Google is committed to advancing racial equity for Black communities. See how.


Localization (L10N for short) is a process for adapting the software so that it is usable in a particular region; or a number of regions. Fuchsia has a workflow that allows program authors to equip their programs with language-specific resources (a.k.a localized assets).

At present, Fuchsia's localization support is limited, when compared to the scope of all localization features out there. However, even in its limited scope (though the scope will grow) enough small scale decisions have been made that it is useful to document them in the form of a specification.

This specification is by no means complete or final. We reserve the right to modify it in the future, and though we will make a best effort to evolve it in backward compatible ways, there may be cases in which breaking changes could be introduced if the benefits outweigh the potential downsides.

Message Translation

The first functionality that Fuchsia's localization provides is message translation. Conceptually this gives a program the ability to display user-facing messages in the user's language of choice. This is achieved using locale-sensitive formatter printing. The program needs to keep the "localization context" where user's localization preferences are stored. When, in the course of program execution, there comes a point that a message "Hello world!" needs to be displayed on screen in the user's native language (for example Spanish would be encoded as "es", for European Spanish, and "es-419" for Spanish as used in the Americas), the program can look up a translation by providing an abstract [Lookup] service with the original message and the desired translation. Conceptually this will amount to a line in the code that matches this general pattern:

[Lookup].String({locale-ids=["es-419"]}, MSG_Hello_World) ⇒ (yields the translation)

In the above example, [Lookup] can be any sort of callable endpoint: it could be a library exposed function, or could be an interface point for an RPC stub that fetches the translation over the network. The abstract operation of "fetching a translated message" is here called String to distinguish it from other possible calls to typed data, such as StringArray or others.

Note, messages can get quite a bit more elaborate than that, which is why we typically don't want the program authors to handle them directly, but rather through message IDs.

Two more things are of note:

  1. The language identifiers are specified as Unicode locale IDs (hence the named parameter locale-ids in the example), and multiple such locale IDs can be provided at once. This is because users may have more than a single preferred language, and may have a hierarchy of languages by preference. This allows the localization system to choose the best available message, possibly in more than a single language in a single session.
  2. The messages are not specified as their string representation in code. Rather, they are referred to by a unique message identifier. In the example above, this was arbitrarily named MSG_Hello_World. While schools of thought differ on whether strings should be internalized or externalized, we opted for the latter. Our main reasons were to keep the source code free from linguistic concerns, which makes the translation toolchain somewhat easier to maintain, and makes translations at scale easier to manage.

Two main questions arise from the example above:

  1. What does the concrete interface to the [Lookup] service look like in the programmer's language of choice? And,
  2. How do translations make their way to my program so that they are available to use?

We will answer these in turn.

Lookup API

The Lookup API library is used to obtain translated strings. A simplified view of the Lookup API in C++ is as follows:

class Lookup {
  enum class Status {
    // No error.
    OK = 0,
    // The resource was unavailable as requested.
  static fit::result<std::unique_ptr<Lookup>, Lookup::Status>
    New(const std::vector<std::string>& locale_ids);
  fit::result<std::string_view, Lookup::Status> String(uint64_t message_id);

The actual API can be seen in the file lookup.h, and is essentially the same as shown above, except that it contains documentation, construction and testing overhead. At the time of this writing, only a high level C++ API is available for use. We will be adding high level APIs in other languages as need arises. A low-level C API is available as a basis for writing FFI bindings to this functionality in other languages. As a special case, rust does not need the FFI bindings since the low-level implementation is in rust and can be interfaced with directly; but an actual rust API has not been formulated yet.

A basic usage of the Lookup API looks like this:

std::vector<std::string> locale_ids = {"nl-NL"};
auto result = Lookup::New(locale_ids);
if (result.is_error()) {
auto lookup = result.value();
auto lookup_result = lookup.string(42);
if (lookup_result.is_error()) {
  // handle error
std::string_view message = lookup_result.value();
// Use `message`.

The example is taken from the lookup.h documentation. Knowing the API, this example is fairly straightforward, save for one thing: the call lookup.string(...) uses a magic number 42 to look up a message. It is fair of you as a programmer to ask where this number comes from. The next section addresses this question.

Localization workflow

Since it is impractical to request of a programmer to know by heart a magic key referring to a specific message, it stands to reason that the localization system should provide an ergonomic way to refer to these keys in a symbolic manner (remember the abstract example for MSG_Hello_World above.

Following the best practices for 18n and l10n, the source strings live in an XML file (here, named strings.xml), explained below. An example of a strings.xml file is shown below. The goal of this file is to declare all externalized strings that our program uses, and give them a locally unique name. The strings will be used as a basis for translation, and names will be used as a basis for the symbolic constants used to refer to particular messages.

<!-- comment -->
<?xml version="1.0" encoding="utf-8"?>
<resources xmlns:xliff="urn:oasis:names:tc:xliff:document:1.2">
  <!-- comment -->
      >string with
an intervening newline</string>

The file strings.xml goes through a series of transformations in which language-specific varieties of the same file are produced by translators. The input-output behavior of the translation process is: strings.xml file goes in, with strings written in some source (human) language, and multiple flavors of strings.xml come out, each translated in a particular single language. The entire translation process can be quite involved, in a large organization it can involve farming tasks out to translators who may live around the world, and scores of dedicated translation tooling. but the precise mechanics of the box does not matter too much to us as consumers as long as the input-output behavior of the process is upheld, and we're generally aware that the translation could take a while. The resulting files are converted into a machine-readable form, and shipped alongside a Fuchsia program within the same Fuchsia package. An important feature of Fuchsia packages is that they are inherently not an archive, but rather a manifest that points to files by their content hash. So multiple programs can share the same files, and languages closely related ("en-US", "en-GB") can potentially share message disk space. The following diagram shows a compact overview of the lifecycle of strings.

The above image shows the localization flow. Since XML files are annotated they are not directly suitable for machine translation, so we convert to JSON files, for which we can reuse available libraries to load them, and construct a map from a key to message string.  These strings can then be used as format strings in `MessageFormat`.


We are reusing the Android string resources XML format to represent localizable strings. Since we will be adding nothing to the strings.xml format, the full discussion of the features is delegated to the string resources page.

While all that XML in the above diagram makes this discussion look like it just emerged from some wormhole connected straight to the 1990s, XML is actually a very good fit to describe annotated text. strings.xml is a format that has been time-tested in Android so we know it will be adequate, and developers are familiar with it.

For example, a string resource can be declared with annotations interleaved into the source text.

<!-- … -->
<string name="title"
   >Best practices for <annotation font="title_emphasis">text</annotation> look like so</string>
<!-- … -->

Above: An example interleaving of translation text and annotation._

It is possible to interleave text that should be protected from translation, like so:

<string name="countdown">
  <xliff:g id="time" example="5 days"
    >{1}</xliff:g> until holiday</string>

Above:: An example of an interleaving of a fenced-off parameter, annotated with an example value and guarded with a tag that is not part of the string resources data schema._

We can also define our own additions to the data schema if we so need, and interleave that data schema transparently in an existing schema.

There are some necessary constraints on the contents of the file above:

  • Every name attribute in the file must be unique.
  • Name identifiers may contain uppercase and lowercase ASCII letters, digits and underscores, but may not start with a number. So for example,_H_e_L_L_o_wo_1_rld is allowed, but 0cool is not.
  • No two name-message combinations in the file may repeat.

For the time being there are no provisions for having multiple strings files in a project.

Message identifiers

The message identifiers (the "magical" numeric constants for each message) are generated based on the contents of the strings.xml file. Every string message gets a unique identifier, which is computed based on the one-way hash on name and the contents of the message itself. This identifier assignment ensures that it is vanishingly unlikely for two different messages to accidentally have the same resulting identifier.

The generation of these messages is automated by GN build rules in Fuchsia, but is ultimately performed by a program called strings_to_fidl. This program generates FIDL intermediate representation for the message IDs, and the regular FIDL toolchain is used to produce language-specific versions of that info. As an example, the C++ flavor would be a header file with the following content:

namespace fuchsia {
namespace intl {

namespace l10n {
enum class MessageIds : uint64_t {
  STRING_NAME = 42u,
  STRING_NAME_2 = 43u,
  STRING_NAME_3 = 44u,

}  // namespace l10n
}  // namespace intl
}  // namespace fuchsia

The precise values assigned to each particular enum value in the example above are not relevant. The generation method is also not relevant at this time, since all identifiers are generated at compile time and there is no opportunity for version skew. We may for now safely assume that an identical name-content combination will always have the same message ID assigned.

It is fairly easy to include the resulting file into a C++ program. A minimal example is given below, but refer to the fully worked-out example for the precise details of the wire-up. The library parameter fuchsia.intl.l10n is provided directly by the author as a flag to strings_to_fidl; or if the appropriate GN template is used, as a parameter to the GN template.

#include <iostream>

// This header file has been generated from the strings library fuchsia.intl.l10n.
#include "fuchsia/intl/l10n/cpp/fidl.h"

// Each library name segment between dots gets its own nested namespace in
// the generated C++ code.
using fuchsia::intl::l10n::MessageIds;

int main() {
  std::cout << "Constant: " << static_cast<uint64_t>(MessageIds::STRING_NAME) << std::endl;
  return 0;


The FIDL and C++ code generation makes the message IDs available to the program authors. On the packaging side, we also must provide the localized asset for each language we support. At present the encoding for this information is JSON. This was done for expedience, but a number of improvements can be made on that decision to improve performance and security.

Generating this information is delegated to the program named strings_to_json, which merges the original strings.xml with a language specific file (for example, a French translation lives in strings_fr.xml). Again, for builds driven by GN, the invocation of strings_to_json is encapsulated in a build rule.

Example contents of a generated JSON file are given below.

  "locale_id": "fr",
  "source_locale_id": "en-US",
  "num_messages": 3,
  "messages": {
    "42": "le string",
    "43": "le string 2",
    "44": "le string\nwith intervening newline"

The JSON format has the following fields currently defined. In case the table below goes out of date, the source of truth for the JSON structure is the strings model.

Field Type Description
locale_id Locale ID (string) The locale for which the messages are translated.
source_locale_id Locale ID (string) The locale of the source message file.
num_messages Positive integer The number of messages that were present in the original strings.xml. This allows us to estimate quickly the quality of the translation by comparing that number of messages with the number of messages that are present in the JSON file.
messages Map [u64->string] A map from message ID to the appropriate message.


The generated JSON file from the previous section must be bundled together with the Fuchsia program so that it can be found at program runtime. This is done by the regular Fuchsia build rules, such as those in package.gni.

We have established some conventions for packaging resources (i.e. localized assets). The schema is intended to be extensible to other asset types, and also to be able to support _combinations _ of asset types which are sometimes useful to have when expressing more complex relationships between device and locale (a Hebrew icon version for a 200dpi display). All paths below are relative to the package's data directory and are found under /pkg/data on a running system.

Path Description
assets/ Stores all assets. This is similar to how the meta/ directory contains package manifests and other metadata. In the future, this directory could contain conventional indices.
assets/locales Stores data specifically for locales
assets/locales/fr-fr Stores data for particular locales. The locale names are individual directories in BCP47 format. Each program contributes a single JSON file to this directory, named program.json, where the program part of the name is chosen by the author. We will, at some point, probably need to ensure that package and library names for files here do not collide. Also, due to Fuchsia's packaging strategy, it may pay to have many smaller files storing translations instead of one large one, for ease of update.