Does anyone know what the expected range of the "penalty" GiST method
is? (i.e. Is the legal range documented anywhere? Failing that, what
does existing GiST-based code expect?)

While rewriting gistchoose() in gist.c to be less obfuscated, it
occurred to me that (a) I don't think the existing code will work as
intended if a negative penalty is returned (b) it would be good to
define a "minimum legal penalty". Once the minimum penalty has been
returned for a particular location, GiST can know that all other
locations where it might insert the node will have a penalty greater
than or equal to that value, so gistchoose() can bail-out earlier.

Therefore, I'd like to make "0.0" is the minimum legal penalty, and
require all GiST "penalty" methods to return values >= 0. I think
requiring the penalty != NaN would also be a good idea. Comments?

Note that making this change to CVS tip results in a regression failure
in contrib/btree_gist. The regression.diffs are attached. The regression
tests for rtree_gist, tsearch, tsearch2 and pg_trgm succeed after making
the modification.

-Neil

*** ./expected/bit.out	2004-11-09 11:20:51 +11:00
--- ./results/bit.out	2004-11-10 15:44:52 +11:00
***************
*** 33,38 ****
--- 33,39 ----
  (1 row)
  
  CREATE INDEX bitidx ON bittmp USING GIST ( a );
+ ERROR:  illegal penalty from GiST penalty method: -116307439489975337053247905800904507392.000000
  SET enable_seqscan=off;
  SELECT count(*) FROM bittmp WHERE a <   '011011000100010111011000110000100';
   count 

======================================================================

*** ./expected/varbit.out	2004-11-09 11:20:51 +11:00
--- ./results/varbit.out	2004-11-10 15:44:52 +11:00
***************
*** 33,38 ****
--- 33,39 ----
  (1 row)
  
  CREATE INDEX varbitidx ON varbittmp USING GIST ( a );
+ ERROR:  illegal penalty from GiST penalty method: -92381335565846851340979378715858305024.000000
  SET enable_seqscan=off;
  SELECT count(*) FROM varbittmp WHERE a <   '1110100111010'::varbit;
   count 

======================================================================

---------------------------(end of broadcast)---------------------------
TIP 1: subscribe and unsubscribe commands go to [EMAIL PROTECTED]

Reply via email to