We currently monitor and plot CPU, memory and I/O usage on a per-host basis.
In order to monitor resource usage on a service-basis, we should enable cgroups accounting in systemd, either globally or on specific units.
DefaultCPUAccounting, DefaultBlockIOAccounting and DefaultMemoryAccounting can be used to enable cgroups accounting system-wide, see systemd-system.conf(5). Similar settings (CPUAccounting and friends) can be used to enable the feature on selected units. See systemd.resource-control(5).
cadvisor can be used to export all available metrics to prometheus, using id as the label to distinguish between cgroups.
container_cpu_system_seconds_total{id="/system.slice/rsyslog.service"} 0.04 [...] container_memory_rss{id="/system.slice/ssh.service"} 40960