the reason is that LMDB allows only 1 write transaction
per thread, and this doesn’t work well with multiplexed
requests. there is probably going to be some performance hit,
but now transactions are scoped for each cache operation