On macOS, hold back parallel jobs on the kernel's memory pressure level, not on free memory

Fact

A job runner that starts as many jobs as there are cores ignores memory, and on macOS free memory cannot tell it when to stop: it sits at a few hundred MB on a healthy machine, because the rest is cached or compressed. sysctl -n kern.memorystatus_vm_pressure_level can: 1 is normal, 2 is warning, 4 is critical. It turns to warning as the kernel starts compressing and swapping hard. The pressure level is what Activity Monitor's graph and DispatchSource.makeMemoryPressureSource report.

Why it matters

Browser test shards take 1 to 1.5 GB each. On a machine already low on memory, a suite that finishes in under three minutes ran for over twenty, with load average near 300 while CPU sat a third idle: threads waiting on page-ins count towards the load.

How to apply

• Before starting another job while some are already running, read the level. At 2 or above, wait (poll every couple of seconds or when a job ends). Never hold the first job, so the run cannot stall.
• Expect it to oscillate: the runner drains a few jobs, then refills. That costs a little time on a busy machine and nothing on a quiet one.
• On Linux, os.freemem() (MemAvailable) is a usable signal instead.
• Load average alone misleads here: check vm.swapusage, the compressor size in top -l 1, and CPU idle before concluding something is CPU-bound.