Given the impact of Kubernetes on my field over the past decade, it is admittedly weird how little I’ve used it. But such is the effect of the Cave of Wonders that is life inside Amazon, where smart people have worked hard to produce some very impressive internal tools. Combine that with the world-changing AWS infrastructure, and you have a lot of teams for whom Kubernetes just wasn’t the most obvious choice.
Salt Lake City’s 2024 conference was my first exposure to KubeCon (extremely long official name: “KubeCon + CloudNativeCon North America”). Exhausting as these conferences are, they are also energizing in terms of what we learn, and exciting in terms of the new possibilities they expose us to. My new colleague Pieter Kasselman is right about a lot of things, and one of these is that the real magic happens in between the presentations. I saw that first-hand as one of the audience members for Nadin El-Yabroudi and Eli Nestorov’s talk asked some very good questions related to the IETF WIMSE working group that Pieter co-chairs.
My big takeaway is that the landscape of the Cloud Native Computing Foundation (CNCF) eerily resembles that of AWS’s service offering. The big difference is that the CNCF-sponsored products are open-source, and don’t come with infrastructure behind them. But that’s their appeal to their customers, in that they can use any and all the cloud infrastructure providers, or run the same products on their own hardware. Computing in large organizations is nothing if not diverse, after all.
Something else that kept coming up, either explicitly or implicitly, were products that would help make applications and workloads truly infrastructure-agnostic. “Provider arbitrage” came up a few times. That’s the idea of moving your work from one place to another depending on where it’s cheaper to run. This is a simple concept, and Kubernetes provides a lot that makes this easier, but as John Salvatier has said in a different context, “Reality has a surprising amount of detail.” This kind of thing is especially difficult for websites, but long-running and compute-intensive applications like AI training make it more appealing.