lima-docker-lean

lima-docker-lean: Docker on macOS with near-zero idle cost

A measured, minimal Docker host for Apple Silicon Macs, built on Lima and Apple’s Virtualization.framework. When idle it writes nothing to disk and uses about 1% of one core. Unlike a stock Lima VM, it lets macOS take back about half of the RAM the VM holds after a load. It is a real Docker Engine with docker.sock, so testcontainers, docker compose, buildx and werf work unchanged.

This is not another wrapper. It is one Lima template plus the analysis behind every line of it, with the raw benchmark output included.

Results

Apple M5, 16 GB, macOS 26.6, Lima 2.2.0, 4 vCPU / 6 GiB guest. Single runs; see limitations.

	Stock Alpine on Lima	This template
Idle host CPU (empty VM)	1.31% of a core	1.00–1.12%
Idle disk writes	~25 KB/min (why)	0 B
`limactl restart` → ready	~26 s	~14 s
Host RAM really held after a Postgres + FoundationDB load	~4.2 GB	~2.1 GB (how)
pgbench (8 clients) / FoundationDB writes	~15k tps / ~246k keys/s	same

Quick start

brew install lima docker            # Lima >= 2.2 and the docker CLI (+ compose/buildx plugins)
git clone https://github.com/heaprip/lima-docker-lean && cd lima-docker-lean

limactl start --name=dev ./templates/minimal-docker.yaml
limactl restart dev                 # once, to boot with init_on_free=1

docker context create lima-dev --docker "host=unix://$HOME/.lima/dev/sock/docker.sock"
docker context use lima-dev
docker run --rm -p 8080:80 nginx:alpine   # published ports appear on localhost

For testcontainers, point DOCKER_HOST at the same socket (or use the context). When you are not working, limactl stop dev brings the cost to zero.

What the template does

Each of these is explained and measured in the analysis:

/var/log on tmpfs + commit=60. Lima’s guest agent logs a never-converging SyncTime line every 10 s, and the ext4 journal amplifies it to ~25 KB/min.
Kernel init_on_free=1 + an idle-only page-cache trim. VZ never returns guest memory, but zeroed pages compress to almost nothing in the macOS compressor.
No phantom ttyAMA0 getty. The cloud image respawns a getty on a device VZ does not have. This was the largest single CPU leak.
GRUB_TIMEOUT=0, syslogd in a 64 KB ring buffer (it cannot simply be disabled), and chronyd and cloud-init-hotplugd off.
Docker limits: capped container logs, BuildKit GC, fstrim on boot (the disk is a sparse file).
Kept on purpose: acpid, without which limactl stop is a hard kill.

Findings in short

Apple container machine (1.3 / 1.4.1) is not usable as a Docker host yet. It has no port or socket publishing, a new IP on every start, and $HOME is mounted read-write by default.
LinuxKit / DIY microVM: there is no ready image for Lima, and there is little to gain. dockerd + containerd + the guest agent are almost all of the guest’s idle RAM, and the OS services people usually strip add up to ~5 MB.
Immutable ISO (alpine-lima) vs cloud image: after tuning, the two are equally quiet and fast, but only the cloud image lets you change the kernel command line, which is what init_on_free needs. The ISO variant is kept in templates/minimal-docker-iso.yaml.
OrbStack is the one alternative that returns RAM to macOS on its own. It is proprietary and paid for commercial use, and it was not benchmarked here.
FoundationDB never idles. fdbserver alone keeps the VM at ~3–4% of a core, so stop it when you do not need it.

Repository layout

templates/   minimal-docker.yaml (recommended), minimal-docker-iso.yaml (immutable alternative)
docs/en/     analysis.md: full write-up with methodology and limitations
docs/ru/     analysis.md: the same in Russian
bench/       idle.sh, disk-writes.sh, wakeups.sh, load.sh, memtest.sh, fdbbench.py
results/     raw output of every run quoted in the docs

Reproduce

./bench/idle.sh dev 120 60      # empty idle: host CPU, guest disk I/O, memory
./bench/load.sh dev             # Postgres + FoundationDB: idle with containers, load, idle after
./bench/memtest.sh dev          # real host RAM cost after load (stops the VM!)

License

MIT. Measurements and write-up by @heaprip, assisted by Claude.

This site is open source. Improve this page.