I think it's a fallacy to say that you can or should build an application layer that's completely DBMS agnostic. Even if you are very careful to only write SQL queries with features that are part of the official SQL standard, you're still coupled to your particular DBMS's internal implementations for query compilation, planning, optimization, etc. At enterprise scale, there's still going to be plenty of queries that suddenly perform like crap, after a DBMS swap.

In my mind, standardization for things like ODBC or Hibernate or Entity Framework or whatever else isn't meant to abstract away the underlying DBMS, it's meant to promote compatibility.

Not to mention that you're tying your own hands by locking yourself out of non-standard DBMS features, that you could be REALLY useful to you, if you have the right use-cases. JSON generation and indexing is the big one that comes to mind. Also, geospatial data tables.

For context, my professional work for the past 6 years is an Oracle/.NET/Browser application, and we are HEAVILY invested in Oracle. Most notably, we do a LOT of ETL, and that all runs exclusively in the DBMS itself, in PL/SQL procedures orchestratedbbybthe Oracle job scheduler. Attempting to do this kind of data manipulation by round-tripping it into .NET code would make things significantly worse.

So, my opinion could definitely be a result of what's been normalized for me, in my day job. But I've also had a few other collaborative side projects where I think the "don't try and abstract away the DBMS" advice holds true.