The evolution of Fb’s iOS app structure

Fb for iOS (FBiOS) is the oldest cell codebase at Meta. Because the app was rewritten in 2012, it has been labored on by 1000’s of engineers and shipped to billions of customers, and it could help a whole lot of engineers iterating on it at a time.

After years of iteration, the Fb codebase doesn’t resemble a typical iOS codebase:

• It’s filled with C++, Goal-C(++), and Swift.
• It has dozens of dynamically loaded libraries (dylibs), and so many courses that they’ll’t be loaded into Xcode without delay.
• There’s nearly zero uncooked utilization of Apple’s SDK — every thing has been wrapped or changed by an in-house abstraction.
• The app makes heavy use of code era, spurred by Buck, our customized construct system.
• With out heavy caching from our construct system, engineers must spend a complete workday ready for the app to construct.

FBiOS was by no means deliberately architected this manner. The app’s codebase displays 10 years of evolution, spurred by technical selections essential to help the rising variety of engineers engaged on the app, its stability, and, above all, the consumer expertise.

Now, to rejoice the codebase’s 10-year anniversary, we’re shedding some mild on the technical selections behind this evolution, in addition to their historic context.

2014: Establishing our personal cell frameworks

Two years after Meta launched the native rewrite of the Fb app, Information Feed’s codebase started to have reliability points. On the time, Information Feed’s information fashions had been backed by Apple’s default framework for managing information fashions: Core Data. Objects in Core Information are mutable, and that didn’t lend itself properly to Information Feed’s multithreaded structure. To make issues worse, Information Feed utilized bidirectional information movement, stemming from its use of Apple’s de facto design sample for Cocoa apps: Model View Controller.

In the end, this design exacerbated the creation of nondeterministic code that was very troublesome to debug or reproduce bugs. It was clear that this structure was not sustainable and it was time to rethink it.

Whereas contemplating new designs, one engineer investigated React, Fb’s (open supply) UI framework, which was turning into fairly fashionable within the Javascript group. React’s declarative design abstracted away the tough crucial code that induced points in Feed (on net), and leveraged a one-way information movement, which made the code a lot simpler to purpose about. These traits appeared properly suited to the issues Information Feed was dealing with. There was just one drawback.

There was no declarative UI in Apple’s SDK. 

Swift wouldn’t be announced for a few months, and SwiftUI (Apple’s declarative UI framework) wouldn’t be introduced till 2019. If Information Feed wished to have a declarative UI, the workforce must construct a brand new UI framework. 

In the end, that’s what they did.

After spending just a few months constructing and migrating Information Feed to run on a brand new declarative UI and a brand new information mannequin, FBiOS noticed a 50 p.c efficiency enchancment. A couple of months later, they open-sourced their React-inspired UI framework for cell, ComponentKit. 

To this present day, ComponentKit continues to be the de facto alternative for constructing native UIs in Fb. It has offered numerous efficiency enhancements to the app by way of view reuse swimming pools, view flattening, and background structure computation. It additionally impressed its Android counterpart, Litho, and SwiftUI.

In the end, the selection to exchange the UI and information layer with customized infra was a trade-off. To attain a pleasant consumer expertise that might be reliably maintained, new staff must shelve their trade information of Apple APIs to be taught the customized in-house infra. 

This wouldn’t be the final time FBiOS must decide that balanced finish consumer expertise with developer expertise and pace. Going into 2015, the app’s success would set off what we seek advice from as a function explosion. And that offered its personal set of distinctive challenges.

2015: An architectural inflection level

By 2015, Meta had doubled down on its “Mobile First” mantra, and the FBiOS codebase noticed a meteoric rise within the variety of day by day contributors. As increasingly merchandise had been built-in into the app, its launch time started to degrade, and other people started to note. Towards the top of 2015, startup efficiency was so gradual (practically 30 seconds!) that it risked being killed by the cellphone’s OS.

Upon investigation, it was clear that there have been many contributing components to degraded startup efficiency. For the sake of brevity, we’ll focus solely on those that had a long-term impact on the app’s structure:

• The app’s ‘pre-main’ time was rising at an unbounded price, because the app’s measurement grew with every product.
• The app’s ‘module’ system gave every product ungoverned entry to all of the app’s resourcing. This led to a tragedy of the commons issue as every product leveraged its ‘hook’ into startup to carry out computationally costly operations in order that preliminary navigation to that product could be snappy.

The adjustments that had been wanted to mitigate and enhance startup would basically alter the best way product engineers wrote code for FBiOS.

2016: Dylibs and modularity

In response to Apple’s wiki about improving launch times, numerous operations should be carried out earlier than an app’s ‘principal’ operate will be referred to as. Typically, the extra code an app has, the longer this can take.

Whereas ‘pre-main’ contributed solely a small subset of the 30 seconds being spent throughout launch, it was a selected concern as a result of it will proceed to develop at an unbounded price as FBiOS continued to amass new options.

To assist mitigate the unbounded development of the app’s launch time, our engineers started to maneuver massive swaths of product code right into a lazily loaded container often known as a dynamic library (dylib). When code is moved right into a dynamically loaded library, it isn’t required to load earlier than the app’s principal() operate.

Initially, the FBiOS dylib construction seemed like this:

Two product dylibs (FBCamera and NotOnStartup) had been created, and a 3rd dylib (FBShared) was used to share code between the assorted dylibs and the principle app’s binary.

The dylib answer labored superbly. FBiOS was in a position to curb the unbounded development of the app’s startup time. Because the years glided by, most code would find yourself in a dylib in order that startup efficiency stayed quick and was unaffected by the fixed fluctuation of added or eliminated merchandise within the app.

The addition of dylibs triggered a psychological shift in the best way Meta’s product engineers wrote code. With the addition of dylibs, runtime APIs like NSClassFromString() risked runtime failures as a result of the required class lived in unloaded dylibs. Since lots of the FBiOS core abstractions had been constructed on iterating by means of all of the courses in reminiscence, FBiOS needed to rethink what number of of its core techniques labored.

Except for the runtime failures, dylibs additionally launched a brand new class of linker errors. Within the occasion the code in Fb (the startup set) referenced code in a dylib, engineers would see a linker error like this:

Undefined symbols for structure arm64:
  "_OBJC_CLASS_$_SomeClass", referenced from:
      objc-class-ref in libFBSomeLibrary-9032370.a(FBSomeFile.mm.o)

To repair this, engineers had been required to wrap their code with a particular operate that would load a dylib if needed:

Out of the blue:

int principal() 
  DoSomething(context);

Would appear like this:

int principal() 
  FBCallFunctionInDylib(
    NotOnStatupFramework,
    DoSomething,
    context
  );

The answer labored, however had fairly just a few code smells:

• The app-specific dylib enum was hard-coded into numerous callsites. All apps at Meta needed to share a dylib enum, and it was the reader’s accountability to find out whether or not that dylib was utilized by the app the code was working in.
• If the incorrect dylib enum was used, the code would fail, however solely at runtime. Given the sheer quantity of code and options within the app, this late sign led to a number of frustration throughout growth.

On prime of all that, our solely system to safeguard towards the introduction of those calls throughout startup was runtime-based, and lots of releases had been delayed whereas last-minute regressions had been launched into the app.

In the end, the dylib optimization curbed the unbounded development of the app’s launch time, but it surely signified an enormous inflection level in the best way the app was architected. FBiOS engineers would spend the following few years re-architecting the app to easy among the tough edges launched by the dylibs, and we (finally) shipped an app structure that was extra sturdy than ever earlier than.

2017: Rethinking the FBiOS structure

With the introduction of dylibs, just a few key elements of FBiOS needed to be rethought:

• The ‘module registration system’ might now not be runtime-based.
• Engineers wanted a option to know whether or not any codepath throughout startup might set off a dylib load.

To handle these points, FBiOS turned to Meta’s open supply construct system, Buck.

Inside Buck, every ‘goal’ (app, dylib, library, and so forth.) is asserted with some configuration, like so:

apple_binary(
  title = "Fb",
  ...
  deps = [
    ":NotOnStartup#shared",
    ":FBCamera#shared",
  ],
)

apple_library(
  title = "NotOnStartup",
  srcs = [
    "SomeFile.mm",
  ],
  labels = ["special_label"],
  deps = [
    ":PokesModule",
    ...
  ],
)

Every ‘goal’ lists all info wanted to construct it (dependencies, compiler flags, sources, and so forth.), and when ‘buck construct’ is known as, it builds all this info right into a graph that may be queried.

$ buck question “deps(:Fb)”
> :NotOnStartup
> :FBCamera

$ buck question “attrfilter(labels, special_label, deps(:Fb))”
> :NotOnStartup

Utilizing this core idea (and a few particular sauce), FBiOS started to provide some buck queries that would generate a holistic view of the courses and capabilities within the app throughout construct. This info could be the constructing block of the app’s subsequent era of structure.

2018: The proliferation of generated code

Now that FBiOS was in a position to leverage Buck to question for details about code within the dependency, it might create a mapping of “operate/courses -> dylibs” that might be generated on the fly.

  "capabilities": 
    "DoSomething": Dylib.NotOnStartup,
    ...
  ,
  "courses": 
    "FBSomeClass": Dylib.SomeOtherOne
  

Utilizing that mapping as enter, FBiOS used it to generate code that abstracted away the dylib enum from callsites:

static std::unordered_map<const char *, Dylib> functionToDylib 
   "DoSomething", Dylib.NotOnStartup ,
   "FBSomeClass", Dylib.SomeOtherOne ,
  ...
;

Utilizing code era was interesting for just a few causes:

• As a result of the code was regenerated primarily based on native enter, there was nothing to test in, and there have been no extra merge conflicts! On condition that the engineering physique of FBiOS might double yearly, this was a giant growth effectivity win.
• FBCallFunctionInDylib no-longer required an app-specific dylib (and thus might be renamed to ‘FBCallFunction’). As an alternative, the decision would learn from static mapping generated for every software throughout construct.

Combining Buck question with code era proved to be so profitable that FBiOS used it as bedrock for a brand new plugin system, which finally changed the runtime-based app-module system.

Transferring sign to the left

With the brand new Buck-powered plugin system. FBiOS was in a position to exchange most runtime failures with build-time warnings by migrating bits of infra to a plugin-based structure.

When FBiOS is constructed, Buck can produce a graph to indicate the placement of all of the plugins within the app, like so:

From this vantage level, the plugin system can floor build-time errors for engineers to warn: 

• “Plugin D, E might set off a load of a dylib. This isn’t allowed, for the reason that caller of those plugins lives within the app’s startup path.”
• “There isn’t a plugin for rendering Profiles discovered within the app … which means that navigating to that display screen won’t work.”
• “There are two plugins for rendering Teams (Plugin A, Plugin B). Certainly one of them must be eliminated.”

With the outdated app module system, these errors could be “lazy” runtime assertions. Now, engineers are assured that when FBiOS is constructed efficiently, it gained’t fail due to lacking performance, dylibs loading throughout app startup, or invariants within the module runtime system.

The price of code era

Whereas migrating FBiOS to a plugin system has improved the app’s reliability, offered sooner alerts to engineers, and made it attainable for the app to trivially share code with our different cell apps, it got here at a price:

• Plugin errors are usually not on Stack Overflow and will be complicated to debug.
• A plugin system primarily based on code era and Buck is a far cry from conventional iOS growth. 
• Plugins introduce a layer of indirection to the codebase. The place most apps would have a registry file with all options, these are generated in FBiOS and will be surprisingly troublesome to seek out.

There isn’t a doubt that plugins led FBiOS farther away from idiomatic iOS growth, however the trade-offs appear to be price it. Our engineers can change code utilized in many apps at Meta and ensure that if the plugin system is glad, no app ought to crash due to lacking performance in a hardly ever examined codepath. Groups like Information Feed and Teams can construct an extension level for plugins and ensure that product groups can combine into their floor with out touching the core code.

2020: Swift and language structure

Whereas most of this text has centered on architectural adjustments stemming from scale points within the Fb app, adjustments in Apple’s SDK have additionally compelled FBiOS to rethink a few of its architectural selections.

In 2020, FBiOS started to see an increase within the variety of Swift-only APIs from Apple and a rising sentiment for extra Swift within the codebase. It was lastly time to reconcile with the truth that Swift was an inevitable tenant in FB apps. 

Traditionally, FBiOS had used C++ as a lever to construct abstraction, which saved on code measurement due to C++’s zero overhead principle. However C++ doesn’t interop with Swift (but). For most FBiOS APIs (like ComponentKit), some form of shim must be created to make use of in Swift — creating code bloat.

Right here’s a diagram outlining the problems within the codebase:

With this in thoughts, we started to kind a language technique about when and the place numerous bits of code must be used:

In the end, the FBiOS workforce started to advise that product-facing APIs/code shouldn’t comprise C++ in order that we might freely use Swift and future Swift APIs from Apple. Utilizing plugins, FBiOS might summary away C++ implementations in order that they nonetheless powered the app however had been hidden from most engineers.

One of these workstream signified a little bit of shift in the best way FBiOS engineers considered constructing abstractions. Since 2014, among the greatest components in framework constructing have been contributions to app measurement and expressiveness (which is why ComponentKit selected Goal-C++ over Goal-C).

The addition of Swift was the primary time these would take a backseat to developer effectivity, and we count on to see extra of that sooner or later.

2022: The journey is 1 p.c  completed

Since 2014, FBiOS structure has shifted fairly a bit:

• It launched numerous in-house abstractions, like ComponentKit and GraphQL.
• It makes use of dylibs to maintain ‘pre-main’ occasions minimal and contribute to a blazing-fast app startup.
• It launched a plugin system (powered by Buck) in order that dylibs are abstracted away from engineers, and so code is well shareable between apps.
• It launched language pointers about when and the place numerous languages must be used and started to shift the codebase to mirror these language pointers.

In the meantime, Apple has launched thrilling enhancements to their telephones, OS, and SDK:

• Their new telephones are quick. The price of loading is way smaller than it was earlier than.
• OS enhancements like dyld3 and chain fixups present software program to make code loading even sooner.
• They’ve launched SwiftUI,  a declarative API for UI that shares a number of ideas with ComponentKit.
• They’ve offered improved SDKs, in addition to APIs (like interruptible animations in iOS8) that we might have constructed customized frameworks for.

As extra experiences are shared throughout Fb, Messenger, Instagram, and WhatsApp, FBiOS is revisiting all these optimizations to see the place it could transfer nearer to platform orthodoxy. In the end, we’ve seen that the best methods to share code are to make use of one thing that the app offers you totally free or construct one thing that’s just about dependency-free and might combine between all of the apps.

We’ll see you again right here in 2032 for the recap of the codebase’s 20-year anniversary!