I've thought about the same problem and think I have an explanation. SQL inherently avoids operations that cannot scale. This means that while the same operation can be written much more simpler as code if you're doing with pandas or spark, trying to do it with SQL would be less forgiving if your underlying data schema is not optimal or if your logic isn't scalable. But more often than not, the SQL query will run faster and more reliably given similar amounts of compute power to the two engines.

Of course you can always screw up and write suboptimal SQL that will take forever to finish, but it's just harder to do unless you're really trying to be dumb about it.

So I do think there's still a lot of merit in trying to learn and use SQL more.