Browser journeys sometimes stall together on the laptop for about a minute #396
Loading…
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
What happens
On the laptop the browser step sometimes crawls for about a minute and fails a few journeys: 2.3 min against the usual 15 s, and it happened on a quiet machine too. Which journeys fail changes each time. On 2026-11-05, 3 of 5 runs with nothing else running did it. CI has not shown it.
What the samples show
During a stall the load reached 35 to 47 with 40 to 46% of the CPU idle, and
toptook 14 to 26 s to answer instead of 3. It was first seen after 8-worker repeat runs pushed the laptop into swap, though swap stayed flat during the later stalls.Not known
The cause. It is not the time limit (#393), and the journeys it hits pass alone, 8 of 8.
Done when
The cause is found, and the browser step passes 10 runs in a row on a quiet laptop.
Found while verifying #393.
Cause. macOS sometimes takes the performance cores away from the scheduler, leaving everything on the four efficiency cores.
kern.sched_recommended_coresdrops from 255 to 15. No test is at fault: which journeys fail only depends on which are running.Evidence (all on
3e69aa61or later)Not known. Heat is the likely reason, but macOS gives none. It is no simple threshold: the same spinners later ran 345 s without it, and the thermal level does not predict it.
Shipped in PR 398
tools/test.shends with the cores it had, when it was held back.docs/testing.mdhas a section, "A hot laptop".run.spec.ts:5clicked a tab under its close button, and the console missed what landed before its stream opened. That second journey fails on the old build and passes with the fix.Against the done-when. The 10 runs in a row were not run, at the owner's word. The evidence for the limits is one hot run, which passed all 39 journeys in 53.5 s on 1.7 performance cores. A stall still takes minutes, and at extreme load no limit will hold.
The PR closes this issue when it merges.