Worker: Use cgroups to set memory limits on child processes

[josephjclark]

• prlimit doesn't work
• max-old-space-size doesn't work (doesn't cover native allocations, buffer stuff)
• maybe cgroups will work

[josephjclark]

Alternative idea from Claude: the parent process can poll the rss size of the child process:

But the parent process can check. It's idle while the child is crunching. You can poll /proc/{child.pid}/statm (or status) from the pool — it's just a file read, very cheap, and completely independent of what the child is doing. If RSS exceeds the limit, send SIGKILL from the parent.

The child could be in the middle of a synchronous JSON.parse on a 2GB payload — the parent doesn't care, it just reads a file and kills the process.

That's actually a clean fit with the existing pool architecture: exec() already manages the child lifecycle and has the PID. You'd add an interval there, read /proc/{pid}/statm, and if RSS is over threshold, kill the worker and reject with OOMError — same path as the existing OOM handling.

Linux-only, but you're already Linux-only for prlimit.

Want to go that route?

wait the parent process can check?
Yes. /proc/{pid}/statm is a file the kernel maintains for every process — any process can read it, not just the process itself. So the parent can do:

fs.readFileSync(`/proc/${child.pid}/statm`, 'utf8')

at any time and get the child's current memory stats, completely independently of what the child is doing. No cooperation from the child required.

Nothing beats statm for cheap polling.