I had the AWS re:Invent keynote livestream on in the background this morning while I was supposed to be doing actual work, which is exactly how these things get me every year. Andy Jassy is up on stage in Vegas announcing a pile of new services, and the one that actually made me put my coffee down was Amazon SageMaker: a fully managed thing for building, training, and deploying machine learning models, with hosted Jupyter notebooks baked right in so you dont have to provision your own instance just to open a notebook.
On paper thats genuinely useful. Ive spent entire weekends fighting CUDA driver versions on a spare desktop just to get a model training locally, so "click a button and get a notebook with the right libraries already installed" is not nothing. They also showed off something called DeepLens, a little wireless camera aimed at developers, running around $249, meant to let you deploy vision models straight to a physical device instead of just a cloud endpoint. Kind of a fun toy. I dont need one. I say that about most gadgets and then buy them anyway in January when theres a sale, so ask me again in six weeks.
Heres my actual complaint though, and its the same complaint I have every year during this week of announcements: the AWS pricing page for anything new is basically unreadable on launch day. Per-hour training costs, per-request inference costs, storage costs for the model artifacts, a separate line for the notebook instance itself sitting there idle racking up charges if you forget to shut it down. I went and looked at the console after the keynote just to get a feel for it and gave up after ten minutes of tab-switching between service docs. I dont think Im bad at this. I think its built for a team with a dedicated cloud spend person, not someone tinkering on a Saturday.
Which is sort of funny given the timing, because this is happening the same week Adobe put out their Cyber Monday numbers, $6.59 billion in US online sales on Monday alone, the biggest single shopping day online this country has ever had. So the backdrop to "here is our new managed ML platform" is literally "we just processed an absurd amount of the internets holiday shopping traffic without falling over." Its a flex, and it works as one. AWS is not hurting for credibility right now.
But it does put the whole keynote in a certain light for me. Everything Amazon ships this week is scaled for companies doing Cyber-Monday-sized traffic, and most of what I build is not that. My side projects run at, generously, a few hundred requests a day. For that kind of workload I still just want a box I can SSH into, not a managed pipeline with six billable components. Last month I moved a little Flask app that scrapes bus schedules onto a small VPS through Tricknowtech, pushed it up with git, pointed a domain at it, done in about twenty minutes including the coffee I made while it deployed. No notebook instance idling overnight, no separate inference endpoint bill I forget about until the invoice shows up. Its not glamorous and it will never scale to Black Friday numbers, but it doesnt need to.
Im not saying SageMaker is a bad idea, I think for a team actually shipping production ML it probably saves real time, especially the parts handling model versioning and deployment which is the part everyone always underestimates. I might even spin up the free tier this weekend just to poke at it, mostly out of curiosity about how the training job logs look. But watching a two hour keynote about managed infrastructure while running my own tiny stuff on a five dollar server kind of crystallizes something for me about where Im actually at as a developer right now, which is: still small, still fine with that, occasionally jealous of the tooling anyway.
Reinvent runs through Friday so there will probably be more of this. Andy Jassy is not done talking.