Image pull performance — registry mirrors, benchmark, cross-run layer cache

[adrianlungu]

Apple Container CLI's image pull is noticeably slower than Docker. This issue tracks the investigation + tuning work.

Known contributors

• Apple Container CLI's registry client is newer and less tuned than Docker's (HTTP/2 multiplexing, layer parallelism, etc.)
• ext4 volume formatting overhead — Apple formats each volume as ext4 on first touch; Docker uses overlayfs which skips the FS format step
• In full-isolation mode every VM is fresh, so layer cache doesn't carry across runs
• `gocker pull` gained the knobs in v0.7.0 (`--platform`, `--max-concurrent-downloads`/`-j`, `--progress` with TTY auto-detect), but defaults haven't been tuned

What to do

1. Benchmark: `time gocker pull postgres:16-alpine` vs `time docker pull ...` with cold caches. Break down where time goes (DNS, HTTP, extract, ext4 format) to know which slice is actually our problem.
2. Registry mirror config in `~/.gocker/config.yaml` — `registryMirrors: [mirror.gcr.io, ...]`. For people on slow links or behind a caching proxy.
3. Cross-run layer cache in full mode — today every fresh VM re-pulls. Explore sharing a snapshotter or pre-seeding layers from the host.
4. Skip multi-platform manifest fetches if Apple CLI is pulling both amd64 and arm64 — should halve bytes on multi-arch images.

Scope

Investigation-first. Numbers tell us which fix actually matters. Do (1) before implementing anything else.