Re: <> join selectivity estimate question

On Fri, Mar 17, 2017 at 6:49 PM, Thomas Munro
<thomas(dot)munro(at)enterprisedb(dot)com> wrote:
> Right. If I temporarily hack neqjoinsel() thus:
>
> result = 1.0 - result;
> +
> + if (jointype == JOIN_SEMI)
> + result = 1.0;
> +
> PG_RETURN_FLOAT8(result);
> }

I was looking into this problem. IMHO, the correct solution will be
that for JOIN_SEMI, neqjoinsel should not estimate the equijoin
selectivity using eqjoinsel_semi, instead, it should calculate the
equijoin selectivity as inner join and it should get the selectivity
of <> by (1-equijoin selectivity). Because for the inner_join we can
claim that "selectivity of '=' + selectivity of '<>' = 1", but same is
not true for the semi-join selectivity. For semi-join it is possible
that selectivity of '=' and '<>' is both are 1.

something like below
----------------------------

@@ -2659,7 +2659,13 @@ neqjoinsel(PG_FUNCTION_ARGS)
SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
Oid eqop;
float8 result;

+ if (jointype = JOIN_SEMI)
+ {
+ sjinfo->jointype = JOIN_INNER;
+ }
/*
* We want 1 - eqjoinsel() where the equality operator is the one
* associated with this != operator, that is, its negator.

We may need something similar for anti-join as well.

--
Regards,
Dilip Kumar
EnterpriseDB: http://www.enterprisedb.com