[bug] parenthesized left operand of UNION/EXCEPT/INTERSECT is rejected

[therealpandey]

Summary

A set operator after a parenthesized SELECT never binds: parseSelectQuery checks for UNION/EXCEPT/INTERSECT before consuming the closing ), so any parenthesized left operand is rejected.

Reproduces on v0.5.5 and on master (44047b0).

Reproduction

_, err := parser.NewParser("SELECT a FROM ((SELECT 1 AS a) UNION ALL (SELECT 2 AS a))").ParseStmts()

line 1:32 expected ')', but got '<keyword>'
SELECT a FROM ((SELECT 1 AS a) UNION ALL (SELECT 2 AS a))
                               ^^^^^

Rejected:

SELECT a FROM ((SELECT 1 AS a) UNION ALL (SELECT 2 AS a))       -- expected ')', but got '<keyword>'
SELECT a FROM ((SELECT 1 AS a) UNION DISTINCT (SELECT 2 AS a))  -- expected ')', but got '<keyword>'
SELECT a FROM ((SELECT 1 AS a) EXCEPT (SELECT 2 AS a))          -- expected ')', but got '<keyword>'
SELECT a FROM ((SELECT 1 AS a) INTERSECT (SELECT 2 AS a))       -- expected ')', but got '<keyword>'
(SELECT 1 AS a) UNION ALL SELECT 2 AS a                         -- unexpected token: "("
SELECT 1 UNION ALL (SELECT 2) UNION ALL SELECT 3                -- <EOF> or ';' was expected, but got: "UNION"
SELECT a FROM (((SELECT 1 AS a)))                               -- expected keyword <"SELECT">, but got '('

Accepted:

SELECT a FROM (SELECT 1 AS a UNION ALL SELECT 2 AS a)
SELECT a FROM (SELECT 1 AS a UNION ALL (SELECT 2 AS a))
SELECT a FROM ((SELECT 1 AS a))
SELECT 1 AS a UNION ALL (SELECT 2 AS a)

ClickHouse accepts these

clickhouse local -q "SELECT a FROM ((SELECT 1 AS a) UNION ALL (SELECT 2 AS a))"

Every rejected statement above runs on 26.8.1.337.

Proposed fix

parseSelectQuery in parser/parser_query.go:

hasParen := p.tryConsumeTokenKind(TokenKindLParen) != nil
selectStmt, err := p.parseSelectStmt(p.Pos())
if err != nil {
	return nil, err
}
switch {
case p.tryConsumeKeywords(KeywordUnion):
	// ...
}
if hasParen {
	if err := p.expectTokenKind(TokenKindRParen); err != nil {
		return nil, err
	}
}

1. On (, recurse into parseSelectQuery and consume ) first, then bind the set operator, attaching the right operand to the tail of the chain the left operand already carries.
2. Dispatch statements starting with ( to parseSelectQuery in parseStmt (parser/parser_table.go).
3. Fixtures under parser/testdata/query/.

[therealpandey]

The fix in #312 ended up deviating from the proposal in this issue's description, and it's worth recording why.

The proposed fix parsed the parenthesized operand and attached the operator's right operand to the tail of the chain the operand already carried — flattening (A UNION ALL B) INTERSECT C into the same AST as A UNION ALL B INTERSECT C, with the parentheses dropped. That makes every statement in this issue parse, and for pure UNION/EXCEPT sequences the flattening is harmless because those associate left at equal precedence.

What it silently gets wrong is INTERSECT. ClickHouse gives INTERSECT higher precedence than UNION and EXCEPT, so once the parentheses are gone, formatting the AST back to SQL produces a query with a different result set:

$ clickhouse local -q "(SELECT 1 UNION ALL SELECT 2) INTERSECT SELECT 2"
2
$ clickhouse local -q "SELECT 1 UNION ALL SELECT 2 INTERSECT SELECT 2"
1
2

With the flattening fix, parser.Format would turn the first query into the second — a parse→format round trip that changes what the query returns, which is worse than the parse error it replaced. The mirrored case exists on master today, independent of this issue: SELECT 1 INTERSECT (SELECT 2 UNION ALL SELECT 1) already parses, flattens, and formats to a query returning different rows.

Since the flat UnionAll/UnionDistinct/Except/Intersect chain on SelectQuery has no way to express grouping, #312 adds one: a parenthesized operand is kept as a SelectQuery whose new Paren field wraps the inner query, the operator after ) binds to the wrapper, and FormatSQL reproduces the parentheses. That also fixes the pre-existing right-operand case above, and every statement in the PR's format golden now round-trips to exactly its input text.