database - NOSQL Design, without denormalization? -

i've started designing 'schema' of distributed store database.

i keep having mental debates on how denormalize. understand how it, , why increase performance if denormalization matches queries minimize gathering data multiple places...

...but, it's said pre-mature optimization bad. advantages of design relational, references instead of duplicate data embedded clear: elegant, flexible, no worries keeping duplicate data consistent, etc.

so wondering whether it's reasonable strategy design schema in relational way, using application layer gather data necessary, , change later if needed.

if traffic becomes issue, on technology can scale horizontally design changes (isolation, denormalization).

seems best choice among:

• start rdbms, move distributed store if needed
• start distributed store, full denormalized design (scale-ready)
• start distributed store relational design, denormalize + isolate if needed

thoughts?

thanks

i'd go option 2, assuming there no serious disadvantages scalable schema.

if know need use distributed store later, may go resembling final system start rather having deal multiple schema versions. nosql designs not more complex relational designs, different. many of nosql platforms mature enough easy use initial development sql is.

you may find don't need complex horizontal scaling - lot of joins in typical relational database there because doesn't support multiple values or hierarchical structures.