Обсуждение: selectivity calculation for or_clause is wrong ?

Hi all,

I have a question about the calculation of selectivity.

I see the following code in set_rest_selec() in clausesel.c.
               cost_clause = clausenode->selectivity;
    /*                * Check to see if the selectivity of this clause or any
'or'                * subclauses (if any) haven't been set yet.                */               if (cost_clause <= 0 ||
valid_or_clause(clausenode))              {
 

Why is valid_or_clause(clausenode) necessary ?

This means that even if selectivity is set,set_rest_selec()
calls compute_clause_selec() if the target clause is a
valid_or_clause.
compute_clause_selec() would add the selectivity of
elements of or_clause to the current selectivity.

AFAIC,compute_clause_selec() is called twice at least   ( from add_restrict_and_join_to_rel() in initsplan.c      and
set_rest_selec()in clausesel.c)
 
and seems to accumulate the result by repetition if
the target clause is a valid_or_clause.

Regards.

Hiroshi Inoue
Inoue@tpf.co.jp

"Hiroshi Inoue" <Inoue@tpf.co.jp> writes:
> Why is valid_or_clause(clausenode) necessary ?

Looks like a waste of cycles to me too.

If the subclauses of an OR could get rearranged during optimization
then this might be a necessary check, but AFAIK they don't.
        regards, tom lane

> -----Original Message-----
> From: Tom Lane [mailto:tgl@sss.pgh.pa.us]
> Sent: Wednesday, July 21, 1999 12:32 AM
> To: Hiroshi Inoue
> Cc: pgsql-hackers
> Subject: Re: [HACKERS] selectivity calculation for or_clause is wrong ? 
> 
> 
> "Hiroshi Inoue" <Inoue@tpf.co.jp> writes:
> > Why is valid_or_clause(clausenode) necessary ?
> 
> Looks like a waste of cycles to me too.
>

It's not only a waste of cycles.
For exmaple,

1.explain select key1 from b where someitem in (1);
  NOTICE:  QUERY PLAN:
  Seq Scan on b on b  (cost=1638.49 rows=261 width=4)

2.explain select key1 from b where someitem in (1,2);
  NOTICE:  QUERY PLAN:
  Seq Scan on b on b  (cost=1638.49 rows=773 width=4)

3.explain select key1 from b where someitem in (1,2,3);
 NOTICE:  QUERY PLAN:
 Seq Scan on b on b  (cost=1638.49 rows=1274 width=4)


rows of each plan  261 : 773 : 1274  not =  1 : 2 : 3.
It's nearly = 1 :3 :5.

elements of or_clause except its first element are evaluated 
twice and the results are accumlated.

Regards.

Hiroshi Inoue
Inoue@tpf.co.jp

"Hiroshi Inoue" <Inoue@tpf.co.jp> writes:
>>>> Why is valid_or_clause(clausenode) necessary ?
>> 
>> Looks like a waste of cycles to me too.
>
> It's not only a waste of cycles.
> [ snip ]
> rows of each plan  261 : 773 : 1274  not =  1 : 2 : 3.
> It's nearly = 1 :3 :5.
> elements of or_clause except its first element are evaluated 
> twice and the results are accumlated.

BTW, I fixed this...
        regards, tom lane