On Thu, Feb 1, 2018 at 11:39 AM, Peter Geoghegan <pg@bowt.ie> wrote:
> There is also the matter of subselects in the update targetlist, which
> you similarly claim is only a problem of having the wrong error
> message. The idea that those are unsupported for any principled reason
> doesn't have any justification (unlike WHEN ... AND quals, and their
> restrictions, which I agree are necessary). It clearly works with
> Oracle, for example:
>
> http://sqlfiddle.com/#!4/2d5405/10
>
> You're reusing set_clause_list in the grammar, so I don't see why it
> shouldn't work within MERGE in just the same way as it works in
> UPDATE.
Actually, I now wonder if there is a good reason for restrictions
(e.g. no subselects) on WHEN ... AND quals, too. See this SQL fiddle
from SQL Server:
http://sqlfiddle.com/#!18/8acef/27
I started looking at SQL Server's MERGE to verify that it also does
not impose any restrictions on subselects in a MERGE UPDATE's
targetlist, just like Oracle. Unsurprisingly, it does not. More
surprisingly, I noticed that it also doesn't seem to impose
restrictions on what can appear in WHEN ... AND quals. Most
surprisingly of all, even the main join ON condition itself can have
subselects (though that's probably a bad idea).
What this boils down to is that I don't think that this part of your
design is committable (from your recent v14):
>> + * As top-level statements INSERT, UPDATE and DELETE have a Query,
>> + * whereas with MERGE the individual actions do not require
>> + * separate planning, only different handling in the executor.
>> + * See nodeModifyTable handling of commandType CMD_MERGE.
>> + *
>> + * A sub-query can include the Target, but otherwise the sub-query
>> + * cannot reference the outermost Target table at all.
>> + */
>> + qry->querySource = QSRC_PARSER;
>> + joinexpr = makeNode(JoinExpr);
>> + joinexpr->isNatural = false;
>> + joinexpr->alias = NULL;
>> + joinexpr->usingClause = NIL;
>> + joinexpr->quals = stmt->join_condition;
I am willing to continue to engage with you on the concurrency issues
for the time being, since that is the most pressing issue for the
patch. We can get to this stuff later. Note that I consider cleaning
up the Query and plan representations to be prerequisite to commit.
--
Peter Geoghegan