Dual Power

One day, the sysadmin's scream cut through the office.

Every piece of equipment in the datacenter had stopped responding. The VPN. The bastion. The emergency path we kept on standby. One by one, we ruled them out. What remained was the bleak conclusion. Every server had lost power.

x86 servers threw away the lavish redundancy of SPARC machines. But they kept dual power supplies. One PSU dies, the server keeps running. At the server level, the redundancy is solid.

A brief aside. I once brought a watt meter into a datacenter out of curiosity. Two cables plugged in. How does the load split? Even distribution. A server drawing 200W pulls 100W through each PSU. Yank one cable and the surviving unit picks up the full 200W. No blip. This was a specific Dell model — not every server behaves the same.

The catch is one layer out. A datacenter contract covers more than the rack. Network line, power — each comes as its own contract. The contracts are surprisingly flexible: you can even pull a residential fiber line into a rack. But the power side is bound by numbers. Total capacity drawn into the rack, and physical tap count. Spare capacity is useless if you've run out of taps. Spare taps are useless if you exceed the capacity.

Taps come in two feeds: A and B. You plug each PSU into a separate feed. One feed goes down, every server stays alive on the other. That was the plan.

Fill up the taps without thinking, and when one feed drops, the surviving feed exceeds its rated capacity. Server-level redundancy can be perfect — exceed the contracted capacity and the breaker still trips. The contracted rack goes dark in an instant. That was the scream.

I never caused it myself. But ever since that day, no server gets racked until I've run the numbers.

Cloud era now. Few people touch hardware anymore. Nobody walks into a datacenter carrying a watt meter. Probably.