This cannot be right. The Satoshi Bitcoin client always stored blockchain as the plain flat files. The database engines were used only for indexing into those files. The recent LevelDB-based backend worsened the situation by explicitly storing (also in plain files) the some precomputed data to undo the transaction confirmations in case of a reorganization.

The proper way to modularize the storage layers is to completely encapsulate all data and functions to access blockchain, without the crossing the abstraction layers inside the upper-level code.

I stress that "storage layers" need to be plural. The mempool is also a storage layer and also needs to be properly abstracted. In a proper, modular implementation the migration of transactions between those storage layers (for unconfirmed and confirmed transactions) will ideally require setting one field in the transaction attributes, e.g. bool confirmed;.