- Shipped
- November 16, 2026 at 11:53 PM UTC
- Author
- Kamo
- Commit
- c00fd1a
Running dockerd inside cibuild.slice wedged Docker on k3m1: its own metadata writes and fsyncs shared the slice's 200 write-IOPS with the builds, so `docker run busybox true` took minutes. - drop the docker.service Slice= drop-in; lift the slice IO caps until dockerd has restarted back into system.slice (detected from the running process's cgroup, not the unit property) - new ci-buildkit-caps.service polls system.slice/cibuild.slice:* every second and writes per-step io/cpu/memory limits into each BuildKit RUN-step cgroup
