You get this problem with any sufficiently tall sandwich of abstractions. SQLAlchemy has a query builder, the DSL is easily directly translatable to actual SQL

You get this problem with any sufficiently tall sandwich of abstractions. SQLAlchemy has a query builder, the DSL is easily directly translatable to actual SQL statements in your head. But since you have the full power of Python at your disposal, it gets quite convenient to abstract away some of the boilerplate. And then some of the common parameters. And then some more stuff, just for good measure. Then you try to bite the sandwich, and half the ingredients go sideways: change one parameter, and the generated SQL can look wildly different, etc. I still think DSLs are superior to macros though; a smart DSL transformer will catch a whole class of errors, while a macro engine like M4 could happily spit out garbage with syntax errors.