"Jim C. Nasby" <jnasby@pervasive.com> writes:
> If there was more information than the tuplestore could keep in memory,
> then a TIDstore might be faster, but only if it resulted in reading from
> the heap sequentially, or very near it.
That's easily arranged, use a bitmap indexing data structure.
I think we could probably even live with the structure becoming lossy
under memory pressure: AFAICS, all rows modified by a single query ought
to have the same XMIN/CMIN (or XMAX/CMAX for deleted rows), so it should
be possible to verify whether a particular row is one of the interesting
ones or not.
I think the hard part of this task is designing the API for access to
the rowsets from triggers.
regards, tom lane