This is based on github.com/ibuildthecloud/systemd-docker Please go to http://github.com/ibuildthecloud/systemd-docker for more info about it.
However - There is one important change in this. This no longer moves the cgroups (or) notifies the systemd of the PID of the container process. Instead I run the service as 'simple' systemd service. The systemd-docker will, in a loop, keep watching for the health of the container. If there is any error, it exits.
I also specify the ExecStartPre and ExecStop, ExecStopPost - to take care of any case where the container didnt stop/remove clearly.
Here is an example:
# more /usr/lib/systemd/system/my.service
[Unit]
Description=my test
After=docker.service
Requires=docker.service
[Service]
ExecStartPre=-/usr/bin/docker kill busybox
ExecStartPre=-/usr/bin/docker rm -f busybox
ExecStart=/usr/pensando/bin/systemd-docker run --name busybox --rm busybox /bin/sh -c "while true; do echo Hello World; sleep 1; done"
ExecStop=/usr/bin/docker stop busybox
ExecStopPost=/usr/bin/docker rm -f busybox
Restart=always
RestartSec=10s
Type=simple
NotifyAccess=all
TimeoutStartSec=120
TimeoutStopSec=15While my production setup uses docker and runs the services with systemd, I wanted ability to unit-test my code. I started off with running the whole setup with docker (hence using docker in docker). However the cgroup treatment is really painful. A service running in docker in docker has very odd path in /proc/../cgroup which actually does not exist in the same path if mounting the file. Hence moving the container to systemd cgroup does not work.
Given such a case, my above solution seems to work for me. If you see any issues, please open an issue or send a pull-request.