Re: Boom filters for hash joins (was: A design for amcheck heapam verification)

On Mon, Sep 18, 2017 at 5:13 PM, Peter Geoghegan <pg(at)bowt(dot)ie> wrote:
> On Mon, Sep 18, 2017 at 2:07 PM, Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
>> On Mon, Sep 18, 2017 at 1:29 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>>> Uh, why does the planner need to be involved at all?
>>
>> Because it loses if the Bloom filter fails to filter anything. That's
>> not at all far-fetched; consider SELECT * FROM a.x, b.x WHERE a.x =
>> b.x given a foreign key on a.x referencing b(x).
>
> Wouldn't a merge join be a lot more likely in this case anyway? Low
> selectivity hash joins with multiple batches are inherently slow; the
> wasted overhead of using a bloom filter may not matter.
>
> Obviously this is all pretty speculative. I suspect that this could be
> true, and it seems worth investigating that framing of the problem
> first.

ISTR Tomas Vondra doing some experiments with this a few years ago and
finding that it was, in fact, a problem.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company