log1x

Budgeting rack power for a pantry, not a data centre

19 May 2026 · power

The rack is rated for far more than the room is. Every capacity conversation in a flat is really a conversation about one breaker and one extension lead.

Start from the breaker, not the kit

The pantry shares a 10 A circuit with the hallway and, it turns out, the outlet the kettle lives on. Nominally that is a bit over two kilowatts. In practice the number worth planning against is much smaller, because the circuit has to survive the worst simultaneous moment, not the average one.

ItemSteadyWorst case
Storage node, four disks96 W210 W at spin-up
Compute node145 W240 W under build load
ARM sensor board6 W9 W
Switch and router34 W41 W
UPS overhead and charging55 W190 W recharging after an outage
Total336 W690 W

The three numbers that actually bite

Spin-up, not run

Four disks starting together draw roughly double their running figure for a few seconds. Staggered spin-up costs about eleven seconds of extra boot time and removes the spike entirely, which is the cheapest improvement in this whole note.

Recharge after an outage

The moment power returns is the worst moment on the circuit: everything restarts at once and the UPS begins charging hard at the same time. If the budget only works while the battery is full, it does not work.

The kettle

Two kilowatts, on the same circuit, entirely outside my control. This is not a joke line. It is why the steady figure sits near a third of the circuit rather than two thirds.

What the UPS actually buys

At 336 W the line-interactive unit gives roughly nine to eleven minutes, which sounds thin until you look at what it is for. It is not there to keep the lab running through an outage. It is there to ride out the two-second dips that happen a few times a winter, and to give the storage node time to shut down cleanly on anything longer.

on battery, 30 s    -> ignore, log it
on battery, 120 s   -> stop non-essential services
on battery, 240 s   -> ordered shutdown, storage last
battery below 40%   -> ordered shutdown regardless of runtime

The last line exists because a battery two summers past its date reports runtime with more optimism than it can deliver.

Where the headroom went

The original plan had a third node in the budget. It did not survive contact with the recharge case: adding 140 W steady pushed the worst-case simultaneous moment past what I am willing to put on a shared domestic circuit. It waits for a dedicated line, or it does not happen.

In a flat, the constraint is never the rack. It is the wall.

← All notes