This sounds a lot like Codd's framing of the relational model "problem" in his 1970 ACM paper. [0] His solution was to decouple logical from physical access. That framing and the solution still seem correct to me for the data lake use case.
This implies that any lower level storage formats we pick can morph to better optimized forms in a way that does not alter query semantics. It's not possible to solve this without adding refinements to the framing, such as whether data is mutable/immutable, and adding infrastucture to the solution, like metadata management and alternate projections of data. Vertica/C-Store introduced the latter a couple of decades ago. [1]
This implies that any lower level storage formats we pick can morph to better optimized forms in a way that does not alter query semantics. It's not possible to solve this without adding refinements to the framing, such as whether data is mutable/immutable, and adding infrastucture to the solution, like metadata management and alternate projections of data. Vertica/C-Store introduced the latter a couple of decades ago. [1]
Plus ça change...
[0] https://www.seas.upenn.edu/~zives/03f/cis550/codd.pdf
[1] https://web.stanford.edu/class/cs345d-01/rl/cstore.pdf