6 months from now: top ten reasons why it isn’t a problem AWS costs us twice as much as self hosting.
12 months from now: how we saved 75% by migrating our app back to our “legacy” data center.
anyone intentionally migrating infrastructure to the cloud in 2024 hasnt seen the bill, or is spending the taxpayers pound.
"Gov.uk notify" isnt a critical business service. it doesnt need multi-az or multi region failover. running a docker container itself isnt a feat of achievement.
you could save money over the longterm by running a pair of pizzaboxes in Cardiff and Edenborough running orchestrator. hell, you could arguably run notify in a pensioners basement off a pi powered by a solar panel.
Doesn't it depend on the kind of notifications being sent? If I was some government department using this service I'd certainly appreciate some sort of guarantees about my notifications going out to users in a timely fashion. It doesn't seem hard to imagine there are users for who this service is business critical.
Having your own datacenter requires someone to manage that hardware. There are costs for the space you are in.
But you also have to have the hardware for your peak utilization. If you are either very bursty or have significant times of little to no activity you should be able to scale up and down.
Yes a simple 1:1 is likely going to cost more (depending on your scale, if your Cloud bill is under $200 not sure you could really do much better) but then your not utilizing the cloud to its advantage.
That is before going into the savings with things like spot inferences for processes and don’t have to be real time.
Yes this doesn’t take into account what else they might be running to share resources with. But not every organization is going to be running a bunch of different workloads that can fully utilize their hardware at different times
Here's a less simplified example: https://world.hey.com/dhh/the-big-cloud-exit-faq-20274010
And if you click their first article about making the decision, they even acknowledged what I said about the ability to scale up and down being a major feature of the cloud and then said that does not apply to them anymore but it also was a big advantage for them at one point.
They are an example where moving to their own hardware made sense but for many companies it doesn’t.
For me I need to be able to spin up several hundred high power GPU instances for a few hours and then it’s quiet for a couple week.
The couple thousand dollars (if that thanks to spot) it costs to run that workload is far better than the cost of that same hardware for machines that would not be doing anything most of the time.
So yes you are still over simplifying the situation ignoring that there are real reasons that a company would use the cloud, which your example even references.
Search for "Let me give you a concrete example" if you don't want to read the whole thing.
If you want, I can go ahead and find a "medium" sized example for when you respond back that this example was too small.
> For me I need to be able to spin up several hundred high power GPU instances for a few hours and then it’s quiet for a couple week.
That's great. For you and your application. Nobody is saying there is NO application where the cloud makes sense. I certainly didn't say that. But, there are many, many, applications where IaaS/PaaS ends up being much more expensive than on prem and where the flexibility of the cloud is, if not completely irrelevant, just not worth the extra cost.
I'm guessing the scale-out needs of the UK's notification app are going to fall squarely in the category of "way more expensive to run in the cloud," but, hey, who knows? Maybe they sign their notifications in some blockchain ledger on every second Tuesday from 8 - 9:17 AM and need to rent some GPUs occasionally.
Which is not true.
I am reluctant to try to make any assumptions about the workflow here since I would assume they had ran the numbers to estimate what their cost would be. Possibly looking at their utilization.
The nature of what they built does seem to be like it would be a burst application. But the details on that are not here and are just assumptions on both of our parts.
Admittedly I responded do you because of this response to your post also:
> anyone intentionally migrating infrastructure to the cloud in 2024 hasnt seen the bill, or is spending the taxpayers pound.
Which again is a simplification of the situation. It’s a blanket statement that seems more anti cloud than reality.
Yes there are workloads that make sense to run on your own hardware, but many, many exist that don’t make sense either.
You do not have to go all the way to running your own DC. There are plenty of bare-metal providers that will provide you the machine with full root access while handling all the hardware maintenance, DC, logistics, etc, all for a flat monthly fee and no junk fees like bandwidth charges, etc.
> But you also have to have the hardware for your peak utilization.
The scaling up/down argument only works if the cloud resources were reasonably priced. With how overpriced they are (add in all the other junk fees like bandwidth, storage, etc) and you'll work out that a stack sized for peak load at a conventional non-cloud host is still cheaper to run than the minimum-load-sized stack at a cloud. The only situation the "scaling up/down" argument works out for is if you need large amounts of hardware for very infrequent events like an AI training task or data transformation job.
And yes, AWS EC2 is about 20% the cost of a fully managed bare-metal provider where you never even have to physically see your server if you don't want to. People have weird notions about how expensive servers are these days.
Colo/Cooling is indeed less flexible, but the cost benefits are so insane that it doesn't need to be very flexible.