Energy Incident Register
Register / Data centre / EIR-2289
← Back to the list
Data centreJapanEIR-2289

Rotary UPS failure disables cooling at Azure data centre near Tokyo

Mar 2017Published 12 Oct 2026Responsible editor: Chief Editor

SignificantElectrical
Estimated damageModelled by the register's adjusterHow? Know the real figure?
Figure hidden: shown to subscribersLow confidence
out of service for about •• daysShown in the record →Full estimate
From the recordRoot cause hiddenFailed part hidden

Companies named in this entry: Microsoft

What happened

On Friday 31 March 2017, Microsoft Azure services across Japan were affected for several hours after a failure at the Japan East data centre in Saitama Prefecture, near Tokyo. The facility is managed by a third party. Microsoft's status page said that one rotary uninterruptible power supply (RUPS) system failed in a way that caused the power distribution feeding all of the data centre's air handler units to fail. With cooling lost, temperatures rose across the whole facility and some resources were shut down automatically to avoid overheating and protect data integrity.

Third-party vendor teams and Microsoft site staff restarted the air handlers and used outside air to force cool the building. Affected services included virtual machines and services that depend on Azure Storage. The failed RUPS unit was sent away for analysis.

Contributing factors

Microsoft said the cooling system was designed for N+1 redundancy and the power distribution design was running at N+2. The single RUPS failure affected the power feed to all air handlers, so the redundancy ratings did not prevent loss of cooling across the facility.

Lessons learned

  1. Map which cooling units share a single power distribution path, because redundancy ratings for cooling and power can hide a common point of failure.
  2. Test UPS failure modes, not only switchover, so that a fault in one unit cannot disconnect cooling for the whole hall.
  3. Have automatic load shedding and a tested fallback cooling method, such as outside air or portable units, ready for when mechanical cooling is lost.

More lessons on outages, backup supply and restoration, from every incident on the register.

Source

datacenterdynamics.com

Other entries that name Tokyo

Published 2026-10-12. Written from public reporting. Descriptive, not investigative. See the data accuracy disclaimer.

See the register's stories first in Google: add it as a preferred source.

Chart to reuse with this story: Data centre incidents by year, free with a credit (PNG).

Data centre incidents in Japan: 5 entries, with a free alert for the next one.

Browse the full register or submit an incident.