- Shipped
- September 3, 2026 at 4:33 AM UTC
- Author
- Kamo
- Commit
- bc17024
Builds went from 7-9 minutes to 30+ a few commits ago and nobody had a reason why. Measured on the runner rather than guessed: `next build` sat at exactly 4.39 GB RSS — the 4096 MB heap plus native overhead — unchanged across repeated samples, at a sustained 211% CPU, 36 minutes in and still going. Flat RSS pinned at the ceiling with the CPU pegged and nothing progressing is V8 spending its time collecting garbage instead of compiling. That is the failure mode a heap cap has: it does not error when it is too small. It degrades, gradually, until a build takes four times as long and looks like a hung runner or a slow machine. The runner is neither — 48 cores at load 20, 68 GB free, docker filesystem 39% used. The app simply grew past what 4 GB can build comfortably. 4096 was a default that had not been revisited; the host has 123 GB, so the ceiling was never scarce, only stale.