A cooling loop is commissioned, not switched on. The sequence below is the one that keeps warranty intact and keeps the first fill from becoming the first incident.

1. Pressure test dry, then wet. Prove the piping holds at above working pressure before any electronics are in the rack. Record the decay curve; a slow bleed found now is a leak found later at a much worse moment.
2. Flush and fill. Circulate, filter, and dump the first volume — installation debris is the single largest source of microchannel blockage. Fill with the specified fluid and inhibitor package, not with whatever the building loop contains.
3. Balance the manifolds. Set differential pressure, then verify per-rack flow against design at full population. Unbalanced racks are invisible until they are loaded.
4. Verify leak detection end to end. Trigger each sensor deliberately, confirm the alarm reaches both the BMS and the cluster's own alerting, and confirm what the system does automatically — some designs isolate a rack, some only alarm. Know which you bought.
5. Soak under synthetic load. Run a GPU burn across the hall and watch supply temperature, ΔT, and pump duty converge. This is also when residual air load reveals itself at the transceivers.

Redundancy is a CDU-level property. Pumps inside a CDU are typically N+1 and swap on a schedule to equalise runtime. The unit as a whole is usually not redundant unless you bought it that way — pairing sidecar units across a row so that either can carry the row at reduced ΔT is the standard pattern, and it is a purchasing decision made long before anyone notices it is missing. Ask two questions at design time: what carries the load when one CDU is out for filter service, and how long does the rack have before shutdown if flow stops entirely? The second number is small — tens of seconds at full load, not minutes — which is why leak isolation and orderly shutdown are wired into the control system rather than left to an operator.

The leak runbook should be short enough to execute badly at three in the morning: alarm identifies zone; power is dropped to the affected trays before anything else; the zone is isolated at the manifold valve; the fluid is contained and mopped, not chased; the affected UQDs are capped; and nothing is re-energised until the source is found and the loop has held pressure again. Rack-scale systems from the current generation include tray-level and rack-level leak detection precisely so that the zone is small — use it, and resist the instinct to keep the job running while somebody investigates.