, looks it up in the record of redirects, and when resolved injects the language code back into the URL.

To address the cases discussed in this section, we adopt the following methodology when serving translated pages:

1. If the English URL returns 200 OK, the translated URL also returns 200 OK. This works because we of the trees we create during startup, based entirely on English content.
2. If the English URL returns 301, the translated URL also returns 301.

The {% data ... %} liquid tag is an exception, which utilizes the data/ directory. We do not preload data/** files, they're looked up on-the-fly when rendering. It's highly likely these files do not exist in English or are very out of date.

For example, when rendering a translated page you might find a Liquid reference to a reusable, {% data reusables.saml.saml-accounts %}. That reference is looked up on-the-fly, which should resolve to data/reusables/saml/saml-accounts.md in the translation directory. If it doesn't exist, it will then use the English version of the reusable. If it does exist, it will attempt to render it and fall back to the English if there's a Liquid rendering error.

Alternatively, suppose the translated content below:

Detta exemplet är på Svenska för skojs
skull {% data variables.support.help %}.

If the translated data/variables/support.yml is corrupted, and we can't fall back to an equivalent English file because it's been removed or renamed. In this case, we display the empty string in its place; we don't surface the error.

The translation repos can only be read by us. We do not and cannot modify their contents. When errors in translations are found, we communicate them to the translators and wait for the fixes to appear in the translated repos.

So if a translation urgently needs to be updated and we can't wait for the usual translation process, the best solution is to implement string replacement operations directly in the code. For example, with the translateTree function in src/frame/lib/page-data.ts.

When a file is renamed in github/docs-internal, say from foo.md to bar.md, the translation repos keep both foo.md and bar.md. Files are not deleted from translation repos. This limitation prevents us from scanning the directories for files like we do with the English content and it is the motivation for the creation of language-specific trees.

Translations depend on the same global store of asset images as the English content. Translation repos contain only content/ and data/ directories, but no assets/.

translations/de-de
├── content
└── data

As a result, images should not be deleted from the global store until we've confirmed they are no longer used in a translation. We have a workflow that checks for orphaned assets and removes them when they're no longer used.

This section assumes you want your local environment to replicate the structure of translations in production.

1. Clone the (private) translation repos into the translations/[lang-code] directories.

   cd docs-internal
   mkdir translations && cd translations
   
   git clone <es-es-repo-url> es-es
   git clone <ja-jp-repo-url> ja-jp
   git clone <pt-br-repo-url> pt-br
   git clone <zh-cn-repo-url> zh-cn
   git clone <ru-ru-repo-url> ru-ru
   git clone <fr-fr-repo-url> fr-fr
   git clone <ko-kr-repo-url> ko-kr
   git clone <de-de-repo-url> de-de

2. Start the development server with npm run start-all-languages. If you need to enable specific languages, you can set ENABLED_LANGUAGES. For example, to start the server with English, Japanese, and Spanish only, use: ENABLED_LANGUAGES=en,ja,es NODE_ENV=development nodemon server.ts

Note: If you ever need to place translations in a different location, use the TRANSLATIONS_ROOT environment variable. We do this for tests.

A script in src/languages/scripts can be used to update the translation repos.