> Within the data to import most rows have 20 till 50 duplicates. Sometime
> much more, sometimes less.
In that case (source data has lots of redundancy), after importing the
data chunks in parallel, you can run a first pass of de-duplication on the
chunks, also in parallel, something like :
CREATE TEMP TABLE foo_1_dedup AS SELECT DISTINCT * FROM foo_1;
or you could compute some aggregates, counts, etc. Same as before, no WAL
needed, and you can use all your cores in parallel.
From what you say this should reduce the size of your imported data by a
lot (and hence the time spent in the non-parallel operation).
With a different distribution, ie duplicates only between existing and
imported data, and not within the imported data, this strategy would be
useless.