The cloud's between your ears, not in your data centre
Adopting hyperscale tech needs disposable servers and the courage to toss them
Hyperscale data centres are wondrous. Gartner's Sydney IT Operations and Data Centre summit this week heard that the likes of Google and Facebook have a sysdamin:server ratio of around 1:24,000, compared to the 1:500 common in the most efficient and virtualised enterprise bit barns. Hyperscale data centres can run at a power usage efficiency rating of a miserly 1.09. These data centres never stop and hardly ever wobble, thanks to designs that assume hardware will sometimes fail and therefore treat servers with the same reverence the rest of us reserve for empty soft drink cans.
Because hyperscale data centres run so efficiently, Gartner's analysts say they keep being asked if smaller and more mundane bit barns should give themselves a hyperscale makeover.
The subject came up in several sessions and the answer fell into three camps: you might be at hyperscale already, probably not and yes, a bit, if you are willing to change your culture and work practices.
The you might already be at hyperscale camp says technologies brewed in the bowels of cloud-scale operators are filtering down to mainstream users. Analyst Arun Chandrasekaran pointed out that Hadoop is probably Exhibit A for this trend, and that items like virtual SANs would probably not be on the agenda had hyperscale operators not decided arrays aren't for them. If you're using this kind of stuff, welcome to hyperscale even if you operate it on a more modest scale.
The probably not argument came from Gartner's Jo Skorupa, who said pointed out that the likes of Google don't do everything in new and exotic ways. “Not even Google is Google-y all the time,” he said. “They run Oracle and employ database administrators and have a back office. They do not operate in just one way for infrastructure and apps.”
So there's no need to rush to hyperscale for everything just because the likes of Google can point to impressive metrics.
Skorupa also says companies like Google “know what they do well, focus on that and outsource the rest.” For Google, running data centres is obviously critical. Outfits like high-frequency traders have an obvious reason to spend millions on technology to shave the occasional millisecond. Both probably spend more than three times on IT than the average organisation.
Making the investment to adopt hyperscale technologies therefore has a price tag lots of ordinary IT shops may never be able to justify.
The yes camp emphasises culture, because adopting hyperscale ideas needs some blue sky thinking.
“People don't innovate like they do at hyperscale because they are punished for failure,” said Gartner veep Cameron Haight, citing Clause 229 of the Code of Hammurabi - If a builder has built a house for a man, and has not made his work sound, and the house he built has fallen, and caused the death of its owner, that builder shall be put to death - as the prevailing mindset for data centre builders.
Innovative companies, he said, have a culture that accepts failure as a consequence of attempts to innovate rapidly. Haight cited this presentation explaining Netflix's culture as an example of how businesses behave when they prepare to work at hyperscale. The document says Netflix encourages “smart risks” and explains myriad ways in which the company gives its workers autonomy, even to the extent of not enforcing a fixed amount of leave. Instead, staff are trusted to take as much leave as they need. Trusting staff to do the right thing means defining processes and insisting people follow them becomes much less important, leaving people to be creative and innovate.
Don't write off the Netflix document as new age management mumbo jumbo. It's unusually sincere and cogent.
DevOps was another topic often mentioned as something hyperscale operators practice, as the dialogs and development style it engenders do make a difference.
Analyst Michael Warrilow also pointed out that getting to “infrastructure as code” sounds simple, but isn't.
“Policy as policy as infrastructure so infrastructure self-adjusts is the way to do it, not just code as infrastructure,” he said.
To mimic hyperscale operators you'll also need to create shared resource pools and buy capacity in anticipation of demand.
That last ploy sounds hard to get past most managers, especially in these times when extra capacity can be had for a pittance, by the hour, from … hyperscale data centre providers. ®