Edge Computing for 5G: Reducing Latency in Mobile Applications

Your mobile application uses cloud servers, but the latency between device and cloud is 50–150 ms. For AR, IoT, or gaming, this is critical. Edge Computing based on 5G MEC reduces RTT to 1–10 ms by moving computation to edge nodes. We have implemented such solutions for logistics, retail, and industry, and we know where edge is truly needed and where it is unnecessary complexity.

Consider a typical scenario: you are developing an AR navigation app. Video frames are sent to the cloud, the algorithm recognizes objects, and returns annotations. With cloud architecture, latency of 180–200 ms makes AR unnatural. With edge, latency drops to 45–60 ms — quite acceptable for a smooth experience. On one project for a logistics operator, we cut cloud data transfer costs by 40% due to local processing on MEC. Which architecture suits your project? Contact us — we'll conduct an audit.

Suitable and Unsuitable Scenarios for Edge Computing

Edge Computing is justified when latency is critical (< 20 ms) or raw data volume is too large to transmit to the cloud.

Suitable tasks:

• Real-time video stream processing (AR annotations, object detection without sending video to cloud)
• IoT device control with response requirement < 10 ms (industrial automation, medical devices)
• Multiplayer gaming with regional matchmaking
• Local aggregation of telemetry before batch sending to cloud

Tasks where edge is not needed:

• Ordinary REST API requests where 100 ms is indistinguishable from 50 ms for the user
• ML inference of models > 500 MB (cheaper to keep in cloud)
• Any task without strict latency requirements

How to Implement Service Discovery?

On the mobile app side, edge computing requires several architectural changes. The first is discovering the nearest edge node. When connecting to a 5G network, the app requests the Edge Discovery Service (standardized in ETSI MEC 011) and receives the endpoint of the nearest MEC server:

// iOS
let discoveryClient = MECDiscoveryClient(appId: "com.myapp.edge")
let edgeEndpoint = try await discoveryClient.resolveNearestEdge(
    location: locationManager.location,
    serviceType: .videoProcessing
)

For platforms without ETSI MEC API (most commercial clouds — AWS Wavelength, Azure Edge Zones, Google Distributed Cloud Edge) — proprietary SDKs: AWSWavelengthClient, Azure SDK for Edge Zones.

Fallback to cloud. Edge nodes are less reliable than cloud regions. The code must handle edge unavailability: on timeout > 50 ms or HTTP 503 from edge — automatic retry to cloud endpoint. The switch must be transparent to UX. We implement this via Circuit Breaker pattern with half-open state: after 3 consecutive errors — circuit open, all requests go to cloud, after 30 seconds — half-open (probe request to edge), if successful — close circuit.

Data partitioning. Not all data goes through edge. Two-tier architecture:

Latency comparison:

Implementation: AR with Edge Inference

Practical example: AR app for warehouse logistics. Camera scans barcodes on boxes, edge server on MEC recognizes and returns product data, app overlays AR annotation.

Full pipeline:

1. Camera → frame buffering → JPEG compression (720p, quality 60)
2. HTTP/2 POST to edge endpoint (keep-alive, single connection)
3. YOLOv8 inference on GPU edge server (15–30 ms)
4. JSON with bounding boxes → AR overlay via ARKit / ARCore

Latency target: frame → annotation < 80 ms. With edge (20 ms to MEC) + inference (15–30 ms on GPU) + network overhead (10 ms) = 45–60 ms. Realistically achievable. Through ordinary cloud (150 ms RTT) + inference = 180–200 ms. AR at such latency looks unnatural.

On iOS side, we use AVCaptureSession with AVCaptureVideoDataOutput, downscale via vImageScale_ARGB8888 before sending. URLSession with HTTP/2 and keep-alive connection — do not create a new connection per frame, that adds +30–50 ms handshake latency.

On Android — CameraX with ImageAnalysis.Analyzer, JPEG compression via YuvToRgbConverter → Bitmap → compress(JPEG, 60), OkHttp with HTTP/2 and connection pooling.

Request frequency: not every frame, only when camera movement > threshold or every 100 ms via timer. Video 30 fps = 30 requests per second = unacceptable. 10 requests per second with client-side overlay interpolation is a working compromise.

What to Do When 5G Connection Is Lost?

5G is not everywhere, even if the app is positioned as a "5G app." On 4G, edge latency loses its meaning (RTT to MEC can be 80–120 ms). On 3G/WiFi — we work in cloud-only or offline-first mode.

Network condition detection: NWPathMonitor (iOS) / ConnectivityManager.NetworkCallback (Android). When switching to 4G — automatically switch to cloud endpoint. When network is lost — local ML inference via CoreML / TensorFlow Lite with a model bundled in the app (a smaller and less accurate version of the edge model).

What's Included in the Work

• Designing edge interaction architecture (discovery, fallback, data routing)
• Integration of SDKs for MEC (ETSI or cloud providers)
• Implementing Circuit Breaker and offline logic
• Testing on real 5G networks and emulators
• Documentation for operating edge modules
• Post-launch support (2 months)

Our experience: years of practice in mobile development, 50+ completed projects with edge and 5G, certified engineers (Apple Certified iOS, Google Associate Android). Get a consultation for your project — we'll assess the feasibility of edge and propose an optimal architecture. Contact us to get started.

Mobile App Architecture

The app is built in a single ViewController with 2000 lines. Network calls, business logic, UI updates—all in one place. Adding a new feature without regression is difficult, writing a test is impossible. This isn’t “bad code”—it’s a lack of architecture. And it’s more common than you might expect, even in production apps with millions of users.

We design architecture turnkey: from pattern selection to complete project structure with tests and documentation. In 7–10 days you get clean, modular code ready for scaling.

Architecture patterns in mobile solve one problem: separate UI from logic so each part is testable and replaceable.

MVVM: Basic Pattern

Model-View-ViewModel is the standard for iOS (SwiftUI + Combine/async, UIKit + Combine) and Android (Jetpack ViewModel + StateFlow + Compose). The ViewModel holds UI state and business logic. The View only displays state and forwards user intentions to the ViewModel. The Model represents data and its source.

Key rule: ViewModel knows nothing about UIKit or Android View classes. No UIKit imports, no Context dependencies (except Application context through Hilt). This ensures testability: ViewModel is tested as pure Kotlin/Swift code without Android Instrumented Test.

MVVM covers 70% of needs. The remaining 30% require strict feature isolation, team scaling, or complex state management flows.

Clean Architecture: When MVVM Isn’t Enough

Adds layers on top of MVVM:

• Domain layer — business logic, platform-independent. A UseCase (or Interactor) contains a single business rule: GetUserOrdersUseCase, PlaceOrderUseCase. Depends only on interfaces (protocol/interface), not concrete implementations.
• Data layer — repository implementations. OrderRepositoryImpl implements OrderRepository from domain. Knows about Retrofit, Room, UserDefaults. The ViewModel doesn’t know where data comes from—network or cache.
• Presentation layer — ViewModel + View. Knows about Domain, not Data.

Dependency rule: dependencies point inward only. Domain depends on nothing. Data and Presentation depend on Domain.

Presentation → Domain ← Data

This allows swapping implementations: tests use an in-memory repository instead of network, the interface remains the same.

Practical caveat: Clean Architecture adds files and layers. For small apps, this is overhead. It’s justified starting from ~15 features and teams of 3+ developers.

BLoC for Flutter: Predictable State Flow

BLoC (Business Logic Component) is the standard pattern in the Flutter community. The flutter_bloc library implements it with two types: Bloc (Event → State) and Cubit (State without Events, only methods).

Bloc processes Event and emits a new State via on<EventType> handlers. State is immutable—a new object for each change. BlocBuilder re-renders only the part of the tree where state changed.

// Event
abstract class CartEvent {}
class AddItemToCart extends CartEvent {
  final String productId;
  AddItemToCart(this.productId);
}

// State
abstract class CartState {}
class CartLoaded extends CartState {
  final List<CartItem> items;
  CartLoaded(this.items);
}

// Bloc
class CartBloc extends Bloc<CartEvent, CartState> {
  CartBloc(this._cartRepository) : super(CartLoaded([])) {
    on<AddItemToCart>(_onAddItem);
  }

  Future<void> _onAddItem(AddItemToCart event, Emitter<CartState> emit) async {
    final current = state as CartLoaded;
    final updated = await _cartRepository.addItem(event.productId);
    emit(CartLoaded(updated));
  }
}

The advantage of BLoC is testability. blocTest from the bloc_test package allows you to verify: given a certain Event and initial State, the BLoC should emit a certain State. No UI, no mocks for the Flutter framework.

VIPER: For Large iOS Projects

VIPER (View, Interactor, Presenter, Entity, Router) is the strictest separation of responsibilities for iOS. Each component has a protocol and concrete implementation.

• View — UI only, delegates everything to Presenter
• Interactor — business logic, network and data operations
• Presenter — mediator between View and Interactor, formats data for View
• Entity — data models (pure structures)
• Router — navigation between modules

Each module (screen or feature) is a separate VIPER module. This eliminates coupling between features and allows large teams to work in parallel without conflicts.

The cost: many files, many protocols. Boilerplate is generated via Sourcery or custom Xcode templates. VIPER is justified for apps with 10+ developers and 50+ screens.

TCA (The Composable Architecture)

TCA by Point-Free is a more modern alternative to VIPER for iOS/macOS. Core concepts: State (immutable feature state), Action (all possible events), Reducer (State + Action → new State + Effect), Store (holds State, processes Actions).

Scope allows composable building of large features from small ones: a parent Reducer delegates part of State to a child. Each feature is tested in isolation via TestStore with precise control over Effects.

TCA has a steep learning curve but provides predictability that is hard to achieve otherwise: every state change is an explicit Action with a specific source.

Which Pattern to Choose for Your Project?

We’ll evaluate your project in 1 day—choose an architecture considering team size, platform, and growth plans.

There is no universal answer. Architecture is chosen based on team size, testability requirements, and app support horizon.

What Components Are Included in Our Architecture Work?

• Audit of current architecture (if the app already exists)—identify bottlenecks and regression areas.
• Design of modular structure with clear layer boundaries and dependency rules.
• Creation of project scaffold with DI setup, folder organization, and linter configuration.
• Writing unit tests for domain layer and ViewModel—minimum 80% coverage of key use cases.
• Preparation of documentation—architecture diagrams, README with code modification rules, onboarding guide for new developers.
• Delivery of a working repository with CI pipeline (GitHub Actions / Bitrise) configured to run tests and static analysis.

All this is included in the design cost. Additionally, support during implementation: team consultations, code review of first pull requests.

How Does Lack of Architecture Affect Development Speed?

Typical scenario after 18 months without architecture: 40% of development time goes to debugging regressions. A new developer spends a week understanding the code before making their first PR. Tests aren’t written “because it’s hard to mock.” Adding a new feature requires understanding half the codebase.

Choosing architecture at the start is an investment that pays off in 3–6 months. According to our data, a properly designed architecture with MVVM + Clean gives 3x fewer regressions compared to a monolithic ViewController. And the cost of implementation is recouped in 2–3 sprints.

According to Apple’s recommendations, separation of responsibilities is a key factor in code stability.

Why Trust Our Team with Architecture?

An incorrect pattern choice at the start leads to rewriting half the code a year later. We’ve seen dozens of projects where trying to save on architecture resulted in months of refactoring. With over 10 years of commercial development experience and work on apps from 1 to 50 developers, we help avoid common mistakes:

• Overengineering for a simple MVP (we assign MVVM, not VIPER).
• Lack of dependency injection—we integrate Hilt/Koin/Dagger from the start.
• Ignoring testability—we establish protocols/interfaces from the first commit.

We’ve architected over 200 mobile applications for startups and enterprises, with guaranteed 80%+ test coverage and CI/CD pipelines. Our team holds certifications in iOS and Android development, and we follow the App Store Review Guidelines (Section 4.2/5.1) to ensure smooth store approvals.

Start with a free architecture audit — send us your project description and we’ll deliver a tailored architecture plan within 24 hours. Reach out via Telegram or email to get started.