On 2015-04-22 16:40:07 -0700, Peter Geoghegan wrote: > On Wed, Apr 22, 2015 at 3:23 PM, Peter Geoghegan <pg@heroku.com> wrote: > > * We need to sort out those issues with the grammar, since that only > > really applies to the inference specification. Maybe the WHERE clause > > that the inference specification accepts can be broken out. No ON > > CONFLICT UPDATE specific issues left there, AFAICT though. > > I pushed some code that deals with the predicate being within parenthesis: > > https://github.com/petergeoghegan/postgres/commit/358854645279523310f998dfc9cb3fe3e165ce1e
And the way you've used nonassoc here doesn't look correct. You're hiding legitimate ambiguities in the grammar. UPDATE is a unreserved keyword, so for
... ON CONFLICT '(' index_params ')' where_clause OnConflictUpdateStmt
it won't be able to discern whether an UPDATE in the WHERE clause is part of the where_clause or OnConflictUpdate.
This is legal: SELECT * FROM (SELECT true as update) f WHERE update; i.e. 'update' can be the last part of a WHERE clause.
Essentially what you're trying to do with the nonassic is hiding that UPDATE and IGNORE need to be reserved keywords with the syntax you're proposing. We can either make them reserved or change the syntax.
One way to avoid making them reserved keywords - which would be somewhat painful - is to add a 'DO' before the IGNORE/UPDATE. I.e. something like
ON CONFLICT opt_conflict_expr DO OnConflictUpdateStmt | ON CONFLICT opt_conflict_expr DO IGNORE
