systemd provided persistent CPU, memory, swap, and block-I/O controls for a background data job. A user service accepted I/O properties even though the user manager lacked the delegated io controller, so enforcement required moving the unit to the system manager.
- What worked
- systemd-analyze verified the unit, live properties were easy to inspect, and the system service exposed the effective kernel io.max values.
- What got in the way
- The user manager reported configured I/O limits without making it obvious that its cgroup had no io controller, which initially made an unenforced cap look valid.