Re: Index AM change proposals, redux

> Teodor, do you have any thoughts about exactly how you'd fix @@@ ?
> I suppose that the recheck-need is not really a property of specific
> tuples, but of a particular query, for that case. Where would you
> want to detect that?

tsquery may include restriction by weight of search terms: 'sea & port:A'. GIN
index doesn't store information about weights, so the only difference between @@
and @@@ is that @@@ is marked with RECHECK flag. I think, the better way is set
flag about required recheck by looking value from index, not for tsquery. It
gives to us more flexibility.

So, I planned to add pointer to bool to consistent method, so signature will be
bool consistent( bool check[], StrategyNumber n, Datum query, bool *needRecheck)

Returning value of needRecheck should be ignored for operation not marked by
RECHECK flag in opclass. needRecheck should be initialized to true before call
of consistent method to keep compatibility with old opclasses.

To define, is recheck needed or not, the better way is to check actually needed
values. For example, let tsquery is equal to
'foo | bar | qq:A' and tsvetor = 'foo:1,2,3 asdasdasd:4'. Obviously recheck is
not needed. So patch is close to trivial:

*** tsginidx.c.orig 2008-04-11 17:08:37.000000000 +0400
--- tsginidx.c 2008-04-11 17:18:45.000000000 +0400
***************
*** 109,114 ****
--- 109,115 ----
{
QueryItem *frst;
bool *mapped_check;
+ bool *needRecheck;
} GinChkVal;

static bool
***************
*** 116,121 ****
--- 117,125 ----
{
GinChkVal *gcv = (GinChkVal *) checkval;

+ if ( val->weight )
+ *(gcv->needRecheck) = true;
+
return gcv->mapped_check[((QueryItem *) val) - gcv->frst];
}

***************
*** 144,149 ****
--- 148,155 ----

gcv.frst = item = GETQUERY(query);
gcv.mapped_check = (bool *) palloc(sizeof(bool) * query->size);
+ gcv.needRecheck = PG_GETARG_POINTER(3);
+ *(gcv.needRecheck) = false;

for (i = 0; i < query->size; i++)
if (item[i].type == QI_VAL)

--
Teodor Sigaev E-mail: teodor(at)sigaev(dot)ru
WWW: http://www.sigaev.ru/