Respectfully disagree on this. SRE is a huge complex realm unto itself. Just understanding how all the cloud components and environments and role systems work together is multiple training courses, let alone how to reliably deploy and run in them.
Lambda functions, for example: you have to understand their performance and scalability characteristics — in turn requiring knowledge of things like the latency added by crossing the boundary between a managed shared service cluster and a VPC — in order to understand how and where to factor things into individual deployable functions.
...and then those same devs being expected to directly debug the behavior of this "appliance" in a client environment (think: someone consuming the "appliance" through the Amazon Marketplace, where this launches the workload into an EKS cluster in the customer's own VPC, with the customer in control of defining that cluster's node pools);
...where this can involve, for example, figuring out that a seemingly-innocent bounded-size Redis cache deployment, needs 10x its steady-state memory, when booting from a persisted AOF file... for some godforsaken reason.
Lucas helped make one of my favorite Google Historical Artefacts, a crayon chart of search volume. They had to continuously rescale the graph in powers of ten due to exponential growth.
I miss pre-IPO Google and the Internet of that time.
Source: I was one at WebTV in 1996, and I worked with people who did it at Xerox PARC and General Magic long before then.
A second major reason is automation. If you read the linked site long enough you'll find that in the early days of Google, SREs did plenty of manual work like deployments while manually watching graphs. They were indispensable then simply because even Google didn't have enough automation in their systems! You can read the story of Sisyphus https://www.usenix.org/sites/default/files/conference/protec... to kind of understand how Google's initial failure of adopting standardized automation ensured job security for SREs.