- Shipped
- November 14, 2026 at 9:43 PM UTC
- Author
- Kamo
- Commit
- 18b221a
With no prune running, `docker rm` of a finished CI job container still took 15-100 s: it waits for a containerd GC pass (leases.Delete -> gcScheduler.ScheduleAndWait), and each pass spent that time in the mark phase (getMarked/scanRoots) walking a metadata DB bloated by BuildKit build history that was never trimmed: moby_history held 2,694 leases and 18,701 blobs back to the last dockerd start (2026-09-28), ~8x everything else. - builder.history: maxAge 48h, maxEntries 100. - builder.gc: replace the default 4-rule policy (rule #0 capped local contexts/cache mounts at 2.57 GiB, deleting cache after builds) with one hard backstop: max 80 GB cache, prune when < 40 GB free. Routine cleanup is the idle-window k1m1-docker-prune.sh. Apply only in a quiet CI window: restarting docker also stops forgejo-runner.