Re: Boolean partitions syntax

On 2018/04/11 10:44, Tom Lane wrote:
> Kyotaro HORIGUCHI <horiguchi(dot)kyotaro(at)lab(dot)ntt(dot)co(dot)jp> writes:
>> At least partition bound *must* be a constant. Any expression
>> that can be reduced to a constant at parse time ought to be
>> accepted but must not be accepted if not.
>
> My point is that *any* expression can be reduced to a constant,
> we just have to do so.

Currently transformPartitionBoundValue() applies eval_const_expressions()
by way of calling expression_planner(). However passing to it, say, an
expression representing random() is unable to reduce it to a Const because
simplify_function/evaluate_function won't compute a mutable function like
random(). So, that currently results in an error like this (note that
this is after applying Horiguchi-san's latest patch that enables
specifying random() as a partition bound in the first place):

create table foo_part partition of foo for values in ((random())::int);
ERROR: specified value cannot be cast to type integer for column "a"
LINE 1: ...table foo_random partition of foo for values in ((random()):...
^
DETAIL: The cast requires a non-immutable conversion.
HINT: Try putting the literal value in single quotes.

The error is output after the following if check in
transformPartitionBoundValue fails:

/* Fail if we don't have a constant (i.e., non-immutable coercion) */
if (!IsA(value, Const))

I think what Tom is proposing here, instead of bailing out upon
eval_const_expressions() failing to simplify the input expression to a
Const, is to *invoke the executor* to evaluate the expression, like the
optimizer does in evaluate_expr, and cook up a Const with whatever comes
out to store it into the catalog (that is, in relpartbound).

Thanks,
Amit