Case Study: ESP-Point — IoT Hardware + iOS Sync, Offline-First Architecture

Problem: Bidirectional realtime sync between an ESP32 microcontroller (DIY IoT device) and an iOS app: the device uploads sensor data, and the app sends commands + firmware updates to the device. Offline-first is mandatory — the device is used in locations without internet access.

Result: 50ms BLE handshake median, 3-day offline survival (local cache + delayed sync), OTA firmware update with a 99.2% success rate, 12K devices sold.

1. Architecture

1[ESP32 Device]
2    ↓ BLE (peripheral)
3[iOS App — central]
4    ↓ Wi-Fi when available
5[Firebase Realtime DB + Cloud Storage]
6    ↓ sync queue
7[Web Dashboard for fleet ops]

The iOS app acts as a proxy / router: a BLE \<-\> Cloud bridge.

2. BLE State Machine: The Real Issue

BLE doesn't have just one "easy demo." State machine fragmentation is the real production issue:

States:

• unpairedScanning
• pairing
• paired (connected)
• paired-disconnected (cached)
• firmwareUpdating
• error (reachable from every state)

For each transition, we resolved 2-3 race conditions in iOS CBCentralManager callbacks:

1. didDisconnect + didFailToConnect overlap — UUID-based dedup
2. scanForPeripherals background restoration — set CBCentralManagerOptionRestoreIdentifierKey
3. On the Bluetooth state transition .poweredOff → .poweredOn, scanning doesn't auto-resume — explicit re-call to scanForPeripherals

Result: 50ms handshake median. Connection drop rate 0.8% (industry standard 4-6%).

3. Offline Storage: Realm + Sync Queue

3-day offline target. Every sensor reading (hourly) is written to device storage, then pushed to a sync queue in Realm:

1class SyncQueue: Object {
2    @Persisted(primaryKey: true) var id: ObjectId
3    @Persisted var payload: Data
4    @Persisted var createdAt: Date
5    @Persisted var retryCount: Int = 0
6    @Persisted var lastError: String?
7}

Sync worker:

1final class SyncWorker {
2    func runOnce() async {
3        let realm = try await Realm()
4        let pending = realm.objects(SyncQueue.self)
5            .where { $0.retryCount < 5 }
6            .sorted(by: \.createdAt)
7            .prefix(50)
8 
9        for item in pending {
10            do {
11                try await uploadToFirebase(item.payload)
12                try realm.write { realm.delete(item) }
13            } catch {
14                try realm.write {
15                    item.retryCount += 1
16                    item.lastError = error.localizedDescription
17                }
18            }
19        }
20    }
21}

The sync worker is triggered via NWPathMonitor network reachability. In the background, the app retries via BGAppRefreshTask starting day 1.

4. OTA Firmware Update

8-12MB binary blob transfer over BLE:

• Bluetooth chunk size: 244 bytes (BLE 5 max safe)
• Sequence number + CRC32 per chunk
• Acknowledgment per 10-chunk batch
• Total transfer: 4-6 minutes (for 10MB firmware)
• Resume on disconnect support

Success rate: 99.2%. Failure modes:

• 0.5%: connection drop right before the last chunk
• 0.2%: device flash corruption (re-upload OK)
• 0.1%: power loss at a critical moment (manual recovery)

The OTA UI matters — progress bar + "do not close app" warning + battery indicator. +18% satisfaction in App Store reviews.

5. Test Strategy

• Unit tests: Sync queue state machine, BLE manager state transitions (mocked CBCentralManager)
• Integration tests: Realm migration, sync conflict resolution
• Device farm: 6 physical ESP32 units + 4 iPhone devices (iOS 16-18) — physical testing in the CI pipeline
• Snapshot tests: UI state for each BLE state

Coverage: Domain layer 92%, BLE manager 78%, UI 65%. Total 78%.

6. iOS App Architecture

• SwiftUI iOS 16+
• BLEManager: @MainActor actor (Bluetooth callbacks run on the main thread)
• DataLayer: Realm primary + Firebase secondary (offline-first)
• Sync orchestrator: BG task + foreground refresh

Memory baseline: 105MB. Peak (during OTA): 220MB (firmware buffer).

7. Lessons

1. Physical BLE testing is mandatory. The simulator can't simulate BLE. A physical device in CI is a must.

2. CBCentralManagerOptionRestoreIdentifierKey is critical for background relaunch — without it, the connection is lost after the app is killed.

3. State machine documentation. Every transition of the 6-state machine is logged for production debugging.

4. Realm migration empties the sync queue on schema changes (this wasn't validated with a test → 200 users lost data).

5. Don't give users an OTA "abort" button — risk of flash corruption. Pause/resume only.

6. Firebase rules can throttle offline write throughput — requireAuth: false write rule + custom token validation.

7. Battery profiling. Constant BLE scanning = 8 hours of battery life. Connection-on-demand pattern saves 35%.

Result: IoT + iOS sync is 10x harder than the "happy path" demos suggest. State machine + offline storage + OTA = 3 months of technical debt if not planned upfront.