The Memory Leak That Ate Our Cluster: A Postmortem

Headless browsers fail in ways that look harmless in tests but collapse under real load.

Our Puppeteer jobs leaked memory slowly enough to appear stable yet fast enough to crash full production runs. Before the fix, we chased OOMs, stalled sessions, and inconsistent execution without knowing why. After exposing the leak, we cut memory usage by more than half and added execution-lifecycle guardrails that keep our browser-based feature reliable under pressure.

This talk walks through the debugging workflow, the tooling that surfaced the issue, and the backend patterns that now shape how we run headless browsers in production.

You’ll leave with techniques to detect leaks early, monitor resource degradation, and design browser-backed features that behave predictably under real traffic and load.