On Mon, Jun 3, 2019 at 5:10 PM Melanie Plageman
<melanieplageman@gmail.com> wrote:
> I was talking to Jeff Davis about this on Saturday, and, he felt that
> there might be a way to solve the problem differently if we thought of
> the left join case as performing an inner join and an antijoin
> instead.
>
> Riffing on this idea a bit, I started trying to write a patch that
> would basically emit a tuple if it matches and write the tuple out to
> a file if it does not match. Then, after iterating through the outer
> batch the first time for the first inner chunk, any tuples which do
> not yet have a match are the only ones which need to be joined against
> the other inner chunks. Instead of iterating through the outer side
> original batch file, use the unmatched outer tuples file to do the
> join against the next chunk. Repeat this for all chunks.
I'm not sure that I understanding this proposal correctly, but if I am
then I think it doesn't work in the case where a single outer row
matches rows in many different inner chunks. When you "use the
unmatched outer tuples file to do the join against the next chunk,"
you deny any rows that have already matched the chance to produce
additional matches.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company