You can get pretty far using LMDB, specifically, its compound-key feature is quite good, and gets you a decent chunk of the relational model. Combined with Message Pack encoding, you can get an extremely concise data layer.

https://github.com/kriszyp/lmdb-js/discussions/170

I think there's a reason Tail Scale used a literal JSON file for up till 150 MB. of data: https://tailscale.com/blog/an-unlikely-database-migration/

So, I think the unconventional wisdom is correct.