Kenneth Marshall <ktm@rice.edu> writes:
> ... This is the rough plan. Does anyone see anything critical that
> is missing at this point?
Sounds pretty good. Let me brain-dump one item on you: one thing that
hash currently has over btree is the ability to handle index items up
to a full page. Now, if you go with a scheme that only stores hash
codes and not the underlying data, you can not only handle that but
improve on it; but if you reduce the bucket size and don't remove the
data, it'd be a step backward. The idea I had about dealing with that
was to only reduce the size of primary buckets --- if it's necessary to
add overflow space to a bucket, the overflow units are still full pages.
So an index tuple up to a page in size can always be accommodated by
adding an overflow page to the bucket.
Just a thought, but AFAIR it's not in the archives anywhere.
regards, tom lane