The Shakespearian question of our age: To cloud or not to cloud
Don't need flexibility and scaling? Then don’t pay for what you don’t need
Opinion On-prem or not on-prem, that is the question. Hamlet, musing on something similar, decried the slings and arrows of outrageous fortune, dramatically foreshadowing the outrageous fortunes many report spending on the sea of troubles more commonly known as the public cloud.
There is no shortage of evidence. Are you a mid-sized SaaS provider with a stable workload? You can save millions repatriating your cloudy apps to the castle of on-premises. SEO tools supplier? Your business might be entirely unfeasible if you commit to cloud and nothing but. This isn’t a matter of saving a buck or two, this is detailing an existential threat to your career or even your company.
You are not a mid-sized SaaS provider, nor are you an Asian SEO toolmaker. Yet these examples show some of the weaknesses of the cloud model. If you have a well-defined task with stable computing needs and predictable throughput, then paying for the rapid scalability and flexibility of the public cloud buys you nothing of worth. And indeed, if you can buy a commodity on-prem server and get five to seven years’ use out of it, the savings over renting the same sort of ability over the same sort of time are huge. Amazon, Google and Microsoft all claim to be the right answer to all questions, but in their strengths lie their weaknesses.
The joke reply to "what is the cloud?" - someone else’s computer - is funny because it's true. At heart, the business model for cloud providers is that they can realise efficiencies of scale by running enormous datacenters and efficiently shoe-horning in large numbers of relatively small clients with rapidly changing needs.
The less well you fit that description, the greater the chances that you’ll do well to look at on-prem to run some, maybe a lot, of your tasks. Likewise, renting a whole bunch of software services when the lifetime of your product is long enough to repay in-house development costs may be culpably lazy. Designing your cloud projects from the outset to be easily moved on-prem once they're mature, though, is wisdom on afterburners.
As is being alert to other bad fists. If the cloud isn’t suitable for long-lived limited scope designs, it may also be an uncomfortable home for their polar opposite. AI/ML, although hotly promoted by the providers, can be an uncomfortably chubby capybara for the Amazonian anaconda to swallow. Very large datasets coupled to highly pipelined compute are vital to AI/ML. They also generate monthly hosting bills you can see from Jupiter. Also, they just may not work very well. Such are sensitive to latency, and deterministic latency is not a given in the public cloud.
One upside of the cloud is that you don’t have to worry about the details of where and what is going on to provide the service you’re paying for. The gory details of infrastructure management are magically taken care of. But they can’t be magically taken care of if they need to be highly optimised for your particular tasks, and AI/ML, especially at scale, is like a huge army demanding to be fed and supplied at the highest efficiency possible. Amateurs study strategy, professionals study logistics.
- The UK's bad encryption law can't withstand global contempt
- ChatGPT has mastered the confidence trick, and that's a terrible look for AI
- Killing trees with lasers isn't cool, says Epson. So why are inkjets any better?
- Open source's totally non-secret weapon big tech dares not use: Staying relevant
Another example of this is the ingress and egress of those large datasets. There has been no time in computing history when connectivity outstripped demand. Whatever you’re doing, if you can move more data more quickly to where it’s needed, your job will be easier. That can be hard and expensive to arrange, but much more so if it’s over highly contended paths that a cloud provider needs to keep available and well-greased for lots of customers.
Cloud is the convenience store of enterprise computing. It makes its money by having lots of pre-packaged goods available at any time to whoever has a credit card. It is one of the great victories of the modern world that you can walk out of your front door and in ten minutes be in possession of two lemons, a six pack of craft beer and a family sized pack of tortilla chips. Convenience always costs extra. You’d be foolish to try and run a restaurant on that sort of supply chain, and doomed if you tried to feed your army that way.
The cloud is a great place to experiment, to build and to tune your services. That does not automatically make it the right place to give them a permanent home, nor one that’s immune to critical cost benefit analysis. It may even just be a bad fit for basic architectural reasons, even if those didn’t matter so much in the past.
As hardware costs fall and hosting costs rise, designing cloud deployments from the outset to be easily repatriated to on-prem or hybrid, and regularly modelling that move, may save you from a tragedy of Shakespearian proportions. ®