At a high-level the architectural patterns used are composed of the following pieces:
This section describes each of the components and how they should be used. The source code in the Blueprint app is built using this architecture and can be used as a source code reference.
Model is the data model representing the raw data. This is generally returned from the backend or loaded from the disk cache. Models are typically generated by some code generator such as the GQL generator.
View is the UI component that utilizes Apple's UI frameworks. In almost all cases, views should be written with the modern SwiftUI framework. At the time of writing, SwiftUI does lack some features and flexibilities compared to the old UIKit framework. However given that it is clearly the direction Apple is pushing, it is reasonable and more future-proof to adopt it. In rare cases where SwiftUI cannot satisfy the product needs, UIKit may be used and then embedded into a SwiftUI view hierarchy.
TypeErasedView
is a drop in replacement for SwiftUI's AnyView
. AnyView
has two major issues that the
TypeErasedView
solves:
- A view hierarchy containing many
AnyView
instances can have rendering performance issues. - Wrapping an
AnyView
within anotherAnyView
can erase certain modifiers such asonAppear
andonDisappear
. This is quite error-prone as different portions of the codebase can unintentionally cause more than one wrapping.
Using a TypeErasedView
enables many architectural benefits:
- Each feature can be written in its own package without exposing the individual classes as
public
. A singlepublic
factory can be used to return aTypeErasedView
to the parent view package for presentation. - A parent feature containing many child views does not need to be modified when new child views are added or modified, when the child views are constructed as plugins.
- A view/feature's lifecycle can be managed by retaining and releasing the
TypeErasedView
reference.
ViewModel is a helper object that encapsulates the logic of transforming data models into presentation models that the
view can display. In SwiftUI, ViewModels are ObservableObject
types that publish presentation data to views. It is
important to note that ViewModels should not handle business logic such as network or user interactions. That logic
is encapsulated by the Interactor.
Presentation data is not only texts and images that are displayed on the screen. It also includes states of the view such as navigation states. For example, a ViewModel may publish an enum property. The corresponding view observes the property and shows different subviews depending on the value of the property. In this case, the ViewModel's property can be considered as view states.
ViewModels are optional objects in this architecture. Not all Views have a corresponding ViewModel. Simple views that encapsulate static UI elements do not have to have a corresponding ViewModel.
An Interactor is an object, typically a class, designed to encapsulate the application business logic such as network and user interactions. An interactor may subscribe to a network data stream to fetch models from the backend as a view first appears. It may send network requests to the backend when a user performs an action. After network data is received, the Interactor passes this data to the ViewModel. The ViewModel then transforms the network data into presentation data for the View to observe and display.
An Interactor is bound to the lifecycle of its corresponding View. As the view appears and disappears, the Interactor's
onLoad
, onViewAppear
and onViewDisappear
functions are invoked accordingly. These functions are overridable,
providing extension points for the concrete implementations to perform various logic during these different lifecycle
stages. Concrete implementations of Interactor should extend from the base Interactor
or ViewModelInteractor
class
in the IMVVMFoundation
package, depending on if the interactor has an associated ViewModel.
Interactors are optional objects in this architecture. Not all Views have a corresponding Interactor. Simple views that encapsulate static UI elements do not have to have a corresponding Interactor.
As the diagram above illustrated, the ownership of View, ViewModel and Interactor flow in the opposite direction as the
data. View owns both ViewModel and Interactor. Interactor receives a reference to its ViewModel via its lifecycle
methods such as onLoad
, onViewAppear
, etc. Interactors do not directly own their ViewModels.
IMPORTANT
A View must reference its Interactor and ViewModel via the @StateObject
property wrapper. This allows SwiftUI to
properly manage the ownerships.
@StateObject var interactor: MyInteractor
@StateObject var viewModel: MyViewModel
Please see Why use @StateObject for Interactor and ViewModel below for detailed reasoning.
In order to decouple all the different types, initialization should be performed via a factory pattern. This means none of the types, View, ViewModel or Interactor directly instantiates any other types. Instead a separate factory is used to instantiate each type and constructor inject them into each other. This factory is generally the component object in the Dependency Injection system. For example:
public class MyComponent: Component<MyDependency> {
public func make() -> some View {
MyView(interactor: MyInteractor(...), viewModel: MyViewModel(...))
}
}
IMPORTANT
The initialization of Interactor and ViewModel must be embedded within the initialization of their owning View. This
allows the @StateObject
property wrapper to correctly track the references. Initializing either outside of the
View's initialization will result in memory-leaks.
public class MyComponent: Component<MyDependency> {
public func make() -> some View {
let interactor = MyInteractor(...) // !!! Memory-leak !!!
return MyView(interactor: interactor, viewModel: MyViewModel(...))
}
}
Please see Why embed Interactor and ViewModel initialization below for detailed reasoning.
If the feature does not have a DI component, then a custom factory object can be created to provide the same
make
function implementation as above.
Lifecycle refers to different stages of an object that can be used to perform different operations.
The View's lifecycle is directly controlled by SwiftUI. It has four stages: init
, onAppear
, onDisappear
and
deinit
. Generally a view does not need to utilize these stages beyond the most obvious cases. let
properties are
set during the init
stage and released during the deinit
stage.
The ViewModel only has two stages, init
and deinit
. Generally a ViewModel only needs to be aware of the init
stage. It sets initial values for its published properties.
The Interactor has five stages that it can use to perform business logic: init
, onLoad
, onViewAppear
,
onViewDisappear
and deinit
. The init
stage should only involve setting let
properties and initial values. The
onLoad
stage can be used to setup persistent operations such as subscriptions to data streams. It can also utilize
this stage to perform initial network requests. The onViewAppear
stage can be used to perform business logic that is
only required after its corresponding view has appeared on the screen. Similarly the onViewDisappear
stage can be
used to perform business logic that is only required after its corresponding view has disappeared from the screen.
Finally the deinit
stage can be used to release any resources.
For Combine subscriptions, the Interactor base implementation provides utility operators that help bind a
subscription's lifecycle to either the Interactor's onViewDisappear
or deinit
lifecycle. This allows the
subscription to be automatically cancelled when either the corresponding view disappears or the interactor deinits.
@CancellableBuilder
override func onLoad(viewModel: MyViewModel) -> [AnyCancellable] {
publisher
.sink {
...
}
.bindToDeinit(of: self) // self is the interactor instance.
}
@CancellableBuilder
override func onViewAppear(viewModel: MyViewModel) -> [AnyCancellable] {
publisher
.sink {
...
}
.bindToViewDidDisappear(of: self) // self is the interactor instance.
}
If the Interactor does not have a corresponding ViewModel, it should inherit from the Interactor
base class instead
of the ViewModelInteractor
class. The lifecycle methods are the same, except in this case the viewModel
parameter
is omitted.
Both ViewModel and Interactor are tied to the lifecycle of the View. This means that fundamentally there is only a single lifecycle driven by SwiftUI directly.
IMPORTANT
A View that has an Interactor must bind the interactor via the bind(observer:)
or bind(observer: viewModel:)
modifier of the view. This establishes the lifecycle connection between the View, the Interactor and the ViewModel.
This operation must be performed within the View's body
property.
struct MyView<Interactor: ViewWithModelLifecycleObserver<MyViewModel> & MyViewHandler>: View {
@StateObject var interactor: Interactor
@StateObject var viewModel: MyViewModel
var body: some View {
content
.bind(observer: interactor, viewModel: viewModel)
}
}
struct MyView<Interactor: ViewLifecycleObserver & MyViewHandler>: View {
@StateObject var interactor: Interactor
var body: some View {
content
.bind(observer: interactor)
}
}
There are many reasons why it is a good practice to decouple the View from its Interactor's concrete implementation. The SwiftUI previews section below demonstrates such a use case. The decoupling can be simply implemented via an interactor protocol. The View only references the protocol and never the concrete implementation. The implementation can conform to the protocol.
protocol MyViewHandler {
func doStuff(viewModel: MyViewModel)
}
struct MyView<Interactor: ViewWithModelLifecycleObserver<MyViewModel> & MyViewHandler>: View {
@StateObject var interactor: Interactor
@StateObject var viewModel: MyViewModel
var body: some View {
Button("My Button") {
interactor.dofStuff(viewModel: viewModel)
}
}
}
class MyInteractor: Interactor, MyViewHandler {
func doStuff(viewModel: MyViewModel) {
...
}
}
SwiftUI previews can significantly improve the development process of SwiftUI views. The iMVVM architecture is designed to support building SwiftUI previews. The previews can be created by providing fixture ViewModel data and mocked Interactor types.
class MyViewModel: ObservableObject {
@Published var ...
...
}
protocol MyViewHandler {
func doStuff(viewModel: MyViewModel)
...
}
struct MyView<Interactor: ViewWithModelLifecycleObserver<MyViewModel> & MyViewHandler>: View {
@StateObject var interactor: Interactor
@StateObject var viewModel: MyViewModel
}
#if DEBUG
class MyViewPreviewsInteractor: ViewModelInteractor<MyViewModel> MyViewHandler {
func doStuff(viewModel: MyViewModel) {
viewModel.property = "new value"
}
}
struct MyViewPreviews: PreviewProvider {
static var previews: some View {
MyView(interactor: MyViewPreviewsInteractor(), viewModel: MyViewModel())
}
}
#endif
This section provides further details on some of the information described above.
In SwiftUI @StateObject
is used to declare a View is the owner of an object. At runtime, SwiftUI replaces old
instances of Views with new ones when the contents of the View changes. By declaring an Interactor with the @StateObject
property wrapper, SwiftUI ensures the same instance of the Interactor is retained in memory and linked to the new
instance of the View when such a replacement occurs.
Because Interactors are typically stateful, this ensures the data already fetched from the backend is retained properly. Otherwise unnecessary data fetches would have to occur for a new instance of the Interactor to populate the View. The exact same reasoning applies to the ViewModel as well.
Beyond properly retaining state, because the Interactor is bound to the View's onAppear
and onDisappear
lifecycle
methods, the Interactor only "activates" when the View's onAppear
method is invoked by SwiftUI. When a View instance
is replaced by SwiftUI due to content changes, as SwiftUI considers the two instances of the View are the same, the new
View instance's onAppear
method is NOT invoked after the replacement. This means if the Interactor is not referenced
as a @StateObject
, a new instance of Interactor is created but it will never "activate" to perform any work. A clear
symptom of this issue is after a View's content changes, the View gets stuck in its loading state.
A quirk of the @StateObject
property wrapper is that its initializer is declared with @autoclosure
. When an
Interactor or ViewModel is declared an @StateObject
and passed into the View's initializer, whatever is passed in
is automatically wrapped with a closure. This allows the @StateObject
property wrapper to lazily instantiate the
actual object after the View has been properly installed in the view hierarchy.
If an instance of the Interactor is instantiated outside the View's initializer:
func make() -> some View {
let interactor = MyInteractor()
return MyView(interactor: interactor)
}
SwiftUI would retain that instance via the closure. The first time when the View and its Interactor is instantiated,
everything works as expected. Once the View updates and SwiftUI instantiates the second instances however, a
memory-leak would occur. As the function that instantiates the View and Interactor is invoked for the second time, a
new instance of the Interactor and a new instance of the View is created. SwiftUI first replaces the old instance of
the View with the new instance. At this time, only a single instance of the View exists. The old instance of Interactor
is still retained by @StateObject
, and so is the new instance. SwiftUI then installs the old instance of Interactor
to the new instance of the View as expected. This completes the update process. However, the new instance of Interactor
which has already been created before the replacement even occurs, is still retained by the @StateObject
closure,
as @StateObject
s are managed as a global cache by SwiftUI. This leaks the new instance of the Interactor.
If the Interactor is properly initialized by embedding into the View's initializer:
func make() -> some View {
MyView(interactor: MyInteractor())
}
whenever the function is invoked, no instances of the Interactor is actually created. The @autoclosure
only retains
the initializer function of the Interactor. Once the replacement is completed, no new instances of the Interactor was
ever created. Therefore, no memory-leak!
This architectural pattern provides a reasonable separation between data, view and business logic without adding too much complexity or overhead. For comparison with other patterns, please see the Alternative architectures section below.
Dependency injection, or DI, is used to decouple unrelated code and enable unit testing. Instead of coupling classes together with concrete implementations, DI should be used to link disparate code together.
In order to achieve compile-time safety, the framework Needle is used. For further details on the benefits of using DI, please see Needle's documentation.
Please refer to Needle's official documentation on how to use the framework.
As mentioned above, the DI component of a feature generally acts as the factory of the feature. It declares a make
function that returns the View with its Interactor already bound.
public class MyComponent: Component<MyDependency> {
public func make() -> some View {
MyView(interactor: MyInteractor(), viewModel: MyViewModel())
}
}
Scopes naturally emerge with iMVVM and DI patterns. A scope can be defined generally in three ways:
- An iMVVM set of classes.
- A node in the DI graph.
- A state in the application.
For example, a basic app can be divided into LoggedIn
and LoggedOut
scopes. From an iMVVM perspective, each scope
has their own iMVVM objects such as LoggedInView
, LoggedInViewModel
, etc. From a DI perspective, both LoggedIn
and LoggedOut
represent nodes in the DI graph. Using Needle specific terminologies, there is a LoggedInComponent
and a LoggedOutComponent
. From an application state perspective, the LoggedIn
scope represents the state where the
user has successfully signed into the app, whereas the LoggedOut
scope represents the state when the user has not
been authenticated.
The plugin pattern allows separation between disjointed parts of the application to be decoupled yet integrated with clean interfaces. The architectural approach we decided on is that "(almost) everything is a plugin". This means every set of features such as items in a feed, tabs in a tab view are implemented as plugins. Utility objects such as workers objects can also be plugins.
Each plugin has its own feature flag. The plugin is only instantiated and integrated with its parent if the corresponding feature flag is turned on. All plugins are defaulted to the "on" state. This means when the feature flag framework fails to retrieve a value for a specific feature flag, the associated plugin is by default instantiated and integrated. This optimistic approach of assuming feature flags are default "on" simplifies the understanding of most runtime code paths. At the same time any instabilities in the feature flag framework will not affect the entire app. This implies turning off a plugin is a best-effort operation, due to the potential instabilities in the feature flag framework.
Please refer to the PluginFramework
package for the base implementations. Most commonly, a set of SwiftUI View
based structures such as Plugin
and ForEachPlugin
provide integrations between standard SwiftUI View
and the
plugin pattern.
struct ParentView: View {
let pluginViewBuilder: (MyFeatureFlag) -> TypeErasedView
var body: some View {
ForEachPlugin(featureFlags: MyFeatureFlags.allCases, viewBuilder: pluginViewBuilder)
}
}
In order to codify the Xcode project configurations and avoid project file merge conflicts, all apps' Xcode projects should be generated from manifest files. (Tuist)[https://docs.tuist.io/] is a good tool for project generation.
In order to create a new app, a Project.swift
manifest file should be created to define the structure and
configuration of the Xcode project. Please refer to
Tuist's documentation for details.
The manifest file for the Blueprint app can be used as a quick-start reference.
To streamline the local development process, it is generally a best practice to integrate Needle DI code generation with the Xcode project. Because DI code is both required to compile the application and verify the DI graph is setup correctly, it needs to be integrated as both the app scheme's pre-build action and the target's pre-build phase. In order to fail the build process in case the DI graph is not properly setup, the pre-build phase integration is used. This helps to detect DI graph issues such as missing dependencies during local development. Xcode scheme's pre-action does not fail the build process regardless of the result from the integrated actions.
schemes: [
Scheme(
name: "APP_NAME",
shared: true,
buildAction: .buildAction(
targets: ["APP_TARGET_NAME"],
preActions: [
ExecutionAction(
title: "Generate Needle",
scriptText: "cd \"$SRCROOT\" && /usr/local/bin/needle-generator \"$SRCROOT\"/needle.json",
target: "APP_TARGET_NAME"
),
...
targets: [
Target(
name: "Otter",
platform: .iOS,
scripts: [
.pre(script: "cd \"$SRCROOT\" && /usr/local/bin/needle-generator \"$SRCROOT\"/needle.json", name: "Generate Needle"),
],
...
Since application code relies on GQL generated code to compile, it is generally a best practice to integrate the GQL generator with the Xcode project to streamline the development process. Unlike Needle however, GQL generation is a slower process. Because of this, it is better to only integrate it as the app scheme's pre-build action. This means that if the generation process fails due to validation errors, the build process will not be affected. Fortunately, in most scenarios, if the generation fails, the application source code would not compile properly either.
schemes: [
Scheme(
name: "APP_NAME",
shared: true,
buildAction: .buildAction(
targets: ["APP_TARGET_NAME"],
preActions: [
ExecutionAction(
title: "Generate GQL",
scriptText: "cd \"$SRCROOT\" && /usr/local/bin/gql-generator \"$SRCROOT\"",
target: "APP_TARGET_NAME"
),
...
In order to write unit tests, mocks are necessary in most cases. Mockingbird is a good tool to provide this functionality.
There are times where the generated mocks are insufficient for the tests cases. In these cases, manually writing mocks is the solution.
This section provides additional context on the decisions described above. This information isn't crucial in developing iOS apps. It is provided here for prosperity.
The following alternatives were considered and compared to the chosen iMVVM pattern.
For large scale apps that concentrates many features into a single UI, such as Uber, RIBs is a great pattern to use. For most other apps, smaller or has natural UI separation of concerns, RIBs is too complicated for the job.
MVC isn’t really applicable for the SwiftUI world. It is also pretty much the same as the chosen iMVVM pattern anyways. (B)VIPER is a more complex and more boilerplate version of MVx. It doesn’t really provide much advantages over iMVVM given iMVVM already has good separation of concerns.
At Uber, apps were partially built as plugins. Some features are built directly without using the plugin API. It aimed to address the distinction between “core” features and “non-core” features. The downsides of this approach are:
- Inconsistency across different parts of the app.
- Disagreements around what is core and what is non-core.
- Core features cannot be turned off. To address these issues, the proposal is to build all the features of the app as plugins!
The app contains a basic structure of starting up. From there on, all features are built as plugins. From an architectural perspective, the distinction between core and non-core is completely removed. Only the skeleton of the app remains outside of the plugin API.
Drawing from the experience at Uber, it is very difficult to reason about what the end user actually sees since all plugins are default off. And because the backing feature flag system cannot be 100% reliable, some users may be left in a broken state without us even knowing.
Instead of being pessimistic about our own code, we choose to be optimistic. All plugins are default on! Each plugin has a feature flag automatically generated on the client side. In case of outages, engineers may configure the backend feature flag service to turn off plugins. Since the plugin service is not 100% reliable, turning off plugins is a best effort operation.