App Service

Chrome, and especially Chrome OS, has apps, e.g. chat apps and camera apps.

There are a number (lets call it M) of different places or app Consumers, usually UI (user interfaces) related, where users interact with their installed apps: e.g. launcher, search bar, shelf, New Tab Page, the App Management page, permissions or settings pages, picking and running default handlers for URLs, MIME types or intents, etc.

There is also a different number (N) of app platforms or app Providers: built-in apps, extension-backed apps, PWAs (progressive web apps), ARC++ (Android apps), Crostini (Linux apps), etc.

Historically, each of the M Consumers hard-coded each of the N Providers, leading to M×N code paths that needed maintaining, or updating every time M or N was incremented.

This document describes the App Service, an intermediary between app Consumers and app Providers. This simplifies M×N code paths to M+N code paths, each side with a uniform API, reducing code duplication and improving behavioral consistency. This service (a Mojo service) could potentially be spun out into a new process, for the usual Servicification benefits (e.g. self-contained services are easier to test and to sandbox), and would also facilitate Chrome OS apps that aren't tied to the browser, e.g. Ash apps.

The App Service can be decomposed into a number of aspects. In all cases, it provides to Consumers a uniform API over the various Provider implementations, for these aspects:

  • App Registry: list the installed apps.
  • App Icon Factory: load an app's icon, at various resolutions.
  • App Runner: launch apps and track app instances.
  • App Installer: install, uninstall and update apps.
  • App Coordinator: keep system-wide settings, e.g. default handlers.

Some things are still the responsbility of individual Consumers or Providers. For example, the order in which the apps' icons are presented in the launcher is a launcher-specific detail, not a system-wide detail, and is managed by the launcher, not the App Service. Similarly, Android-specific VM (Virtual Machine) configuration is Android-specific, not generalizable system-wide, and is managed by the Android provider (ARC++).


Talk of the App Service is an over-simplification. There will be an App Service instance per Profile, as apps can be installed for a Profile.

Note that this doesn‘t require the App Service to know about Profiles. Instead, Profile-specific code (e.g. a KeyedService) finds the Mojo service Connector for a Profile, creates an App Service and binds the two (App Service and Connector), but the App Service itself doesn’t know about Profiles per se.

App Registry

The App Registry's one-liner mission is:

  • I would like to be able to for-each over all the apps in Chrome.

An obvious initial design for the App Registry involves three actors (Consumers ⇔ Service ⇔ Providers) with the middle actor (the App Registry Mojo service) being a relatively thick implementation with a traditional GetFoo, SetBar, AddObserver style API. The drawback is that Consumers are often UI surfaces and UI code likes synchronous APIs, but Mojo APIs are generally asynchronous, especially as it may cross process boundaries.

Instead, we use four actors (Consumers ↔ Proxy ⇔ Service ⇔ Providers), with the Consumers ↔ Proxy connection being synchronous and in-process, lighter than the async / out-of-process ⇔ connections. The Proxy implementation is relatively thick and the Service implementation is relatively thin, almost trivially so. Being able to for-each over all the apps is:

for (const auto& app : GetAppServiceProxy(profile).GetCache().GetAllApps()) {

The Proxy is expected to be in the same process as its Consumers, and the Proxy would be a singleton (per Profile) within that process: Consumers would connect to the in-process Proxy. If all of the app UI code is in the browser process, the Proxy would also be in the browser process. If app UI code migrated to e.g. a separate Ash process, then the Proxy would move with them. There might be multiple Proxies, one per process (per Profile).

Code Location

Some code is tied to a particular process, some code is not. For example, the per-Profile AppServiceProxy obviously contains Profile-related code (i.e. a KeyedService, so that browser code can find the AppServiceProxy for a given Profile) that is tied to being in the browser process. The AppServiceProxy also contains process-agnostic code (code that could conceivably be used by an AppServiceProxy living in an Ash process), such as code to cache and update the set of known apps (as in, the App Mojo type). Specifically, the AppServiceProxy code is split into two locations, one under //chrome/browser and one not:

  • //chrome/browser/apps/app_service
  • //chrome/services/app_service

On the Provider side, code specific to extension-backed applications or web applications (as opposed to ARC++ or Crostini applications) lives under:

  • //chrome/browser/extensions
  • //chrome/browser/web_applications

Matchmaking and Publish / Subscribe

The AppService itself does not have an GetAllApps method. It doesn‘t do much, and it doesn’t keep much state. Instead, the App Registry aspect of the AppService is little more than a well known meeting place for Publishers (i.e. Providers) and Subscribers (i.e. Proxies) to discover each other. An analogy is that it‘s a matchmaker for Publishers and Subscribers, although it matches all to all instead of one to one. Publishers don’t meet Subscribers directly, they meet the matchmaker, who introduces them to Subscribers.

Once a Publisher and Subscriber connect, the Pub-side sends the Sub-side a stream of Apps (calling the Subscriber‘s OnApps method). On the initial connection, the Publisher calls OnApps with "here’s all the apps that I (the Publisher) know about", with additional OnApps calls made as apps are installed, uninstalled, updated, etc.

As mentioned, the App Registry aspect of the AppService doesn't do much. Its part of the AppService Mojo interface is:

interface AppService {
  // App Registry methods.
  RegisterPublisher(Publisher publisher, AppType app_type);
  RegisterSubscriber(Subscriber subscriber, ConnectOptions? opts);

  // Some additional methods; not App Registry related.

interface Publisher {
  // App Registry methods.
  Connect(Subscriber subscriber, ConnectOptions? opts);

  // Some additional methods; not App Registry related.

interface Subscriber {
  OnApps(array<App> deltas);

enum AppType {

struct ConnectOptions {
  // TBD: some way to represent l10n info such as the UI language.

Whenever a new Publisher is registered, it is connected with all of the previously registered Subscribers, and vice versa. Once a Publisher is connected directly to a Subscriber, the AppService is no longer involved. Even as new apps are installed, uninstalled, updated, etc., the app's Publisher talks directly to each of its (previously connected) Subscribers, without involving the AppService.

TBD: whether we need un-registration and dis-connect mechanisms.

The App Type

The one Mojo struct type, App, represents both a state, “an app that's ready to run”, and a delta or change in state, “here‘s what’s new about an app”. Deltas include events like “an app was just installed” or “just uninstalled” or “its icon was updated”.

This is achieved by having every App field (other than App.app_type and App.app_id) be optional. Either optional in the Mojo sense, with type T? instead of a plain T, or if that isn't possible in Mojo (e.g. for integer or enum fields), as a semantic convention above the Mojo layer: 0 is reserved to mean “unknown”. For example, the App.show_in_launcher field is a OptionalBool, not a bool.

An App.readiness field represents whether an app is installed (i.e. ready to launch), uninstalled or otherwise disabled. “An app was just installed” is represented by a delta whose readiness is kReady and the old state's readiness being some other value. This is at the Mojo level. At the C++ level, the AppUpdate wrapper type (see below) can provide helper WasJustInstalled methods.

The App, Readiness and OptionalBool types are:

struct App {
  AppType app_type;
  string app_id;

  // The fields above are mandatory. Everything else below is optional.

  Readiness readiness;
  string? name;
  IconKey? icon_key;
  OptionalBool show_in_launcher;
  // etc.

enum Readiness {
  kUnknown = 0,
  kReady,                // Installed and launchable.
  kDisabledByBlacklist,  // Disabled by SafeBrowsing.
  kDisabledByPolicy,     // Disabled by admin policy.
  kDisabledByUser,       // Disabled by explicit user action.

enum OptionalBool {
  kUnknown = 0,

// struct IconKey is discussed in the "App Icon Factory" section.

A new state can be mechanically computed from an old state and a delta (both of which have the same type: App). Specifically, last known value wins. Any known field in the delta overwrites the corresponding field in the old state, any unknown field in the delta is ignored. For example, if an app‘s name changed but its icon didn’t, the delta's field (a base::Optional<std::string>) would be known (not base::nullopt) and copied over but its App.icon field would be unknown (base::nullopt) and not copied over.

The current state is thus the merger or sum of all previous deltas, including the initial state being a delta against the ground state of “all unknown”. The AppServiceProxy tracks the state of its apps, and implements the (in-process) Observer pattern so that UI surfaces can e.g. update themselves as new apps are installed. There's only one method, OnAppUpdate, as opposed to separate OnAppInstalled, OnAppUninstalled, OnAppNameChanged, etc. methods. An AppUpdate is a pair of App values: old state and delta.

class AppRegistryCache {
  class Observer : public base::CheckedObserver {
    ~Observer() override {}
    virtual void OnAppUpdate(const AppUpdate& update) = 0;

  // Etc.

App Icon Factory

Icon data (even compressed as a PNG) is bulky, relative to the rest of the App type. Publishers will generally serve icon data lazily, on demand, especially as the desired icon resolutions (e.g. 64dip or 256dip) aren't known up-front. Instead of sending an icon at all possible resolutions, the Publisher sends an IconKey: enough information to load the icon at given resolutions.

An IconKey augments the AppType app_type and string app_id. For example, some icons are statically built into the Chrome or Chrome OS binary, as PNG-formatted resources, and can be loaded (synchronously, without sandboxing). They can be loaded from the IconKey.resource_id. Other icons are dynamically (and asynchronously) loaded from the extension database on disk. The base icon can be loaded just from the app_id alone.

In either case, the IconKey.icon_effects bitmask holds whether to apply further image processing effects such as desaturation to gray.

interface AppService {
  // App Icon Factory methods.
      AppType app_type,
      string app_id,
      IconKey icon_key,
      IconCompression icon_compression,
      int32 size_hint_in_dip,
      bool allow_placeholder_icon) => (IconValue icon_value);

  // Some additional methods; not App Icon Factory related.

interface Publisher {
  // App Icon Factory methods.
      string app_id,
      IconKey icon_key,
      IconCompression icon_compression,
      int32 size_hint_in_dip,
      bool allow_placeholder_icon) => (IconValue icon_value);

  // Some additional methods; not App Icon Factory related.

struct IconKey {
  // A monotonically increasing number so that, after an icon update, a new
  // IconKey, one that is different in terms of field-by-field equality, can be
  // broadcast by a Publisher.
  // The exact value of the number isn't important, only that newer IconKey's
  // (those that were created more recently) have a larger timeline than older
  // IconKey's.
  // This is, in some sense, *a* version number, but the field is not called
  // "version", to avoid any possible confusion that it encodes *the* app's
  // version number, e.g. the "2.3.5" in "FooBar version 2.3.5 is installed".
  // For example, if an app is disabled for some reason (so that its icon is
  // grayed out), this would result in a different timeline even though the
  // app's version is unchanged.
  uint64 timeline;
  // If non-zero, the compressed icon is compiled into the Chromium binary
  // as a statically available, int-keyed resource.
  int32 resource_id;
  // A bitmask of icon post-processing effects, such as desaturation to
  // gray and rounding the corners.
  uint32 icon_effects;

enum IconCompression {

struct IconValue {
  IconCompression icon_compression;
  gfx.mojom.ImageSkia? uncompressed;
  array<uint8>? compressed;
  bool is_placeholder_icon;

Icon Changes

Apps can change their icons, e.g. after a new version is installed. From the App Service‘s point of view, an icon change is like any other change: Providers broadcast an App value representing what’s changed (icon or otherwise) about an app, the Proxy‘s AppRegistryCache enriches this App struct to be an AppUpdate, and AppRegistryCache observers can, if that AppUpdate shows that the icon has changed, issue a new LoadIcon Mojo call. A new Mojo call is necessary, because a Mojo callback is a base::OnceCallback, so the same callback can’t be used for both the old and the new icon.

Caching and Other Optimizations

Grouping the IconKey with the other LoadIcon arguments, the combination identifies a static (unchanging, but possibly obsolete) image: if a new version of an app results in a new icon, or if a change in app state results in a grayed out icon, this is represented by a different, larger IconKey.timeline. As a consequence, the combined LoadIcon arguments can be used to key a cache or map of IconValues, or to recognize and coalesce multiple concurrent requests to the same combination.

Such optimizations can be implemented as a series of “wrapper” classes (as in the classic “decorator” or “wrapper” design pattern) that all implement the same C++ interface (an IconLoader interface). They add their specific feature (e.g. caching) by wrapping another IconLoader, doing feature-specific work on every call or reply before sending the call forward or the reply backward.

There may be multiple caches, as there may be multiple cache eviction policies (also known as garbage collection policies), spanning the trade-off from favoring minimizing memory use to favoring maximizing cache hit rates. The Proxy may have a single cache, with a relatively aggressive eviction policy, which applies to all of its Consumer clients. A Consumer might have an additional Consumer-specific cache, with a more relaxed eviction policy, if it has additional Consumer-specific UI signals to guide when icon-loading requests and cache hits are more or less likely.

Note that cache values (the IconValue Mojo struct) are, primarily, a gfx.mojom.ImageSkia, which are cheap to share. Copying an ImageSkia value does not duplicate any underlying pixel buffers.

As a separate optimization, if the AppServiceProxy knows how to load an icon for a given IconKey, it can skip the Mojo round trip and bulk data IPC and load it directly instead. For example, it may know how to load icons from a statically built resource ID.

Placeholder Icons

It can take some time for Publishers to provide an icon. For example, loading the canonical icon for an ARC++ or Crostini app might require waiting for a VM to start. Such icons are often cached on the file system, but on a cache miss, there may be a number of seconds before the system can present an icon. In this case, we might want to present a Publisher-specific placeholder, typically loaded from a resource (an asset statically compiled into the binary).

There are two boolean fields that facilitate this: allow_placeholder_icon is sent from a Subsciber to a Publisher and is_placeholder_icon is sent in the response.

LoadIcon's allow_placeholder_icon states whether the the caller will accept a placeholder if the real icon can not be provided quickly. Native user interfaces like the app launcher will probably set this to true. On the other hand, serving Web-UI URLs such as chrome://app-icon/app_id/icon_size will set this to false, as that URL should identify a particular icon, not one that changes over time. Web-UI that wants to display placeholder icons and be notified of when real icons are ready will require some mechanism other than a chrome:://app-icon/etc URL.

IconValue‘s is_placeholder_icon states whether the icon provided is a placeholder. That field should only be true if the corresponding LoadIcon call had allow_placeholder_icon true. When the LoadIcon caller receives a placeholder icon, it is up to the caller to issue a new LoadIcon call, this time with allow_placeholder_icon false. As before, a new Mojo call is necessary, because a Mojo callback is a base::OnceCallback, so the same callback can’t be used for both the placeholder and the real icon.

Provider-Specific Subtleties

Some concerns (like caching and coalescing multiple in-flight calls with the same IconKey) are not specific to any particular Providers like ARC++ or Crostini, and can be solved by the Proxy.

Other concerns are Provider-specific, and are generally solved in Provider implementations, albeit often with non-Provider-specific support (such as for placeholder icons, discussed above). Such concerns include:

  • Multiple icon sources: some icons for built-in VM-based apps (e.g. ARC++ or Crostini) should be served from a compiled-into-the-browser resource instead of from the VM.
  • Pending LoadIcon calls: some LoadIcon calls might need to wait on bringing up a VM.
  • Potential on-disk corruption: for whatever reason, an on-disk file that's meant to hold a cached icon may be missing or invalid. In that case, the Provider should still provide a (placeholder) icon, and trigger Provider-specific clean-up and re-load of the real app icon.

All of these concerns listed should be straightforward to handle, and don't invalidate the overall App Service Publisher.LoadIcon Mojo design, including its non-Provider-specific caching and other optimization layers.

There are also yet another category of concerns that are Provider-specific, but also outside the purview of the App Service. For example, the file system layout of ARC++‘s on-disk icon cache is, from the App Service’s point of view, considered a private ARC++ implementation detail. As long as ARC++'s API remains the same, and if ARC++ can notify the App Service if the App Service needs to reload any or all icons, then any change in ARC++‘s file system layout isn’t a direct concern to the App Service.

App Runner

Each Publisher has (Publisher-specific) implementations of e.g. launching an app and populating a context menu. The AppService presents a uniform API to trigger these, forwarding each call on to the relevant Publisher:

interface AppService {
  // App Runner methods.
  Launch(AppType app_type, string app_id, LaunchOptions? opts);
  // etc.

  // Some additional methods; not App Runner related.

interface Publisher {
  // App Runner methods.
  Launch(string app_id, LaunchOptions? opts);
  // etc.

  // Some additional methods; not App Runner related.

struct LaunchOptions {
  // TBD.

TBD: details for context menus.

TBD: be able to for-each over all the app instances, including multiple instances (e.g. multiple windows) of the one app.

App Installer

This includes Provider-facing API (not Consumer-facing API like the majority of the AppService) to help install and uninstall apps consistently. For example, one part of app installation is adding an icon shortcut (e.g. on the Desktop for Windows, on the Shelf for Chrome OS). This helper code should be written once (in the AppService), not N times in N Providers.

TBD: details.

App Coordinator

This keeps system-wide or for-apps-as-a-whole preferences and settings, e.g. out of all of the installed apps, which app has the user preferred for photo editing. Consumer- or Provider-specific settings, e.g. icon order in the Chrome OS shelf, or Crostini VM configuration, is out of scope of the App Service.

TBD: details.

Updated on 2019-03-20.