Paul Jungwirth wrote:
Missed number of digit, but nevertheless it doesn't matter for idea.
Original coding uses only 8 bytes from 16 to compute penalty which could
cause a problem with index performance. Simple way is just printing each
4bits with %02d modifier into string and then make a numeric value with
a help of numeric_in.
uuid2num(const pg_uuid_t *i)
return *((uint64 *)i);
It isn't looked as correct transformation for me. May be, it's better
to transform to numeric type (UUID looks like a 16-digit hexademical
and follow gbt_numeric_penalty() logic (or even call directly).
Thanks for the review! A UUID is actually not stored as a string of
hexadecimal digits. (It is normally displayed that way, but with 32
digits, not 16.) Rather it is stored as an unstructured 128-bit value
(which in C is 16 unsigned chars). Here is the easy-to-misread
declaration from src/backend/utils/adt/uuid.c:
Or something like this in pseudocode:
numeric = int8_numeric(*(uint64 *)(&i->data)) *
int8_numeric(MAX_INT64) + int8_numeric(*(uint64 *)(&i->data))
Right, but precision of double is enough to represent 1 century
interval with 0.00001 seconds accuracy which is enough for practical
usage. In UUID case you will take into account only half of value. Of
course, GiST will work even with penalty function returning constant but
each scan could become full-index-scan.
The only other 128-bit type I found in btree_gist was Interval. For that
type we convert to a double using INTERVAL_TO_SEC, then call
penalty_num. By my read that accepts a similar loss of precision.
If I'm mistaken about 128-bit integer support, let me know, and maybe we
can do the penalty computation on the whole UUID. Or maybe I should just
convert the uint64 to a double before calling penalty_num? I don't
completely understand what the penalty calculation is all about, so I
welcome suggestions here.
Penalty method calculates how union key will be enlarged if insert will
be produced in current subtree. It directly affects selectivity of subtree.
Teodor Sigaev E-mail: teo...@sigaev.ru
Sent via pgsql-hackers mailing list (firstname.lastname@example.org)
To make changes to your subscription: