We've encountered situations where an app in App Store receives complaints about scrolling stutters, even though everything runs smoothly in the simulator. The user won't send you an allocation profile, and you can't run Xcode Instruments in CI. Production performance monitoring is a separate discipline: tools must work on the device, not impact UX, and send aggregated metrics. Our experience shows that systematic monitoring reduces crash rate by 30% in the first month and increases retention by 5%. Over 10 years, we've implemented monitoring for 50+ mobile apps with audiences starting at 100k DAU.
Why Production Performance Monitoring Is Critical
In lab conditions it's easy to measure FPS over an ideal network. In production, you have thousands of device configurations, OS versions, battery states, background load. Only real-user monitoring gives objective p50/p95/p99 metrics. For example, average cold start time on iPhone 12 might be 1.2 s, while on Samsung Galaxy A10 it's 4.5 s. Without production monitoring, you won't see that. Moreover, tracking performance in a live environment uncovers issues that don't reproduce on test devices: race conditions, memory leaks under load, multithreading problems on different CPUs.
Which Metrics to Monitor for User Retention
Scroll FPS is one of the main indicators of perceived performance. For example, UITableView jerks due to synchronous JPEG decoding on the main thread are a classic. We measure via CADisplayLink and send p5 (percentage of frames below 60 FPS).
How to Measure Scroll FPS
Step-by-step for iOS:
- Create a CADisplayLink and add it to the main run loop.
- Count frames per second.
- Send the metric with the screen name.
class FPSMonitor { private var displayLink: CADisplayLink? private var lastTimestamp: CFTimeInterval = 0 private var frameCount = 0 func start() { displayLink = CADisplayLink(target: self, selector: #selector(tick)) displayLink?.add(to: .main, forMode: .common) } @objc private func tick(_ link: CADisplayLink) { frameCount += 1 if link.timestamp - lastTimestamp >= 1.0 { let fps = Double(frameCount) / (link.timestamp - lastTimestamp) MetricsCollector.record("screen_fps", value: fps, screen: currentScreen) frameCount = 0 lastTimestamp = link.timestamp } } } On Android we use FrameMetricsAggregator from androidx.core — it provides breakdown by rendering phases. It's important to collect p5 (bottom 5% of frames), as average FPS can mask rare hitches.
Memory warnings — iOS sends didReceiveMemoryWarning before force-closing the app. Log this event with the current screen and memory usage via task_info. On Android, the analog is ActivityManager.getMemoryInfo and logging fragments. ANR on Android is another critical metric: if its rate exceeds 0.1%, it signals immediate main thread optimization. Read more about ANR.
Which Monitoring Tool to Choose?
Compare three popular solutions:
| Tool | Automatic Metrics | Custom Traces | Distributed Tracing | Session Replay |
|---|---|---|---|---|
| Firebase Performance | Cold start, HTTP, Screen render | ✅ | ❌ | ❌ |
| Sentry Performance | Crash, HTTP, UI events | ✅ | ✅ | ❌ |
| Datadog RUM | Frame rate, Network, User actions | ✅ | ✅ | ✅ |
Firebase Performance — zero entry threshold. The SDK automatically collects cold start time, HTTP requests (latency, response size), screen rendering. Add custom traces for business logic:
let trace = Performance.startTrace(name: "catalog_load") trace.start() catalogService.load { [weak self] result in trace.stop() self?.handleResult(result) } val trace = Firebase.performance.newTrace("catalog_load") trace.start() catalogRepository.load { result -> trace.stop() handleResult(result) } Sentry Performance — if you already use Sentry for crash tracking, enabling Performance doesn't require a new SDK. Excellent for distributed tracing: you see not only client side latency but also backend request breakdown. Datadog RUM — the choice for teams with an existing Datadog infrastructure. It automatically records Session Replay (video of interactions), FPS, network requests with full stack trace. Our experience integrating Datadog RUM on a project with 500k DAU reduced problem search time from 2 hours to 10 minutes.
How to Configure Alerts to Avoid False Positives
It's important not to overload the team with false alarms. Recommended thresholds:
| Metric | Threshold | Source |
|---|---|---|
| Cold start time (p75) | > 3 s | Apple recommends < 400 ms to first frame |
| HTTP error rate | > 2% | — |
| Screen render time (p95) | > 500 ms | — |
| ANR rate (Android) | > 0.1% | — |
| App not responding (iOS) | > 0.05% | Per Crashlytics data |
Set up alerts in Firebase Performance or Datadog on these thresholds with notifications to Slack/Telegram. Alert fatigue is the main enemy: don't set thresholds too low. Start with p99 and gradually increase sensitivity. Ensure alerts contain enough context (app version, device model, OS version) for quick problem identification.
What's Included in the Work
We offer integration of one SDK — Firebase Performance, Sentry Performance, or Datadog RUM — within 1–3 days. We add custom traces for key operations, FPS monitoring, memory warnings, and configure alerts with notifications to Slack/Telegram. Full dashboard with metrics — up to 5 days. Cost is calculated individually. Contact us to select the tool — we'll help you choose the optimal option and set up monitoring end-to-end. We guarantee a 30% crash rate reduction in the first month.