I hope this won’t be counted as some form of self-promotion, even though I am sharing a post from my own blog.
As a tech worker who works in a Cloud shop, I wanted to elaborate the many reasons why I find working with Clouds terrible, from multiple points of view.
I tried to organize my thoughts in a (relatively long) post, in which both technical aspects and political aspects (which are very related) are covered.
I am sure many people will have different perspectives, and this could be potentially also a nice prompt for a discussion.
And what is simpler and faster, renting rack space in another continent (and buying, shipping, racking and initializing) or editing your terraform file?
Why on another continent? Except maybe VDI, some direct calls to some LLM or some insane scales, there’s nothing really that needs those round trip times.
Also data rules / data privacy. Some things need to have the original in Europe; China & Russia also need their data separated from others.
Because the customer demands it.
Not OP, but they are comparable efforts, especially since it’s a relatively infrequent activity. You can rent dedicated boxes with off-the-sheld hardware almost instantly, if you don’t want to deal with the hardware procurement, and often you can do that via APIs as well. And of course both options are much, much, much cheaper than the Cloud solution.
For sure speed in general is something Cloud provide. I would say it’s a very bad metric though in this context.
Full-ACK.
My last customer (global insurance company) provisions several systems a day. Now moving to hundreds via Jenkins. Frequency is environment dependent.
If your compute needs expand that much everyday, and possibly shrink in others, than your use-case is one that can benefit from Cloud (I covered this in the post).
That said, if provisioning means recycle, then it’s obviously not a problem.
This is a very rare requirement. Most companies’ load is fairly stable and relatively predictable, which means that with a proper capacity planning, increasing compute resources is something that happens rarely too. So rarely that even a lead time for hardware is acceptable.
So if I may ask (and you can tell), what is the purpose of provisioning that many systems each day? Are they continuously expanding?
Agree to disagree. Banking, telecommunications, insurance, automotive, retail are all industries where I have seen wild load fluctuations. The only applications where I have seen constant load are simulations: weather, oil&gas, scientific. That’s where it makes sense to deploy your own hardware. For all else, server less or elastic provisioning makes economic sense.
Edit to answer the last question: to test variable loads, in the last one. Imagine a hurricane comes around and they have to recalculate a bunch of risk components. But can be as simple as running CI/CD tests.
Systems are always overspecced, obviously. Many companies in those industries are dynosaurs which run on very outdated systems (like banks) after all, and they all existed before Cloud was a thing.
I also can’t talk for other industries, but I work in fintech and banks have a very predictable load, to the point that their numbers are almost fixed (and I am talking about UK big banks, not small ones).
I imagine retail and automotive are similar, they have so much data that their average load is almost 100% precise, which allows for good capacity planning, and their audience is so wide that it’s very unlikely to have global spikes.
Industries that have variable load are those who do CPU intensive (or memory) tasks and have very variable customers: media (streaming), AI (training), etc.
I also worked in the gaming industry, and while there are huge peaks, the jobs are not so resource intensive to need anything else than a good capacity planning.
I assume however everybody has their own experiences, so I am not aiming to convince you or anything.
Banking is extremely variable. Instant transactions are periodic, I don’t know any bank that runs them globally on one machine to compensate for time zones. Batches happen at a fixed time, are idle most of the day. Sure you can pay MIPS out of the ass, but you’re much more cost effective paying more for peak and idling the rets of the day.
My experience are banks (including UK) that are modernizing, and cloud for most apps brings brutal savings if done right, or moderate savings if getting better HA/RTO.
Of course if you migrate to the cloud because the cto said so, and you lift and shift your 64 core monstrosity that does 3M operations a day, you’re going to 3nd up more expensive. But that should have been a lambda function that would cost 5 bucks a day tops. That however requires effort, which most people avoid and complain later.
Ofc they don’t run them on one machine. I know that UK banks have only DCs in UK. Also, the daily pattern is almost identical everyday. You spec to handle the peaks, and you are good. Even if you systems are at 20% half the day everyday, you are still saving tons of money.
Between banks, from customer to bank they are not. Also now most circuits are going toward instant payments, so the payments are settled more frequently between banks.
I want to see this happening. I work for one and I see how our company is literally bleeding from cloud costs.
One of the most expensive product, for high loads at least. Plus you need to sign things with HSMs etc., and you want a secure environment, perhaps. So I would say…it depends.
Obviously I agree with you, you need to design rationally and not just make a dummy translation of the architecture, but you are paying for someone else to do the work + the service, cloud is going to help to delegate some responsibilities, but it can’t be cheaper, especially in the long run since you are not capitalizing anything.
Not only it can be cheaper, it is cheaper in most cases… when designed correctly. And if you compare TCO, not hardware vs IaaS.
It can also be much more expensive of course, but that’s almost always a skill issue.