I really don't understand the cloud craze. Everything is more complex to debug, more expensive, and more shitty in all the possible ways you can imagine. I mean i was not exactly a fan of the VPS craze 10-15 years ago, but at least it wouldn't automatically ruin your bank account whenever you got a little traffic.
Kudos to the author for having so much money (thousands in one month?!) to waste. I wish i did too :)
Coming from traditional infrastructure and development methods, you're mostly right. Part of the expectation of the cloud is that you do things _their way_. And even then each cloud provider does things a little differently. However, if you're willing to subscribe to the <insert provider> way of doing things it (and you'll have to trust me here) makes many things easier. Here's a short list:
* networking setup is free/cheap/doesn't require a Cisco cert. you can trust a developer to set things up.
* object storage is so much easier than any file hosting scheme you can come up with
* the path from container-on-a-host to container-in-a-cluster to container-in-{serverless,k8s} is extremely straightforward
* I turn all my dev/test servers off at night and they don't cost me a thing
* consumption based compute will result in a much cheaper solution than a VPS or colo (admittedly there are many assumptions baked into this)
* some core services (like sqs, sns on Amazon) are extremely cheap and have provably reduced development time because you're not having to build these abstractions yourself.
This all being said I'm not advocating an all-in approach without thinking it through, but to do so where it's easy and makes sense.
EDIT: clarity
Bare metal hosts set up the network for you. You may need to know how to configure a local network interface. Even if you actually rack and stack many colos will give you a drop with network set up. You don't need to do what you describe unless you are building your own DC.
> object storage is so much easier than any file hosting scheme you can come up with
That matters if your data volume is truly massive. Only a small percentage have this problem. Also AWS inbound is free so you could upload big data to AWS and warehouse it there if you wanted. Not using big cloud for everything doesn't mean you can't use it for anything.
> the path from container-on-a-host to container-in-a-cluster to container-in-{serverless,k8s} is extremely straightforward
This is the one spot where admittedly you will have to spend more in administration. You'll need to either run your own k8s or Nomad or adopt a different configuration, and you may have to think about it a bit more.
> I turn all my dev/test servers off at night and they don't cost me a thing
You could still do this. Just host live somewhere else. You could also test on a local VM, which is what we do. Obviously that depends on how big your app is.
> consumption based compute will result in a much cheaper solution than a VPS or colo (admittedly there are many assumptions baked into this)
You only see the savings if they are passed onto you. What we've seen is that Moore's Law savings have not been passed on by cloud providers. Look at what you can get at a bare metal host compared to how much the same compute costs in cloud. Years ago the difference would not have been so large.
Bandwidth costs in cloud are insane, and most use asymmetric pricing where inbound bandwidth is free. This is known as "roach motel pricing" for a reason. Data goes in, but it doesn't come out.
> some core services (like sqs, sns on Amazon) are extremely cheap and have provably reduced development time because you're not having to build these abstractions yourself.
Fair, but they make their money back elsewhere. Those are lures to get you locked in so you now have to pay their crazy compute and bandwidth egress charges.
Here's an example. There are more.
And yeah I agree about the "some services are super low cost so you get hooked" thing. Always been my impression of Amazon: they look for what they can apply scale savings on (usually object storage, it seems) and make it cheap and then over-charge for almost everything else.
I bet it still costs you more than my Hetzner one despite me not having to care about turning it on or off. I mean it's great that the cloud gives you this flexibility but you wouldn't need it to begin with if it wasn't so expensive.
As a case in point, I worked in standing up a critical system in a large enterprise a few years ago. We spent about $12M on compute, storage, networking, etc. At operational state, it was about 40% cheaper than AWS. The problem is, it all sat there for 6-18 months filling up before we fully hit that state.
With a cloud provider, you pay a high unit cost but if you engineer intelligently your costs should move with utilization. Except for government, most entities generally want to see opex move with revenue and prefer to minimize capex where possible.
Also, the cloud offers managed software as a service. You don't have to manage your own HA DB cluster or PubSub. It's all just there and it works. That can save you a lot on technical labor costs.
But yes, I do agree with your point. If you don't know what you're doing, you can nuke your budget super quick.
If I were building a new business, I would use both cloud and colo. But I do understand that everyone don't have that luxury.
"This is the cost of scaling, this is the cost of owning our own infra, how does that fit into our budgeting and requirements?"
There are plenty of tools and systems that can present a sufficiently linear cost relationship to load and usage that, should your COGS versus revenue make sense, the marginal cost of increased cloud resources a no-brainer--especially versus always-paid-for hardware. If you don't have such a linear relationship you're as much in the position of deciding whether the project is viable as you are anything else.
The lead time for Hetzner or OVH is measured in minutes and is appropriate for the majority of use cases. Old-school providers like these used to run the entire internet before the cloud craze started.
Colocation & own datacenter is not a good choice for most startups. Seems like a lot of people here miss a step and only go from one extreme to the other. There's a middle-ground of renting bare-metal that gives you amazing price/performance with none of the drawbacks of colocation or running your own DC.
And even if you do, which I think you’ll agree Troy Hunt does.
The opposite, I don't understand why anyone would ever put up a server if they didn't have to.
It's not 'processing power' that's going to be the 'big cost' for most projects.
It's headcount and salary.
If you can materially improve the operating ability of your company, then a few $K in cloud fees is dirt cheap.
I used to work at a 'tech company' that made a physical product and our IT was abysmal. We had to wait weeks for our sysadmins to order blades, get things set up, there were outages etc..
If a project is definitely going to be 'a few linux servers and never more' - even then it would be cheaper and more reasonable to use virtual instances.
The time to 'roll your own' is when the infra. operating costs are a material part of your business.
For example, 'Dropbox' invariably had to roll their own infra, that was inevitable.
Similarly others.
That said - as this article indicates, it's easy to 'over do it' and end up in ridiculous amounts of complexity.
The Amazon IAM security model has always been bizarre and confusing, and the number of AWS services is mind-boggling.
But the core case of EC2+S3 +Networking, and then maybe a couple of other enhanced services for special case works fine.
I also object to what I think is a vast overuse of Cloudflare, I just don't believe that in most scenarios needing to have content at the edge really changes the experience that much.
This only really applies to fully-managed services such as Heroku. Every other cloud still needs a DevOps person according to my experience in many companies.
Just security alone, in terms of managing access to all of those resources, various forms of backup, across regions ... it's just out of reach for most organizations.
In what universe? This frictionless perfect vacuum where traffic comes in a wholly predictable consistent continuum?
Also, we may be taking the problem the wrong way around: do these multigig files need to be accessed by everyone from a web browser? No, it's a dump file used by specific people in specific circumstances. Then why are we using HTTP for this in the first place? In this case, only publishing over Bittorrent/IPFS makes sense and many people will happily seed, pushing costs toward 0 for the publisher (and very close to 0 if you only push to a first circle of mirrors who can then push to more mirrors, some of which can be declared webseeds in your torrent).
Startups growing fast are the secondary audience.
The primary audience is large enterprises where their internal IT costs <<more>> than the cloud costs. Plus internal IT provides those resources after 6 months...
As to difficulty, they “solve” organizational problems by avoiding sticker shock when someone wants 100+k in equip that’s often a huge number of hoops to jump through and possibly months of delays, a giant bill every month and nobody a complains about the electric bill etc.
you can rest assured that even the largest company will come looking for the person responsible for increasing an expense by that large of a percentage. So maybe it doesn't come out of your personal checking account but you will certainly pay for it.
It’s easy to justify having a larger bill with more traffic. “A retail store isn’t going to complain that they need to buy more stock after selling more stuff that’s just a cost of doing business.” Meanwhile it can be hard to justify a capital expenditure just because traffic increased.
This only happens when consumers fail to set budget alerts. Troy could have saved himself $10k with 15min worth of work.
