[IMPALA-7964] Inconsistent typing of pure-constant expressions

• Type: Bug
• Status: Open
• Priority: Minor
• Resolution: Unresolved
• Affects Version/s: Impala 3.1.0
• Component/s: Frontend
• Epic Color: ghx-label-4
• Assignee: Unassigned
• Reporter: Paul Rogers
• Created: 12/Dec/18 21:39
• Updated: 01/Mar/19 17:50

Description

Impala implements a very subtle rule for computing the type of constant expressions. See IMPALA-4213. The desired rules when performing constant folding are:

• Pure-literal numeric expressions take the natural (smallest) type of the result.
• All other expressions preserve the type of the expression.

Examples:

 -- simple literal
100 -> 100:TINYINT

 -- explicit cast
CAST(100 AS SMALLINT) --> 100:SMALLINT

 -- pure literal expression
2 + 3 --> 5:TINYINT

  -- addition is done at next wider type up from that of tinyint_col
tinyint_col + 3 --> SMALLINT

 -- explicit type given, math done with that type, mimics above
CAST(2 AS TINYINT) + 3 --> 5:SMALLINT

 -- Impala uses C++ semantics, with overflow
CAST(257 AS TINYINT) --> 2:TINYINT

At present, some parts of the code follow these rules, but the constant folding rule does not. As a result, PlannerTest fails for subtle changes in unrelated areas. Some PlannerTest results, such as the Kudu test mentioned in IMPALA-4213 do observe the rules. But, some constant folding test do not observe the rules.

Expected that the rules would be followed on all code paths to ensure consistent results.