Show CPU, memory and disk usage per instance and per volume

[traviswu-bigstack]

What type of PR is this?

/kind feature

What this PR does / why we need it

The instance list (/compute/instance and /compute/instance-admin) gains a single Usage column holding three stacked gauges — CPU, MEM, DISK — and the volume list (/storage/volume and /storage/volume-admin) gains an Allocated column. Part of the customer request behind bigstack-oss/cubecos#1449: see per-VM space use in the GUI.

Both read Prometheus directly, one instant query per metric returning every row at once, so adding the columns costs a fixed number of requests per page rather than one per table row on every render.

Gauge colours are the disk alarm's thresholds (tpl_alert_vm_disk.tick): blue below 50, green to 75, yellow to 85, red above. Warn and crit collapse into one red band — a row at 90% and a row at 96% both need attention now. A row that is about to raise an alarm looks like it before the alarm exists.

Disk falls back when an instance has no qemu-guest-agent, rather than reading -. The fallback is block allocation, drawn grey and marked with *, deliberately not on the colour scale — see below.

Which issue(s) this PR fixes

Part of bigstack-oss/cubecos#1449. The series come from the collector in bigstack-oss/cubecos#1450, which also fixes the prometheus_endpoint in skyline.yaml.in — without that fix every Skyline Prometheus query 404s and all of these columns render -.

Special notes for your reviewer

Why the fallback is grey and the volume column is called "Allocated". Block allocation and filesystem usage are not the same measurement and cannot be substituted for one another. Ceph keeps a block allocated once written unless the guest passes TRIM all the way through, so allocation runs ahead of filesystem usage and never comes back down. On a measured instance here the block layer reported 85% while the guest reported 14.8%, and running fstrim inside the guest — which reported 32.1 GiB trimmed, exit 0 — moved rbd du by nothing (confirmed by counting RADOS objects, not just trusting the object map: 4350 objects × 8 MiB). Ceph's rbd_skip_partial_discard frees an object only when the whole 8 MiB is discarded, and on an aged filesystem free space is too scattered for that.

So colouring a fallback row red would flag an idle instance, and those rows cannot raise the disk alarm anyway — it streams guest readings, which is precisely what they lack. On the volume page there is no guest reading to fall back from, since nothing maps a block device back to a guest mountpoint; naming that column "Usage" invites reading a fully allocated thin volume as a full one, so it says "Allocated" and the tooltip gives the absolute figures.

The shared UsageBar component and resources/prometheus/usage.js exist because two pages now draw the same gauge; the per-page modules only hold their own queries.

Additional documentation