Follows the policy of naming containers according
to the protocol they serve.

Fixes issue #13.

With 128m three days of logs are enough to cause death by GC, add 30%
more memory for long-lived VMs.

Fixes issue #19.

Change the deployment model from pull-on-every-restart to
pull-when-Ansible-runs, so that new images don't get automatically
deployed to production. Unifies the binary/configuration deployment
process so that both are Ansible-driven.

Another consequence is that the system no longer depends on the Docker
registry being reachable in order to restart services successfully.

A partial implementation of issue #21, we still can't uninstall
packages without additional metadata, but at least we can stop the
services.

This fixes issue #20 by making the daemon able to open the local LDAP
UNIX socket.

Make it more readable, and avoid the endless line-termination errors
that characterized the previous approach (it's still quite terrible
though).

Addresses issue #18, though not completely (yet).

In debian stable there is still no way to prevent auditd from logging
to disk, but at least we can reduce its I/O usage by disabling flushes.

Speeds up testing environments dramatically. We are going to have to
find a better solution for production environments anyway.

Uses dpkg-eatmydata, which disables fsync() calls.

The "test-base" environment now only contains the frontend and a
simple HTTP server, and it is primarily meant for CI.

The integration tests have been updated to skip tests when the
associated builtin services are disabled.

Before we were only cleaning up if no errors occurred.

The gitlab runner can't write to the source directory.

We can't reliably set the apt_proxy value in the base role defaults,
as we are using it in other roles too. Use "is defined" instead, and
just leave the default unset.

Trust the docker registry running locally on the Vagrant host