HTTP Request Monitoring: Response Time and Error Rate

Note: when the average response time of the GET /api/feed endpoint grows from 200 ms to 1800 ms, users feel the lag, but reviews in the App Store appear a day later. With monitoring, you learn about the problem within 5 minutes. Our team, with over 10 years of mobile development experience, has seen situations where lack of HTTP request monitoring led to losing up to 30% of users. We have implemented more than 50 projects with monitoring—from startups to enterprise apps with 10 million users. That's why we deploy turnkey systems for collecting Response Time and Error Rate metrics. According to analysts, every hour of downtime with high latency costs companies an average of $10,000. A full monitoring setup costs from $1,500 for a basic Firebase integration to $10,000 for a custom enterprise system, but it quickly pays for itself by preventing that hourly loss. Let's break down how this works in practice and which approaches deliver real results.

Problems Without HTTP Request Monitoring

Without monitoring, you only learn about API problems from user reviews or complaints. But complaints come with delay, and every hour of high-latency downtime means losing part of your audience. Monitoring solves three tasks: detecting degradation within 5 minutes, diagnosing the cause (network vs. server), and preventing regressions after releases.

Client-Side vs Server-Side Monitoring: Which to Choose?

Two levels of metric collection provide a complete picture.

Client-side (in-app): response time from the user's device—includes network latency. Reflects real experience but is noisy: one user's poor network doesn't mean a server problem.

Server-side (APM): instrument the server, measure only processing time. Doesn't see network latency but accurately shows backend state.

Correct approach is both levels. Client-side for understanding UX, server-side for diagnostics.

Choice depends on goal. If you need real UX—client-side is mandatory. For backend diagnostics, server-side suffices. Optimal combination: client metrics for alerts that 'user is suffering', server metrics for root cause.

How to Instrument the HTTP Client: Code and Examples

Let me show using React Native. The code is analogous for iOS (URLProtocol) and Android (OkHttpInterceptor).

Axios Interceptor in React Native

import axios, { AxiosInstance, AxiosRequestConfig, AxiosResponse } from 'axios'; type RequestMetric = { endpoint: string; method: string; statusCode: number; durationMs: number; timestamp: number; error?: string; }; const metricsBuffer: RequestMetric[] = []; const FLUSH_INTERVAL_MS = 30_000; const FLUSH_BATCH_SIZE = 50; function createMonitoredAxios(): AxiosInstance { const instance = axios.create({ baseURL: API_BASE_URL }); instance.interceptors.request.use((config: AxiosRequestConfig) => { (config as any).metadata = { startTime: Date.now() }; return config; }); instance.interceptors.response.use( (response: AxiosResponse) => { recordMetric(response.config, response.status, null); return response; }, (error) => { const status = error.response?.status ?? 0; recordMetric(error.config, status, error.message); return Promise.reject(error); } ); return instance; } function recordMetric(config: any, status: number, error: string | null) { const durationMs = Date.now() - (config?.metadata?.startTime ?? Date.now()); const url = config?.url ?? 'unknown'; const endpoint = new URL(url, API_BASE_URL).pathname; metricsBuffer.push({ endpoint, method: (config?.method ?? 'GET').toUpperCase(), statusCode: status, durationMs, timestamp: Date.now(), error: error ?? undefined, }); if (metricsBuffer.length >= FLUSH_BATCH_SIZE) flushMetrics(); } 

We normalize URL to pathname—we don't want thousands of unique metrics like /api/users/123, /api/users/456. A pattern like /api/users/:id is needed.

Client-Side Aggregation: P50/P95/P99

Mean response time is deceptive: 90% of requests at 100 ms and 10% at 5000 ms give an average of 590 ms—doesn't reflect reality. Percentiles are more accurate:

function calculatePercentiles(durations: number[]): { p50: number; p95: number; p99: number } { const sorted = [...durations].sort((a, b) => a - b); const p = (percentile: number) => sorted[Math.floor(sorted.length * percentile / 100)]; return { p50: p(50), p95: p(95), p99: p(99) }; } 

P99 is the response time for 99% of requests. If P99 rises while P50 is stable, there's a problem with slow requests from a small subset of users (specific endpoint, specific OS, specific region).

Sending Metrics: Batching and Prioritization

async function flushMetrics() { if (metricsBuffer.length === 0) return; const batch = metricsBuffer.splice(0, FLUSH_BATCH_SIZE); try { await fetch(`${METRICS_ENDPOINT}/ingest`, { method: 'POST', headers: { 'Content-Type': 'application/json' }, body: JSON.stringify({ metrics: batch, appVersion: APP_VERSION, platform: Platform.OS }), }); } catch { // On failure, put back into buffer but not more than MAX_BUFFER_SIZE metricsBuffer.unshift(...batch.slice(0, MAX_BUFFER_SIZE - metricsBuffer.length)); } } 

We send metrics via a special non-critical fetch—errors should not affect UX. The buffer is limited—no gigabytes accumulated during extended offline.

Ready-Made Solutions: Firebase Performance Monitoring

@react-native-firebase/perf automates most of the work: intercepts fetch/XHR, measures time, sends to Firebase. The console provides a dashboard with percentiles by endpoint. According to Firebase Performance Monitoring documentation, automatic network metric collection is available on Android and iOS.

import perf from '@react-native-firebase/perf'; // Custom trace for a critical operation const trace = await perf().startTrace('checkout_flow'); trace.putAttribute('userId', userId); // ... operation ... await trace.stop(); 

For most apps, Firebase Performance Monitoring is the right choice. For enterprises with self-hosted requirements—Datadog RUM Mobile or custom sending to InfluxDB/Prometheus.

Percentiles: Why This Is the Key Metric?

Average response time might be acceptable, but the 10% slowest requests make the app feel sluggish. P95 and P99 show how much users with poor connections or older devices actually wait. Setting alerts on P95 lets you react to degradation for the majority, and P99 for outliers.

How to Set Up Alerts and Not Miss Degradation?

Define a baseline—weekly average of P95. Set thresholds: P95 > 2× baseline for 5 minutes—warning in Slack. Error Rate > 5% for 10 minutes—critical alert in PagerDuty. It's important to separate client and server metrics: client P95 rise with stable server P95 indicates network issues.

Firebase Performance is integrated 3x faster than a custom Datadog solution—3 days vs 2 weeks. But custom solutions give full control over storage and dashboard customization. The investment in monitoring pays off by reducing user churn and support costs, saving up to 30% of budget.

What's Included

• Instrumentation of HTTP client (Axios/OkHttp/URLSession) for your platform
• Setup of metrics dashboard (Grafana, Firebase Console, or Datadog)
• Scripts for batching and sending metrics with error handling
• Documentation on alerts and thresholds
• Training for the operations team

Process and Timeline

1. Analysis: audit current stack, identify critical endpoints
2. Design: choose approach (client/server), tools, metric schema
3. Implementation: integrate interceptors, implement batching and aggregation logic
4. Testing: verify in real-world scenarios, stress-test buffer
5. Deployment: release, monitor first days, calibrate thresholds

Approximate timelines: Firebase Performance with custom traces and basic alerts—from 1 week. Custom metric system with batching, percentiles, and Datadog/Grafana dashboard—from 2 to 4 weeks. Cost is calculated individually—depends on platform, number of endpoints, and integration complexity. We guarantee integration quality and prompt incident response.