Comment by eksemplar | Hacker News Reader

Comment by eksemplar | Hacker News Reader

Parent Full thread

It could be, but it’s more expensive.

We killed the local datacenter a decade ago when we started renting iron to house our own cloud. The majority of our systems still operate on VMs in that setup, but even that is more expensive than going full Azure/aws. We don’t, because a lot of our legacy systems wouldn’t run in the full cloud and if we need a SQL cluster on our own internal network anyway, then we might as well use it. Though, we’re nearing the point where a DB cluster in Azure will be considered our own network.

Then there’s the stuff we can’t do, like spin up 5000 servers on demand. Say we train some ML to find something specific missing from millions and documents. With Azure we can call Microsoft and plan to execute this, and they’ll give us exactly the resources we need for exactly how long we need it, and you just can’t beat that.

So it’s a cost-benefit thing. By going Azure and AWS, we get to spend more money on our primary function, which isn’t IT.

What kind of patterns do you have that renting iron is more expensive than Azure/AWS? That seems counter-intuitive.

It's not the ML itself, it's the ability to quickly store large amounts of data and troll through them in a few hours and then shut everything down and never use it again.

One example is the case files I told you about. It's basic image recognition really, but it's millions of poorly scanned pages. 10 years ago, we would've gone through it by hand, and it would've taken 8-12 full time employees 3 months to do it.

Now we can have one PHD student do a 2-3 month training project that he/she uses for their research, and we have an algorithm that can go through the case files in 5 hours, provided we rent enough iron to do so. With AWS or Azure that's not a big problem, depending on size you can do it on-demand, but we had to call Microsoft and prepare 5000 server instances when we did it. That was some years ago by the way, so maybe it's on-demand now. Anyway, they went through the case files in 5 hours, and had a lower error margin than our employees would have had.

After those 5 hours, you shut everything down. There is really nothing comparable to that. You can certainly do that without moving other things to AWS or Azure, like I said, but it's often the financially solid decision.

And what's the problem with it exactly? Our own internal network has been breeched more than once, our Azure instances have never been hacked.

Reply on news.ycombinator.com