On Tue, Aug 28, 2018 at 8:02 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> I think this argument is a red herring TBH. The example Robert shows is
> of *zero* interest for dynahash or catcache, unless it's taking only the
> low order 3 bits of the OID for the bucket number. But actually we'll
> increase the table size proportionally to the number of entries, so
> that you can't have say 1000 table entries without at least 10 bits
> being used for the bucket number. That means that you'd only have
> trouble if those 1000 tables all had OIDs exactly 1K (or some multiple
> of that) apart. Such a case sounds quite contrived from here.
Hmm. I was thinking that it was a problem if the number of OIDs
consumed per table was a FACTOR of 1000, not just if it was a POWER of
1000. I mean, if it's, say, 4, that means three-quarters of your hash
table buckets are unused, which seems poor. But maybe it's not really
a big enough problem in practice for us to care? Dunno.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company