Werner Vogal gave a classic Werner keynote… his keynotes are more focused on the soul of the cloud than on specific new services. As always, he started with a video about possible directions we might all go, and as usual, he suggested we go in a different direction. He returned to the video several times during the talk.
A theme of the talk was architecting for cost. He talked about PBS as an organization that is highly cost-constrained. They were hitting limits in terms of both performance and cost. By moving to the cloud and re-architecting their software, they reduced their cost by 80%.
AWS has always emphasized the cloud as a way to save costs, but it has lately pivoted to recognizing that lowering costs also reduces carbon footprint, which increases sustainability. He gave WeTransfer as an example of reducing their carbon footprint by 78% by moving to the cloud. He observed that cost is a proxy for sustainability.
Werner has written a book called The Frugal Architect, where he lays out the principles described in his talk. It’s not as simple as just reducing costs, it’s paying attention to where your costs are and how they scale with revenue. As an example of failing to heed this principle, he referred to the early days of cell phone “unlimited” data plans. The introduction of these plans changed consumer behavior … suddenly streaming Netflix on a mobile device was possible. And data usage grew beyond the carrier's ability to generate associated revenue.
This leads to the maxim that business decisions need to be in harmony with technology decisions. He explained how this led AWS to create Nitro and Firecracker, which underlie EC2 and Lambda. They needed new technologies to provide the scalability that the business required.
He described seven “laws” of building cost-aware architectures in the cloud
- Make Cost a Non-Functional Requirement
- Systems that last Align Cost to Business
- Architecting is a Series of Tradeoffs
- Unobserved Systems Lead to Unknown Costs
- Cost Aws Architectures Implement Cost Controls
- Cost Optimization is Incremental
- Unchallenged Success Leads to Assumptions
The last of these points is perhaps the most interesting. Success leads to complacency, and Werner urged people to discover their hidden beliefs and disconfirm them.
One of their beliefs for many developers is language choice. Werner showed a table of programming languages by energy efficiency. Rust turns out to be massively more efficient than Python, which partially explains AWS’s embrace of the language.
Another of Werner’s guests was Cat Swetal from NuBank. She said, “we were born on AWS”. They set out to revolutionize fund transfers in Brazil. By going to AWS, they changed fund transfers from taking over a day and costing $5 USD to 10 seconds and free.
Paying attention to costs requires observability, and to that end, AWS announced “myApplications” for the management console. This new UI provides a variety of views of one’s application, including not just the usuals such as CPU and memory but also energy usage and cost. A related new tool is CloudWatch Application Signals.
“Sustainability is a freight train coming your way”.
Werner returned the video in an homage to a previous year’s keynote about The Matrix. He spoke with The Architect and the Predictor. They spoke of grand visions of centralized control and monolithic designs to run everything in a hoverboard-based future. In a nod to “Back To The Future” Werner replied that “this vision wouldn’t McFly”.
It was striking that Werner didn’t even mention AI or GenAI until the second hour of his talk. And even then, he went in an unexpected direction. Rather than touting GenAI, he urged people to remember what he called “old AI” i.e. Machine Learning. Not every problem requires a Large Language Model. Many problems are more appropriately addressed through an ML solution such as a classifier.
In the most moving segment of the talk Werner introduced Dr. Rececca Portnoff of Thorn, who spoke of her group’s efforts to combat childhood sexual abuse. They have a system to scan images on the internet and classify them as images of abuse or not. This is a classic classification problem with several unique aspects. Training such a system requires using hundreds of thousands of “good” and “bad” images. It’s actually illegal to store these “bad” images so they have to live in permitted on-premise systems. Using ML Thorn is rescuing children from horrific abuse … something that can make us all proud to work in the cloud.
Another example of “old AI” still being relevant was Precision AI which makes drones that can fly over crop fields and use machine vision to identify individual plants that need weed treatment. This both reduces cost and the runoff of pesticides.
At the very end of the keynote, he gave a few very interesting product announcements, including CDK Constructs for Generative AI, SageMaker Studio Code Editor, Q in CodeCatalyst, and App Composer in VS Code.
Werner’s keynotes have always represented the soul of the company, and this year was no exception.
Brian Tarbox
David Victoria
