I don't get the transaction bit. At least with postgres, a transaction doesn't guarantee that all statements in it see the data at the same point in time (actually, it's not even guaranteed for subqueries).
Also, often, the transactional database servers is more difficult to scale than application servers so from a technical standpoint, it makes sense to do this glue work in app code.
> a transaction doesn't guarantee that all statements in it see the data at the same point in time
This depends on the transaction isolation level. If you use snapshot or serializable this should be the case (but you may have aborted transactions due to optimistic concurrency).
Doesn’t it guarantee consistency from the time the transaction started (assuming read committed isolation)? It guarantees you won’t see something “later” than when your transaction began.
Read committed (which is the default), doesn't guarantee that.
See "Nonrepeatable Read" and "Phantom Read" which are both possible in your documentation page.
Also, often, the transactional database servers is more difficult to scale than application servers so from a technical standpoint, it makes sense to do this glue work in app code.