Technology
Checkly uses AI agents to rewrite critical high-throughput service in Go
The monitoring platform successfully migrated its Results Daemon from Node.js with zero production incidents, reducing running pods by 70%.
The short version
- Checkly used Claude Code to rewrite its Results Daemon, which processes approximately 92 million messages a day, from Node.js (JavaScript) to Go.
- The migration achieved a 70% reduction in running pods, a lighter database load, and zero production incidents during the rollout.
- A comprehensive, black-box test harness built prior to the rewrite was critical to the project's success, though an early deployment revealed a routing gap between local and production queues.
Key facts
- Checkly chose to rewrite its Results Daemon, a high-throughput background worker processing 92 million messages daily, because the legacy vanilla JavaScript Node.js component was becoming a bottleneck under doubling traffic loads.[Hacker News]
- Prior to using AI, Checkly engineers built a black-box test harness using Playwright, Docker Compose, and Toxiproxy to establish byte-to-byte parity based on input-output "golden files" generated from real data.[Hacker News]
- Using Claude Code, the AI agent generated approximately 13,000 lines of deployable Go application code overnight, keeping token usage within a $200 subscription limit.[Hacker News]
- An earlier attempt to write the service using the Claude Opus model failed to meet the company's quality bar and was discarded.[Hacker News]
- An initial production deployment failed because the test harness simulated a simplified three-queue local topology, whereas the real production environment contained 18 queues, requiring a human-supervised refactor of the retry logic.[Hacker News]
- To safely deploy the service, Checkly used feature flags to migrate customer accounts in staged cohorts—starting with free accounts, moving to paid, and ending with enterprise accounts.[Hacker News]