A false smoke alarm nearly drowned Jane Street's only NYC datacenter
A false smoke alarm almost drowned Jane Street's only NYC datacenter
A systems engineer recalls the night a contractor's dust tripped the smoke alarm at Jane Street's Manhattan datacenter, which automatically killed the CRAC units and pressurized the sprinkler pipes. With no cooling, the room heated toward the 150°F threshold that would trigger the sprinklers, so he sprinted from his apartment and yanked every power cord by hand, halting the climb at 143.6°F. He later wrote an OCaml monitor that averages server temperatures and alerts the team.
The machine room had been kept across the hall so that someone could always run in and pull the power cords. I doubt anyone had pictured doing it to save the servers from the goddamn sprinklers.
- reticulates
The article reminds me of when Hurricane Sandy took out a bunch of New York DCs with 3+ feet of water and people were carrying diesel up stairs to keep generators running.
“The superhuman folks at the data center are hauling 55 gallon drums of diesel fuel up 17 flights of stairs. Assuming they're able to sustain a rate of transfer that exceeds the rate of consumption, the data center should run until electricity is restored. The latest estimate from Con Edison for lower Manhattan is 4 days. (The New York Stock Exchange will be open tomorrow, also running on generator power.)”
https://news.ycombinator.com/item?id=4720388 (https://web.archive.org/web/20121101194314/http://status.fog...)
https://arstechnica.com/information-technology/2012/11/how-o...
- hackermailman
In the article he talks about living within a few minutes walk to work and telling too many people so they made him the emergency office guy. Same thing happened to me, foolishly told everyone hey I can see my apartment from here so I would get panicked calls to run in and shut down servers because the information security guy got a tip of an impending exploit about to be released on a holiday weekend with no way to do it remotely. Everyone else lived a ferry ride or commuter train away.
- rdl
Odd that they didn't have at minimum 1) a physical Emergency Power Off (EPO) button inside the datacenter, such that one button could de-energize everything -- usually required as part of building a datacenter due to code variance for the "low voltage" (<4000V in this context...) wiring. 2) EPO interlink to the smoke alarms, such that it would de-energize equipment automatically (maybe with delay) if smoke detected.
- rainbowzootsuit
Interesting. The Sequence of Operations we write for the data halls I work on specifically do not shut down the CRACs on smoke alarm for this reason. The International Mechanical Code (IMC) specifically allows this where a air handler over 2000 cfm serves only one space because it's recognized that it will not spread smoke around a building. 2000 cfm is the threshold where you typically have to install smoke detectors, and typically they do shut down on smoke detection.
The CRAC won't feed a fire oxygen as it's just recirculating the same oxygen depleted air.
- jddj
The episode of their podcast where Ron explores the facilities side of their data centers and offices is worth a listen too
https://signalsandthreads.com/the-thermodynamics-of-trading/
- sroussey
Are there internet connected fire alarms that can be hacked to destroy the contents of some data centers? Like the European freezers?
- cesarb
> every minute, ask every server in each location directly for the temperature from its onboard sensors, take the average, and if it’s elevated, blast an alert out to everyone on the systems team to check for datacenter overheat.
I've once done the opposite: the office "server room" (a small room with three racks, two of them for the patch panels, one for the servers and switches) was cooled by a mini-split which was prone to locking up with the compressor enabled (the relay which controls the compressor got stuck) and freezing, which then stopped the air from circulating due to all the ice. I wrote a script which ran on one of the servers which read the ambient temperature through IPMI to detect when the room was too cold, email a warning, and power off (using ssh) the neighboring servers and then itself.
- sciurus
I wonder why the answer was to turn off the servers instead of turn their cooling back on. Would it not be able to stabilize the temperature fast enough?