That's precisely what we did with Monero. We abstracted our blockchain access subsystem out into a generic blockchainDB class, and then build out implementations for LMDB (default, preferred) and for BerkeleyDB (failover).

Of course, there exists a risk that a portion of the network may be forked off if in an edge-case between implementations, but that's no different to a subset being forked off the network because they're running 32-bit whatever in a 64-bit world.