HTTP Client Optimization Through Strategic Batching and Chunking Patterns

import { fetch, Agent, setGlobalDispatcher } from 'undici';
import { randomUUID } from 'node:crypto';
import { performance } from 'node:perf_hooks';

// Types
interface BatchOperation<T> {
  id: string;
  payload: T;
}

interface BatchResult<T> {
  id: string;
  status: 'success' | 'failed';
  data?: T;
  error?: string;
}

interface BatchConfig {
  url: string;
  maxChunkSize: number;
  minChunkSize: number;
  maxConcurrency: number;
  timeoutMs: number;
  retryAttempts: number;
}

export class BatchProcessor<T> {
  private config: Required<BatchConfig>;
  private agent: Agent;
  private queue: Array<{ id: string; resolve: (res: BatchResult<T>) => void; reject: (err: Error) => void; payload: T }> = [];
  private activeChunks = 0;
  private currentChunkSize: number;
  private latencyHistory: number[] = [];

  constructor(config: BatchConfig) {
    this.config = {
      maxChunkSize: config.maxChunkSize || 50,
      minChunkSize: config.minChunkSize || 10,
      maxConcurrency: config.maxConcurrency || 4,
      timeoutMs: config.timeoutMs || 5000,
      retryAttempts: config.retryAttempts || 2,
      ...config
    };

    // Explicit connection pooling for HTTP/1.1 & HTTP/2
    this.agent = new Agent({
      keepAliveTimeout: 30_000,
      keepAliveMaxTimeout: 60_000,
      pipelining: 1, // Disable pipelining for safety with batch endpoints
      connections: this.config.maxConcurrency * this.config.maxChunkSize
    });
    setGlobalDispatcher(this.agent);

    this.currentChunkSize = this.config.maxChunkSize;
  }

  async add(payload: T): Promise<BatchResult<T>> {
    return new Promise((resolve, reject) => {
      this.queue.push({ id: randomUUID(), resolve, reject, payload });
      this.processQueue();
    });
  }

  private async processQueue(): Promise<void> {
    if (this.queue.length === 0 || this.activeChunks >= this.config.maxConcurrency) return;

    const chunk = this.queue.splice(0, this.currentChunk

Size); this.activeChunks++;

try {
  await this.executeChunk(chunk);
} catch (err) {
  chunk.forEach(op => op.reject(err as Error));
} finally {
  this.activeChunks--;
  this.processQueue(); // Drain remaining queue
}

}

private async executeChunk(operations: typeof this.queue): Promise<void> { const idempotencyKey = randomUUID(); const startTime = performance.now();

for (let attempt = 0; attempt <= this.config.retryAttempts; attempt++) {
  try {
    const response = await fetch(this.config.url, {
      method: 'POST',
      headers: {
        'Content-Type': 'application/json',
        'X-Idempotency-Key': idempotencyKey,
        'X-Batch-Size': operations.length.toString()
      },
      body: JSON.stringify({
        operations: operations.map(op => ({ id: op.id, payload: op.payload }))
      }),
      dispatcher: this.agent,
      signal: AbortSignal.timeout(this.config.timeoutMs)
    });

    if (!response.ok) throw new Error(`HTTP ${response.status}`);

    const results: BatchResult<T>[] = await response.json();
    const latency = performance.now() - startTime;
    this.adaptChunkSize(latency, results);

    results.forEach(res => {
      const op = operations.find(o => o.id === res.id);
      if (op) {
        if (res.status === 'success') op.resolve(res);
        else op.reject(new Error(res.error || 'Partial failure'));
      }
    });
    return;
  } catch (err) {
    if (attempt === this.config.retryAttempts) throw err;
    await new Promise(r => setTimeout(r, 100 * (attempt + 1)));
  }
}

}

private adaptChunkSize(latency: number, results: BatchResult<T>[]): void { this.latencyHistory.push(latency); if (this.latencyHistory.length > 20) this.latencyHistory.shift();

const avgLatency = this.latencyHistory.reduce((a, b) => a + b, 0) / this.latencyHistory.length;
const errorRate = results.filter(r => r.status === 'failed').length / results.length;

// Adaptive logic: shrink if latency spikes or errors increase, grow if stable
if (avgLatency > this.config.timeoutMs * 0.7 || errorRate > 0.1) {
  this.currentChunkSize = Math.max(this.config.minChunkSize, this.currentChunkSize - 5);
} else if (avgLatency < this.config.timeoutMs * 0.3 && errorRate < 0.05) {
  this.currentChunkSize = Math.min(this.config.maxChunkSize, this.currentChunkSize + 2);
}

}

async drain(): Promise<void> { while (this.queue.length > 0 || this.activeChunks > 0) { await new Promise(r => setTimeout(r, 100)); } this.agent.close(); } }

### 2. Usage Example

```typescript
// Mock server endpoint expectation:
// POST /batch
// Body: { operations: [{ id: "uuid", payload: { ... } }] }
// Response: [{ id: "uuid", status: "success" | "failed", data?: ..., error?: ... }]

async function main() {
  const processor = new BatchProcessor<{ userId: string; action: string }>({
    url: 'http://localhost:3000/batch',
    maxChunkSize: 50,
    minChunkSize: 10,
    maxConcurrency: 4,
    timeoutMs: 3000,
    retryAttempts: 2
  });

  const promises = Array.from({ length: 1000 }, (_, i) =>
    processor.add({ userId: `user_${i}`, action: 'update_profile' })
  );

  const results = await Promise.allSettled(promises);
  const successes = results.filter(r => r.status === 'fulfilled').length;
  const failures = results.filter(r => r.status === 'rejected').length;

  console.log(`Completed: ${successes} success, ${failures} failed`);
  await processor.drain();
}

main().catch(console.error);

Pitfall Guide

Symptom	Root Cause	Troubleshooting & Fix
Memory OOM under load	Unbounded queue growth when consumer is slower than producer	Implement backpressure: drop oldest requests, reject new ones, or use a bounded AsyncQueue with await queue.push()
Silent partial failures	Batch endpoint returns 200 OK but marks individual items as failed; client treats entire batch as success	Always parse the response array. Map each id to its status. Never assume batch success without iterating results.
Rate limit spikes (429s)	Fixed chunk size sends bursts that exceed token bucket limits	Enable adaptive chunking (shown in code). Add exponential backoff with jitter: delay = Math.min(base * 2^attempt, max) * (0.5 + Math.random())
TLS handshake storms	Connection pool exhausted or keepAlive disabled, forcing repeated handshakes	Configure Agent with keepAliveTimeout: 30000, connections: maxConcurrency * maxChunkSize. Verify Connection: keep-alive headers.
Idempotency key collisions	Reusing keys across different payloads or batches causes stale responses	Generate keys per batch execution, not per operation. Use randomUUID() or deterministic hash of payload + sequence counter.
Timeout cascades	Chunk size too large for payload serialization/deserialization time	Monitor p95 latency vs timeoutMs. If serialization dominates, reduce maxChunkSize or switch to streaming/chunked transfer encoding.

Debugging Checklist:

1. Enable undici debug logging: DEBUG=undici* node app.js
2. Verify connection pool utilization: agent.stats.connected vs agent.stats.pending
3. Trace batch correlation IDs through server logs to confirm partial failure mapping
4. Use clinic.js or 0x to profile memory leaks in long-running batch processors

Production Bundle

Configuration Matrix

Parameter	Recommended Range	Tuning Guidance
maxChunkSize	20–100	Start at 50. Reduce if payload > 1MB or server deserialization is CPU-bound.
maxConcurrency	2–8	Match to server-side thread pool or event loop capacity. Avoid exceeding connections limit.
timeoutMs	2000–5000	Set to 1.5x expected p95 batch latency. Use per-request AbortSignal.timeout()
retryAttempts	1–3	Only retry on transient errors (5xx, network reset). Never retry on 4xx or idempotency violations.

Observability Setup

Integrate OpenTelemetry metrics to track batch health:

import { metrics } from '@opentelemetry/api';
const meter = metrics.getMeter('batch-processor');
const batchLatency = meter.createHistogram('batch.latency', { unit: 'ms' });
const batchSuccess = meter.createCounter('batch.success');
const batchFailure = meter.createCounter('batch.failure');

// Inside executeChunk():
batchLatency.record(latency, { chunk_size: operations.length });
results.forEach(r => r.status === 'success' ? batchSuccess.add(1) : batchFailure.add(1));

Dashboard Queries:

• rate(batch_failure_total[5m]) / rate(batch_success_total[5m]) → Alert if > 5%
• histogram_quantile(0.95, batch_latency_bucket) → Alert if > 80% of timeout
• agent_connections_total vs agent_connections_active → Detect pool starvation

Runbook & Deployment Checklist

• Verify server batch endpoint supports partial success and returns correlated id arrays
• Configure reverse proxy (nginx/envoy) to allow large request bodies (client_max_body_size / max_request_bytes)
• Set circuit breaker thresholds: open after 50% failure rate, half-open after 30s
• Implement graceful shutdown: call processor.drain() on SIGTERM to flush queue
• Load test with k6 or autocannon simulating 2x expected peak concurrency
• Monitor egress bandwidth; batch payloads should stay under 4MB to avoid TCP segmentation overhead

Batching is not a configuration toggle; it is an architectural contract. When chunking, idempotency, and observability are treated as first-class concerns, HTTP clients transition from latency liabilities to predictable, cost-efficient data pipelines.