Обсуждение: Implement predicate propagation for non-equivalence clauses

On 05/09/18 09:34, Richard Guo wrote:
> Hi,
> 
> As we know, current planner will generate additional restriction clauses 
> from
> equivalence clauses. This will generally lower the total cost because 
> some of
> tuples may be filtered out before joins.
> 
> In this patch, we are trying to do the similar deduction, from 
> non-equivalence
> clauses, that is, A=B AND f(A) implies A=B AND f(A) and f(B), under some
> restrictions on f.

I haven't read the patch in detail, but that really only applies under 
special circumstances. Tom caught me making that assumption just 
recently 
(https://www.postgresql.org/message-id/8003.1527092720%40sss.pgh.pa.us). 
I think the restriction here is that f(x) must be an operator that's in 
the same operator family as the = operator. In a quick read-through, 
it's not clear to me what conditions are in the patch now. Please have a 
comment somewhere to list them explicitly.

> This patch will introduce extra cost for relation scan, due to the
> cost of evaluating the new implied quals. Meanwhile, since the extra
> filter may reduce the number of tuples returned by the scan, it may
> lower the cost of following joins. So, whether we will get a better
> plan depends on the selectivity of the implied quals.
Perhaps we should evaluate the selectivity of the clause, and only add 
them if they seem helpful, based on the cost vs. selectivity?

At least in this case from the regression tests:

>  explain (costs off)
>    select * from ec0 a, ec1 b
>    where a.ff = b.ff and a.ff = 43::bigint::int8alias1;
> -                 QUERY PLAN                  
> ----------------------------------------------
> +                              QUERY PLAN                              
> +----------------------------------------------------------------------
>   Nested Loop
>     ->  Index Scan using ec0_pkey on ec0 a
>           Index Cond: (ff = '43'::int8alias1)
>     ->  Index Scan using ec1_pkey on ec1 b
>           Index Cond: (ff = a.ff)
> -         Filter: (f1 < '5'::int8alias1)
> +         Filter: ((f1 < '5'::int8alias1) AND (ff = '43'::int8alias1))
>  (6 rows)

the new qual is redundant with the Index Condition. If we could avoid 
generating such redundant quals, that would be good.

- Heikki

Richard Guo <riguo@pivotal.io> writes:
> In this patch, we are trying to do the similar deduction, from
> non-equivalence
> clauses, that is, A=B AND f(A) implies A=B AND f(A) and f(B), under some
> restrictions on f.

Uh, *what* restrictions on f()?  In general the above equivalence
does not hold, at least not for any data type more complicated than
integers; and we do not have any semantic model for deciding
which functions it would be correct for.

One simple example to show what I'm talking about is that float8 zero
and minus zero are equal according to float8eq (assuming IEEE float
arithmetic); but they aren't equivalent for any function f() that is
sensitive to the sign or the text representation of the value.
The numeric data type likewise has values that are "equal" without
being identical for all purposes, eg 0.0 vs 0.000.  Or consider
citext.

The existing planner deduction rules for equivalence classes are
carefully designed to ensure that we only generate derived clauses
using operators from the same operator class or family, so that
it's on the opclass author to ensure that the operators have consistent
semantics.  I don't think we can extrapolate from that to any random
function that accepts the datatype.

            regards, tom lane

Hi Richard,

I took a look at the v2, here are some comments:

* This feature needs tests, especially for the cases where opfamilies or 
data types or collations don't match, and other non-obvious cases where 
it shouldn't work.


* Deducing an inequality to a constant is not always helpful. If we know 
that a = b and a = const1 and b < const2, we will deduce b = const1 from 
the EC, and adding b < const2 doesn't improve selectivity and only makes 
the cost estimate worse. One situation where it does make sense is when 
we can detect contradictions in these clauses, e.g. if we know that 
const1 > const2 and therefore know that the above selection clause is 
always false. Looking at the regression test changes, I see that v2 
doesn't do that. I think the handling of deduced inequalities shoud be 
modeled on the flow of generate_base_implied_equalities -> 
process_implied_equality -> distribute_qual_to_rels. This would allow us 
to correctly handle a deduced const1 < const2 clause and turn it into a 
gating Result qual.


* There are functions named generate_implied_quals, gen_implied_quals 
and gen_implied_qual. The names are confusingly similar, we could use 
something like generate_implied_quals_for_clause and 
generate_one_implied_qual for the latter two.


@@ gen_implied_quals
     else if (clause && IsA(clause, ScalarArrayOpExpr))
     {
* When can the clause be NULL?


@@ gen_implied_quals
     item1 = canonicalize_ec_expression(item1,
                                        exprType((Node *) item1),
                                        collation);
     item2 = canonicalize_ec_expression(item2,
                                        exprType((Node *) item2),
                                        collation);
* Why do we do this? If the collation or type of the original expression 
isn't right and we have to add RelabelType, the resulting expression 
won't match the original clause and we won't be able to substitute it 
with other equavalence members. So we can just check that the type and 
collation match.


* In gen_implied_quals, I'd rename item1 => left, item2 => right, em1 => 
orig_em, em2 => other_em, and same for list cells and types. As it is 
now, em1 can actually match both item1 and item2, and em2 is not related 
to either of them, so it took me some effort to understand what's going on.


* In gen_implied_qual, why do we search the clause for a matching 
subexpression? Reading the thread, I thought that we can only do the 
substitution for OpExprs of the same opfamilies as the generating EC. 
This code looks like it can do the substitution an at arbitrary depth, 
so we might change an argument of some unsuitable function, and the 
result will not be correct. What we should probably do is that after we 
matched one side of OpExpr to one EC member, we just replace it with 
another suitable member and add the resulting clause.


@@ gen_implied_qual
     check_mergejoinable(new_rinfo);
     check_hashjoinable(new_rinfo);
...
     /*
      * If the clause has a mergejoinable operator, set the EquivalenceClass
      * links. Otherwise, a mergejoinable operator with NULL 
left_ec/right_ec
      * will cause update_mergeclause_eclasses fails at assertion.
      */
     if (new_rinfo->mergeopfamilies)
         initialize_mergeclause_eclasses(root, new_rinfo);
* It's not an equality clause, so it is not going to be mergejoinable 
nor hashjoinable and we can skip these checks.

-- 
Alexander Kuzmenkov
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company

Hi,

On 2018-11-22 19:15:42 +0300, Alexander Kuzmenkov wrote:
> Hi Richard,
> 
> I took a look at the v2, here are some comments:
> 
> * This feature needs tests, especially for the cases where opfamilies or
> data types or collations don't match, and other non-obvious cases where it
> shouldn't work.
> 
> 
> * Deducing an inequality to a constant is not always helpful. If we know
> that a = b and a = const1 and b < const2, we will deduce b = const1 from the
> EC, and adding b < const2 doesn't improve selectivity and only makes the
> cost estimate worse. One situation where it does make sense is when we can
> detect contradictions in these clauses, e.g. if we know that const1 > const2
> and therefore know that the above selection clause is always false. Looking
> at the regression test changes, I see that v2 doesn't do that. I think the
> handling of deduced inequalities shoud be modeled on the flow of
> generate_base_implied_equalities -> process_implied_equality ->
> distribute_qual_to_rels. This would allow us to correctly handle a deduced
> const1 < const2 clause and turn it into a gating Result qual.
> 
> 
> * There are functions named generate_implied_quals, gen_implied_quals and
> gen_implied_qual. The names are confusingly similar, we could use something
> like generate_implied_quals_for_clause and generate_one_implied_qual for the
> latter two.
> 
> 
> @@ gen_implied_quals
>     else if (clause && IsA(clause, ScalarArrayOpExpr))
>     {
> * When can the clause be NULL?
> 
> 
> @@ gen_implied_quals
>     item1 = canonicalize_ec_expression(item1,
>                                        exprType((Node *) item1),
>                                        collation);
>     item2 = canonicalize_ec_expression(item2,
>                                        exprType((Node *) item2),
>                                        collation);
> * Why do we do this? If the collation or type of the original expression
> isn't right and we have to add RelabelType, the resulting expression won't
> match the original clause and we won't be able to substitute it with other
> equavalence members. So we can just check that the type and collation match.
> 
> 
> * In gen_implied_quals, I'd rename item1 => left, item2 => right, em1 =>
> orig_em, em2 => other_em, and same for list cells and types. As it is now,
> em1 can actually match both item1 and item2, and em2 is not related to
> either of them, so it took me some effort to understand what's going on.
> 
> 
> * In gen_implied_qual, why do we search the clause for a matching
> subexpression? Reading the thread, I thought that we can only do the
> substitution for OpExprs of the same opfamilies as the generating EC. This
> code looks like it can do the substitution an at arbitrary depth, so we
> might change an argument of some unsuitable function, and the result will
> not be correct. What we should probably do is that after we matched one side
> of OpExpr to one EC member, we just replace it with another suitable member
> and add the resulting clause.
> 
> 
> @@ gen_implied_qual
>     check_mergejoinable(new_rinfo);
>     check_hashjoinable(new_rinfo);
> ...
>     /*
>      * If the clause has a mergejoinable operator, set the EquivalenceClass
>      * links. Otherwise, a mergejoinable operator with NULL left_ec/right_ec
>      * will cause update_mergeclause_eclasses fails at assertion.
>      */
>     if (new_rinfo->mergeopfamilies)
>         initialize_mergeclause_eclasses(root, new_rinfo);
> * It's not an equality clause, so it is not going to be mergejoinable nor
> hashjoinable and we can skip these checks.

As this review has not been addressed, I'm marking this patch as
returned with feedback for now. Please resubmit once these are
addressed!

Greetings,

Andres Freund