Imported from jeremylongshore/tons-of-skills-marketplace (skills/.curated/apify-performance-tuning/SKILL.md). Install upstream with npx skills add jeremylongshore/tons-of-skills-marketplace --skill apify-performance-tuning. Copyright stays with the author (MIT).

Apify Performance Tuning

Overview

Optimize Apify Actors for speed, cost, and reliability. Covers Crawlee concurrency settings, memory profiling, proxy rotation strategies, request batching, and crawler selection for different workloads.

The workflow is a repeatable loop: measure a baseline, apply one lever, re-measure. The single highest-impact lever is usually crawler choice — swapping a browser crawler for CheerioCrawler on non-JS pages is a 5-10x speedup on its own. The full six-step walkthrough, with every code block, lives in references/implementation.md.

Prerequisites

• Existing Actor with measurable baseline performance
• Understanding of apify-sdk-patterns
• Access to Actor run stats in Apify Console
• APIFY_TOKEN in the environment for reading run stats via ApifyClient

Instructions

Work the levers in order. Each step is expanded — with copy-paste code — in the reference file linked below.

1. Measure a baseline. Pull runTimeSecs, requestsFinished, memAvgBytes, and usageTotalUsd from the run stats before changing anything. You cannot judge an optimization without a before number.
2. Choose the right crawler. HttpCrawler/CheerioCrawler for static HTML or JSON (low memory, fast); PlaywrightCrawler/PuppeteerCrawler only when the page genuinely needs JavaScript rendering.
3. Tune concurrency. Raise maxConcurrency for Cheerio (up to ~50); keep it low (~3-5) for browser crawlers because each page costs ~200MB. Let autoscaledPoolOptions adjust within the band.
4. Optimize memory. Push data immediately instead of accumulating arrays; for browser crawlers, block images/CSS/fonts in preNavigationHooks and cap concurrent browsers.
5. Right-size the memory allocation. Compute units bill on memory x duration — start low (512 MB for Cheerio) and only raise it if the Actor is memory-starved.
6. Rotate proxies and tune requests. Start on datacenter proxies, fall back to residential on 403/blocked; use a session pool for IP rotation and ban detection.

Full step-by-step walkthrough with all code: references/implementation.md.

The minimal starting skeleton — swap a browser crawler for Cheerio and push immediately:

import { CheerioCrawler } from 'crawlee';
import { Actor } from 'apify';

const crawler = new CheerioCrawler({
  maxConcurrency: 50,             // Cheerio is cheap — parallelize hard
  maxRequestsPerMinute: 300,      // But cap the rate to protect the target
  requestHandler: async ({ $, request }) => {
    await Actor.pushData({ url: request.url, title: $('title').text().trim() });
  },
});

Output

Applying this skill produces:

• A baseline vs. tuned metrics comparison (pages/min, avg/max memory, compute units, cost/run) drawn from the run stats.
• A crawler and concurrency recommendation matched to whether the target pages need JS rendering.
• A memory allocation value sized to the workload, with the compute-unit cost tradeoff made explicit.
• A proxy and session strategy (datacenter-first with residential fallback) for reliability under anti-bot blocking.

Instrument the running crawl to confirm the gains in real time — see references/monitoring.md for the throughput logger and a before/after impact table.

Error Handling

Examples

Slow Playwright crawl on static pages. The Actor renders every page in a browser at 3 pages/min. The pages are server-rendered HTML, so switch to CheerioCrawler and raise maxConcurrency — ~30 pages/min (10x). Code: references/implementation.md Steps 1-2.

Out-of-memory crashes under load. A PlaywrightCrawler at maxConcurrency: 20 OOMs. Drop concurrency to 3, block images/CSS/fonts in preNavigationHooks, and call window.stop() post-navigation. Code: references/implementation.md Step 3.

Getting blocked (403s) mid-crawl. Start on datacenter proxies, and on a 403 re-enqueue the request with a residential proxy and retire the session to force a new IP. Code: references/implementation.md Step 5.

Runs cost too much. A 4GB allocation bills 8x more than needed for HTML parsing. Right-size to 512 MB and re-measure cost/run. Code: references/implementation.md Step 4; impact table in references/monitoring.md.

Resources

• references/implementation.md — full six-step walkthrough with all code blocks
• references/monitoring.md — in-crawl throughput logger + before/after impact table
• Crawlee Performance Guide
• Actor Memory & Performance
• Proxy Management

Next Steps

For cost optimization beyond performance tuning, see the apify-cost-tuning skill in this pack — it covers compute-unit budgeting, storage costs, and scheduling strategies that this skill's memory right-sizing feeds into.