Sure - for example we migrated all our stuff from MySQL to MariaDB.

It was completely painless, because all of the source code and many of the people who wrote that code migrated to MariaDB at the same time. They made sure the transition was effortless. We spent a months second guessing ourselves, weighing all of our options, checking and triple checking our backups, verifying everything worked smoothly afterwards... but the actual transition itself was a very short shell script that ran in a few seconds.

I will never use a proprietary database unless it's one I wrote myself and I'd be extremely reluctant to do that. You'd need a damned good reason to convince me not to pick a good open source option.

My one exception to that rule is Backblaze B2. I do use their proprietary backup system, because it's so cheap. But it's only a backup and it's not my only backup, so I could easily switch.

I'm currently mid transition from MariaDB to SQLite. That one is more complex, but not because we did anything MariaDB specific. It's more that SQLite is so different we have a completely different database design (for one thing, we have hundreds of databases instead of just one database... some of those databases are less than 100KB - the server just reads the whole thing into RAM and slow queries on our old monolithic database are less than 1 millisecond with this new system).

never use anything vendor specific like stored procedures, vendor specific datatypes or meta queries

Yeah we don't do anything like that. All the data in our database is in a JSON type (string, number, boolean, null) with the exception of binary data (primarily images). It doesn't even distinguish between float/int - though our code obviously does. All of the queries we run are simple "get this row by primary key" or "find all rows matching these simple where clauses. I don't even use joins.

Stored procedures/etc are done in the application layer. For example we don't do an insert query anywhere. We have a "storage" object with simple read/write functions, and on top of that there's an object for each model. That model does all kinds of things, such as writing the same data in different places (with different indexes) and catching "row not found" failures with an "ok, lets check if it's in this other place". That's also the layer we do constraints which includes complex business rules, such as "even if this data is invalid — we will record it anyway, and flag it for a human to follow up on".