> In a way, the very existence of the normal forms can be taken as a sign that something is seriously wrong with the relational algebra given how easy it is to create a data model that will produce rubbish for certain queries.
I am of a growing opinion that normal forms below 6th (i.e. 1-5) are simple mistakes taken as performance optimizations. The mathematical side of me says to keep going too - I suspect that there is a "normal form" beyond the 6th that says something along lines of how all keys must be synthetic and that there is no such thing as a business or domain key. I strongly believe the relational algebra/calculus as applied to domain modeling hasn't quite reached a fundamental conclusion yet.
> If the problem with 6th normal form is the underlying storage, then change the underlying storage to optimize for 6NF.
You're god damn right. There is zero reason we cannot build storage layers that can handle this kind of logical layout. My current favorite solution is to simply event source all the facts and keep an in-memory copy of the active business state. You can store pointers to strings and other BLOBs so you don't bloat RAM over something you could stream from NVMe on demand. Most of the pain in higher normal forms is with the join, so having a working set that can fit into RAM is one way to partially address this problem.
I am of a growing opinion that normal forms below 6th (i.e. 1-5) are simple mistakes taken as performance optimizations. The mathematical side of me says to keep going too - I suspect that there is a "normal form" beyond the 6th that says something along lines of how all keys must be synthetic and that there is no such thing as a business or domain key. I strongly believe the relational algebra/calculus as applied to domain modeling hasn't quite reached a fundamental conclusion yet.
> If the problem with 6th normal form is the underlying storage, then change the underlying storage to optimize for 6NF.
You're god damn right. There is zero reason we cannot build storage layers that can handle this kind of logical layout. My current favorite solution is to simply event source all the facts and keep an in-memory copy of the active business state. You can store pointers to strings and other BLOBs so you don't bloat RAM over something you could stream from NVMe on demand. Most of the pain in higher normal forms is with the join, so having a working set that can fit into RAM is one way to partially address this problem.