Hannu Krosing <hannu@skype.net> writes:
> But can't we _define_ such a subset, where we can do a transactionless
> load ?
Sure ... but you'll find that it's not large enough to be useful.
Once you remove all the interesting consistency checks such as
unique indexes and foreign keys, the COPY will tend to go through
just fine, and then you're still stuck trying to weed out bad data
without very good tools for it. The only errors we could really
separate out without subtransaction fencing are extremely trivial
ones like too many or too few fields on a line ... which can be
caught with a sed script.
regards, tom lane