ngosangns Knowledge Base

Case Study Quick Win Optimization

#technology#performance#optimization#cloudflare

Updated: 2026-09-24

Source

Source: Lazy-Style System Optimization: A Tiny Fix for a Huge Gain - Viblo > Author: Minh Monmen. The note keeps both case studies in the same shape, so the context, the problem, the solution and the result line up. The first study starts from a migration to Golang where CPU barely moved. The second starts from a bill for CORS headers alone. A short section lists the wins that arrived without detail. The closing approach section turns all of them into a routine of monitoring, prioritising by impact and reviewing the simple options.

Overview

The article shares small changes that delivered outsized results during system optimization. There are two main case studies. One cuts CPU by 70 percent through an early-return reorder on the request that carries most traffic, while the other saves 63 percent of data transfer by moving CORS headers and preflight handling to a Cloudflare Snippet. Shorter notes record drops in Redis bandwidth, Redis load, and random 5xx errors.

Case Study 1: 70% Less CPU From a One-Line Change

Context

  • System: a few billion requests per day
  • Migrated from NodeJS to Golang
  • Initial result: CPU dropped only 10% (below expectations)
  • Request A accounted for 60-70% of all requests
  • Request A was fully cached in memory, so it triggered no DB queries
  • Latency already reached p99 < 10ms (about 96% lower than NodeJS's 200ms)

Problem

  • Goal: reduce CPU (not latency)
  • Request A has a very small or empty body (status 204, 304)
  • Finding: the processing order was suboptimal

Solution

  • Change the processing order: check the conditions for returning a response early BEFORE running the complex logic
  • Specifically, move the early-return check ahead of the CPU-expensive operations
  • Only one line (the order) changed, plus a few type changes that followed

Result

  • 70% less CPU for request A
  • Cause: avoided unnecessary CPU-intensive operations when the request could return early

Lessons

  • The order of logic processing matters a lot
  • The early return pattern can significantly reduce CPU usage
  • In large systems, every small optimization has a large impact

Case Study 2: 63% Less Data Transfer by Moving CORS Headers to Cloudflare

Context

  • System: 1 billion requests per day
  • Request bodies are small or empty (status 204, 304)
  • Each request is only a few hundred bytes to 1KB

Problem

  • CORS headers were added to every response: 
    access-control-allow-origin: *
access-control-allow-methods: GET,POST,DELETE,PUT,PATCH,OPTIONS
access-control-allow-headers: DNT,Keep-Alive,User-Agent,X-Requested-With,If-Modified-Since,Cache-Control,Content-Type,Range,Authorization
access-control-max-age: 86400
  • These headers take about 300 bytes per request
  • At scale: 300 bytes × 1 billion req/day × 30 days = $810/month just for CORS headers

Solution

  • Move all CORS header handling and CORS preflight requests to Cloudflare
  • Use the Cloudflare Snippet feature (a mini version of Cloudflare Worker)
  • Requirement: the domain must be on Cloudflare's Pro plan ($20/month)
  • Setup: takes only about 5 minutes

Result

  • 63% less data transfer out (down to one third of before)
  • Saved thousands of dollars in data transfer cost
  • Cloudflare handles CORS at the edge, using no bandwidth from the origin server

Lessons

  • In large systems, every byte counts
  • Edge computing can significantly reduce data transfer
  • Cloudflare Snippet is a cost-effective solution for simple tasks such as adding headers

Other Case Studies Mentioned (Without Detail)

  • 90% less Redis bandwidth (from 1Gbps to 100Mbps) after increasing the memory cache duration
  • 99% less Redis load (20k RPS down to 200 RPS) after adding a Bloom filter to the blacklist check
  • 100% fewer random 5xx errors (15k req/day) after fixing the NodeJS keep-alive timeout

Approach

These quick wins are not luck. They come from:

  1. Comprehensive monitoring

    • Build a detailed, complete monitoring system
    • Detect any anomaly

  2. Prioritizing problems

    • Sort and set priorities based on impact
    • Focus on the critical path (the point with the largest impact)

  3. Meticulous review

    • Don't skip the simple possibilities
    • Especially important in large systems
    • Small changes can have a big impact at scale

Conclusion

  • There are many simple changes with big effects
  • The causes usually come from small details that people tend to overlook
  • Systematic investment in monitoring and a structured approach are the key

See also: Request Body Limit And Load Shedding · Hedged Request