docker-ce 29 ships docker-compose-plugin renumbered to v5.x (was v2.x
with docker-ce 26-28). Same Compose v2 codebase under the hood —
Docker just realigned the major number. The verify regex was hardcoded
to `v2\.[0-9]+\.[0-9]+`, so a successful migration on esh-docker-vm
(29.4.1, 16/16 stacks back up clean) reported FAILED on the verify
phase. Switched to `docker compose version --short` parsed for major,
gated `>= 2` — works across future plugin renumbers too.
STATUS.md: mark esh-docker-vm done. ana-docker is the last host.
After nh3-docker's swap, two systemd unit gotchas surfaced that the
playbook now handles automatically:
* The docker.io-era /etc/systemd/system/docker.service.d/override.conf
hardcoded ExecStart=/usr/sbin/dockerd; docker-ce installs at
/usr/bin/dockerd → daemon failed status=203/EXEC.
* The shipped docker-ce unit's ExecStart=dockerd -H fd:// conflicts
with daemon.json hosts: (defined for the 0.0.0.0:2375 homepage
discovery binding) → "conflicting host options".
The "Rewrite docker.service drop-in" step now backs up any existing
override, probes daemon.json for a hosts: setting, and installs an
override that strips -H from ExecStart when needed. Also added an
explicit systemctl reset-failed step to clear the start-rate-limit
state that 3 failed install-time starts leave behind.
configs/homepage/docker.yaml: comment out irv-ml1-docker provider —
20s-per-poll ETIMEDOUTs from the stalled host were drowning homepage's
logs and apparently blocking ana-pfi-docker discovery (the Miniflux
card in the News group wouldn't render until removal). Re-enable when
irv-ml1 is back.
STATUS.md: new "Active migration" section tracking the docker-ce
rollout — nh3-docker done; esh-docker-vm + ana-docker queued.
nh3-docker's daemon kept failing post-package-swap with status=203
even after daemon-reload. Root cause: a stale
/etc/systemd/system/docker.service.d/override.conf from the docker.io
era hardcoding ExecStart=/usr/sbin/dockerd. The override (a) points
at the no-longer-existing path, AND (b) typically also adds
-H tcp://... which now duplicates the hosts: setting in
/etc/docker/daemon.json — dockerd refuses to start when both define
hosts ('conflicting host options').
Daemon.json is the modern way to expose the TCP socket. The
override is redundant and wrong. Move it aside (preserve a
.pre-upgrade copy for forensics), then daemon-reload, then start.
Should let esh-docker-vm and ana-docker upgrades go through cleanly
without the manual debug loop nh3-docker required.
Docker's official package installs dockerd at /usr/bin/dockerd; the
Debian docker.io package put it at /usr/sbin/dockerd. After the apt
swap, the new docker.service unit file is on disk with the right
path, but systemd's cached unit still has the OLD ExecStart pointing
at /usr/sbin/dockerd. Daemon start fails with:
status=203/EXEC "No such file or directory"
Fix is systemctl daemon-reload between install and start. nh3-docker
hit this; adding the step so esh-docker-vm and ana-docker don't.
Multi-line shell with backslash-continued URL had the continuation
line starting at column 0, which breaks YAML's | literal block
('could not find expected :'). Stash the URL into a shell variable
and emit on one logical line.
Three docker hosts on the fleet still run docker.io 20.10.24 (the
Debian bookworm package) which:
* sticks at API 1.41 — newer compose clients (1.52+) refuse to talk
to it without DOCKER_BUILDKIT=0 fallback (caught during the
news-digest deploy on ana-docker today)
* is functionally EOL — docker.io's upstream no longer ships to it
* is missing modern buildx driver versions
This playbook handles a single-host migration: snapshot existing
docker package versions for rollback reference, stop every running
compose stack, apt-remove (NOT purge — preserves /var/lib/docker)
docker.io + plugins, add Docker's signed APT repo for Debian, install
docker-ce + docker-compose-plugin + containerd.io + buildx-plugin,
restart the daemon, bring stacks back up.
Volumes / images / containers survive the swap because:
* /var/lib/docker is preserved by `apt remove` (vs purge)
* both packages default to the overlay2 storage driver
Recommended host order (least → most blast radius):
1. nh3-docker (NH3 site, fewer services)
2. esh-docker-vm (home lab; many services but single-consumer)
3. ana-docker (production-ish; vaultwarden, gitea, synapse,
task-board, miniflux, news-digest, paperless-ng)
Run as `scripts/elway <host> --playbook playbooks/upgrade-docker-ce.yaml`
per host. Verify between hosts via `docker version` + spot-check a
few containers.
Rollback if a daemon won't start or a container errors:
ssh <host> 'sudo apt install --allow-downgrades \$(cat /tmp/docker-pre-upgrade.txt | tr "\n" " ")'
vh