Nobody in the core development team has any significant experience with databases.
By not providing a database abstraction layer for the various storage pools (including mempool) they can have a better grip on the whole project.
Would it make sense to store the
mempool on-disk?
--snip--
LevelDB (kv store) seems pretty much the thing we need for bitcoin-core (there is no need for extensiv querying if you don't want to run a block explorer). What alternatives would you propose today? Would it be worth implementing (balance between (implementation-time+risks <==> benefits-of-a-new-database)? How would you handle migration, testing on serval filesystems?