Cloud and serverless platforms are similarly modern tech, except for deployment and hosting. Do you really think that e.g. AWS and Cloudflare "constantly change their behaviour in the edge cases"?
I mean major Google APIs / SDKs do... And not just in edge cases. Not at all uncommon for a vendor to decide a service is unprofitable and kill it, or to decide to launch a new better version and deprecate the old one / demise in a couple of years. That isn't fun when you heavily depend on it.
When you look at the sheer number of services AWS offer, it feels its only a bad year or a major competitor gaining an advantage and undercutting them on price before there is a risk they consider trimming to a smaller core set of services. I'd bet the VPS offerings aren't what goes...
Or a standard comes out, they adopt that and deprecate the existing offering giving you 2 years to migrate. Having to re-work everything is a major cost to large firms and can kill a startup.
I think there is a difference here between the core behaviour and the edge case behaviour. I guess I would trust that in the core behaviour they do not change on a day-to-day basis but the question is how do they behave when you are pushing the tools outside the intended core use cases. Can you then really trust services that change their core implementation constantly will work for your workflow?
TBH i would probably trust a CDN because I have a fairly simple use for such a service. If I were really pushing these tools, like running a video broadcast service or whatever, I would be much more worried.
What does something "outside the intended use case" even look like for a deployment and hosting platform?
If you're only using them for hosting and deployment there isn't really any lock-in either. That only occurs if you're using their other cloud services, and even then there are many platforms with similar services.
There are lots of cases where deployment needs some kind of customisation. This usually happens either around persistent state or around building in CI. E.g. in one project some years ago, we wanted to test in CI against a subset of the production database. So there has to be a script to appropriately subset this, not easy if you have complicated foreign key relationships, and you have to make sure confidential data did not make it across. We were not the only people to do this I've seen this elsewhere.
Other example, my previous project: the main framework is Django with a React front end. The admin interface is django html templates. But one of those templates has an embedded additional react component because on that admin page we needed more interactivity. So that has to be built. And all of this has to be tested in CI.
What does non-standard deployment look like? Here is one example. For my current project I use wildcard domains so users can create their own tenants on their own subdomain. This would not work on e.g. heroku or similars, at least i don't know how.
All of this could probably be done in Kubernetes but also much simpler with bash scripts and systemd units.
The problem is with functionality that is only available as a remote API, because for whatever reason we wanted to be cloud native rather than rely on free and open source libraries. I cannot pin that dependency version, as best I can choose which version of the API I am talking to, but if they have changed the underlying implementation then I can't request them to roll that back just for me.
