The PagerDuty example seems really unique to me though.

That is, the entire point of PagerDuty is to notify you if your services go down. It seems like it would be colossally stupid if PagerDuty were to say "Sorry we weren't able to notify you about your cluster on AWS going down, because we're on AWS, too!"

For other businesses, I feel like the exact opposite incentives are in play. I've seen this in practice: "Yeah, sorry boss our service is down, but there is an AWS outage, so half of the Internet is down - not much we can do about that!"

Interesting. I didn’t know they did. Though can’t help but thinking that this is all going be a lot less valuable if your ovservability platform is on those mainstream cloud providers.

Not answering the question, but k8s doesn't entirely solve this because of the DB. If you're a SaaS provider that reached this scale, there's a good chance your architecture isn't just stuffing everything in one Postgres instance, so shipping a VM with Mongo, Postgres, Zookeeper, and Cassandra sorta makes sense.

If you'd had ample foresight, it might have been nice to have an architecture where the DB is truly abstracted away. Or not. Maintaining two backend implementations is a pain and error-prone.