apollo-performance-tuning

Overview

Optimize Apollo.io API performance through response caching, connection pooling, bulk operations, parallel fetching, and result slimming. Key insight: search is free but slow (~500ms), enrichment costs credits — cache aggressively and batch enrichment calls.

Prerequisites

• Valid Apollo API key
• Node.js 18+

Instructions

Step 1: Connection Pooling

Reuse TCP connections to avoid TLS handshake overhead on every request.

// src/apollo/optimized-client.ts
import axios from 'axios';
import https from 'https';

const httpsAgent = new https.Agent({
  keepAlive: true,
  maxSockets: 10,
  maxFreeSockets: 5,
  timeout: 30_000,
});

export const optimizedClient = axios.create({
  baseURL: 'https://api.apollo.io/api/v1',
  headers: { 'Content-Type': 'application/json', 'x-api-key': process.env.APOLLO_API_KEY! },
  httpsAgent,
  timeout: 15_000,
});

Step 2: Response Caching with Per-Endpoint TTLs

// src/apollo/cache.ts
import { LRUCache } from 'lru-cache';

// Different TTLs based on data volatility
const CACHE_TTLS: Record<string, number> = {
  '/organizations/enrich': 24 * 60 * 60 * 1000,    // 24h — company data rarely changes
  '/people/match': 4 * 60 * 60 * 1000,              // 4h — contact data changes occasionally
  '/mixed_people/api_search': 15 * 60 * 1000,       // 15min — search results are dynamic
  '/mixed_companies/search': 30 * 60 * 1000,         // 30min — company search
  '/contact_stages': 60 * 60 * 1000,                 // 1h — stages rarely change
};

const cache = new LRUCache<string, { data: any; at: number }>({
  max: 5000,
  maxSize: 50 * 1024 * 1024,
  sizeCalculation: (v) => JSON.stringify(v).length,
});

function cacheKey(endpoint: string, params: any): string {
  return `${endpoint}:${JSON.stringify(params)}`;
}

export async function cachedRequest<T>(
  endpoint: string,
  requestFn: () => Promise<T>,
  params: any,
): Promise<T> {
  const key = cacheKey(endpoint, params);
  const ttl = CACHE_TTLS[endpoint] ?? 15 * 60 * 1000;
  const cached = cache.get(key);

  if (cached && Date.now() - cached.at < ttl) return cached.data;

  const data = await requestFn();
  cache.set(key, { data, at: Date.now() });
  return data;
}

export function getCacheStats() {
  return { entries: cache.size, sizeBytes: cache.calculatedSize };
}

Step 3: Use Bulk Endpoints Over Single Calls

Apollo's bulk enrichment endpoint handles 10 records per call vs 1. Massive performance gain.

// src/apollo/bulk-ops.ts
import { optimizedClient } from './optimized-client';
import PQueue from 'p-queue';

const queue = new PQueue({ concurrency: 3, intervalCap: 2, interval: 1000 });

// Enrich 100 people: 100 individual calls = 100 requests @ 500ms = 50s
// Batch of 10: 10 bulk calls @ 600ms = 6s (8x faster, same credits)
export async function batchEnrich(
  details: Array<{ email?: string; linkedin_url?: string; first_name?: string; last_name?: string; organization_domain?: string }>,
): Promise<any[]> {
  const results: any[] = [];

  for (let i = 0; i < details.length; i += 10) {
    const batch = details.slice(i, i + 10);
    const result = await queue.add(async () => {
      const { data } = await optimizedClient.post('/people/bulk_match', {
        details: batch,
        reveal_personal_emails: false,
        reveal_phone_number: false,
      });
      return data.matches ?? [];
    });
    results.push(...(result ?? []));
  }

  return results;
}

Step 4: Parallel Search with Concurrency Control

export async function parallelSearch(
  domains: string[],
  concurrency: number = 5,
): Promise<Map<string, any[]>> {
  const searchQueue = new PQueue({ concurrency });
  const results = new Map<string, any[]>();

  await searchQueue.addAll(
    domains.map((domain) => async () => {
      const data = await cachedRequest(
        '/mixed_people/api_search',
        () => optimizedClient.post('/mixed_people/api_search', {
          q_organization_domains_list: [domain],
          person_seniorities: ['vp', 'director', 'c_suite'],
          per_page: 25,
        }).then((r) => r.data),
        { domain },
      );
      results.set(domain, data.people ?? []);
    }),
  );

  return results;
}

Step 5: Slim Response Payloads

Apollo returns large person objects (~2KB each). Extract only needed fields to reduce memory.

interface SlimPerson {
  id: string;
  name: string;
  title: string;
  email?: string;
  company: string;
  seniority: string;
}

function slimPerson(raw: any): SlimPerson {
  return {
    id: raw.id,
    name: raw.name,
    title: raw.title,
    email: raw.email,
    company: raw.organization?.name ?? '',
    seniority: raw.seniority ?? '',
  };
}

// Use immediately after API call to free memory
const { data } = await optimizedClient.post('/mixed_people/api_search', { ... });
const slim = data.people.map(slimPerson);  // ~200 bytes each instead of ~2KB

Step 6: Benchmark Your Endpoints

async function benchmark(includePaidEndpoints = false) {
  const endpoints = [
    { name: 'People Search', fn: () => optimizedClient.post('/mixed_people/api_search',
        { q_organization_domains_list: ['apollo.io'], per_page: 1 }) },
    { name: 'Auth Health', fn: () => optimizedClient.get('/auth/health') },
  ];

  // Paid enrichment benchmarks require an approved, budgeted test run.
  if (includePaidEndpoints) {
    endpoints.push({ name: 'Org Enrich', fn: () => optimizedClient.get('/organizations/enrich',
      { params: { domain: 'apollo.io' } }) });
  }

  for (const ep of endpoints) {
    const times: number[] = [];
    for (let i = 0; i < 5; i++) {
      const start = Date.now();
      try { await ep.fn(); } catch {}
      times.push(Date.now() - start);
    }
    const avg = Math.round(times.reduce((a, b) => a + b) / times.length);
    const p95 = times.sort((a, b) => a - b)[Math.floor(times.length * 0.95)];
    console.log(`${ep.name}: avg=${avg}ms, p95=${p95}ms`);
  }
}

Output

• Connection pooling with keepAlive and configurable maxSockets
• LRU cache with per-endpoint TTLs (24h org, 4h contact, 15m search)
• Bulk enrichment via /people/bulk_match (10x fewer requests)
• Parallel search with p-queue concurrency control
• Response slimming reducing memory from ~2KB to ~200B per person
• Benchmarking script measuring avg and p95 latency

Examples

For a search-latency regression, capture a baseline with the free search and health endpoints against a mock or approved sandbox fixture, then introduce connection reuse and a bounded cache behind a feature flag. Compare p95, error rate, cache-hit behavior, and result equivalence before promoting the change. A paid-enrichment benchmark requires an explicit approval flag, a fixed credit budget, and a public organization fixture; otherwise it remains out of the run. Roll back the flag if latency improves at the cost of stale, incorrect, or over-broadly cached results.

Error Handling

| Issue | Resolution | |-------|------------| | High latency | Enable connection pooling, check for stale cache | | Cache misses | Increase TTL for stable data (org enrichment) | | Rate limits with parallelism | Reduce p-queue concurrency | | Memory growth | Lower LRU max entries, slim response payloads |

Resources

• Bulk People Enrichment
• Node.js HTTPS Agent
• LRU Cache
• p-queue

Next Steps

Proceed to apollo-cost-tuning for cost optimization.