Much happened since the last post, but first things first.
Today’s Best Band Ever™ is IDLES. Anger with a purpose.
Difficulty: Hurt me plenty.
How’s the world.
It’s 400 days since Moscovia started a 3-day quick victorious war against Ukraine. And things are very bad in Bakhmut. Since August.
Twitter is run by a clown. Another clown is indicted. Big banks are failing and some are being rescued.
It all would’ve been funny if we were spectators and not participants.
But life goes on, Ukraine will definitely win and hopefully soon. On top of that Ted Lasso’s 3rd season is going strong.
In the tech world, GPT 3/4, LLaMA and Alpaca are all over the place. Once again, thanks to Open Source, the machine learning is experiencing its Stable Diffusion moment. Knowledge must be free.
Probably one of the most interesting developments in the ML world, is Apache TVM. OpenAI created a fascinating system, but the fact that ML advances are restricted to companies with deep pockets is deeply troubling.
In the cloud world, things are quiet and stable. I would even say the cloud has reached a certain maturity level and that’s why we’re not seeing many dramatic developments.
We’re likely entering a rationalisation phase of early or even late majority in the technology adoption cycle, and I think the main question about cloud adoption is whether one actually should move to the cloud. The advantages of cloud are quite obvious but The Cloud™ is not magic and comes at a cost. Those who migrate because <a magazine popular with C*Os> said everybody and their uncle are moving to the cloud, quite often realise that it’s not all amazing.
Bluntly put, lift and shift alone is a bad idea. If one doesn’t progress into replatform, rearchitect etc, they will be spending more and, worse, will likely get stuck half-way and eventually might even retreat from the cloud.
Another curious case study is 37signals and their struggle with cloud costs and their simplification journey.
I think there are several main points in their journey:
MRSK is the tool they’ve developed to orchestrate containers. Its simplicity likely means their architecture is fairly trivial (DB and a web server), and their load is stable, so there’s no need for autoscaling (I do wonder how they deal with failing hardware, though). Very few companies actually need K8S.
$3,201,564 a year is a significant amount of money, but keep in mind, their move out of cloud doesn’t mean they’ll be saving three millions a year. They still will need to either rent or buy hardware and hire people to manage it. Still I expect they might be saving 60% of that, which is nice.
A subtle point is their use of ElasticSearch. While an excellent product for what it offers, it’s notorious in how pre-cloud it is in architecture and management.
In other words, cloud could be a good tool for 37signals, but I don’t think they were using it right. On the other hand, their approach is working for them, and thus they don’t need cloud.
I do believe that this kind of rationalisation and pragmatisation will be fundamental for the cloud adoption and software architecture in general in the coming years.
Take care and stop wars.
PS. This post is surprisingly still 100% GPT-3+ free.
Love that song....