On 2023-06-30 Fr 08:46, Tom Lane wrote:
Andrew Dunstan <andrew@dunslane.net> writes:
On 2023-06-29 Th 18:41, Tom Lane wrote:
Why not make the hash key be the value
itself? Wrap it in a bytea perhaps to avoid needing a bespoke
hash function.
Not sure I understand.
Say the missingval for a particular column is text 'abc'.
We don't actually care which column it is, all we need is a
copy of that datum that will stay put for the rest of the
transaction. So I'm thinking that the lookup key for the
hash table should actually be the contents of the datum,
and we don't need to store anything else at all. (If we
happen to have two columns with the same missingval, they
can perfectly well share this hash entry.) Then there's
no question of invalidation, or at least the existing
invalidation mechanisms for tupdescs do all we need.
OK, I get it. Do we have a routine to wrap a Datum in a bytea, or do I need to write one? It's slightly amusing that the original patch for fast defaults stored the missing value as a bytea, but someone (Andres IIRC) preferred that we store it as a one element array, so that's what we went with.
cheers
andrew
--
Andrew Dunstan
EDB: https://www.enterprisedb.com