Re: [POC] A better way to expand hash indexes.

On Thu, Mar 30, 2017 at 7:23 PM, Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
> I think approach B is incorrect. Suppose we have 1536 buckets and
> hash values 2048, 2049, 4096, 4097, 6144, 6145, 8192, and 8193. If I
> understand correctly, each of these values should be mapped either to
> bucket 0 or to bucket 1, and the goal of the sort is to put all of the
> bucket 0 tuples before all of the bucket 1 tuples, so that we get
> physical locality when inserting. With approach A, the sort keys will
> match the bucket numbers -- we'll be sorting the list 0, 1, 0, 1, 0,
> 1, 0, 1 -- and we will end up doing all of the inserts to bucket 0
> before any of the inserts to bucket 1. With approach B, we'll be
> sorting 512, 513, 1024, 1025, 0, 1, 512, 513 and will end up
> alternating inserts to bucket 0 with inserts to bucket 1.

Oops sorry, yes 2 denominators are different (one used in an insert
and another used in sorting keys) we will end up with different bucket
numbers. I think in patch B, I should have actually taken next 2-power
number of 1536 as the denominator and try to get the mod value. If the
mod value is > 1536 then reduce the denominator by half and retake the
mod to get the bucket within 1536. Which is what effectively Patch A
is doing. Approach B is a blunder, I apologize for that mistake. I
think Patch A should be considered. If adding the members of struct
Tuplesortstate is a concern I will rewrite Patch B as said above.

--
Thanks and Regards
Mithun C Y
EnterpriseDB: http://www.enterprisedb.com