[CALCITE-6074] The size of REAL, DOUBLE, and FLOAT is not consistent

• Type: Bug
• Status: Closed
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: 1.35.0
• Fix Version/s: 1.37.0
• Component/s: core
• Labels: pull-request-available
• Assignee: Mihai Budiu
• Reporter: Mihai Budiu
• Created: 26/Oct/23 01:04
• Updated: 06/May/24 22:19
• Resolved: 09/Jan/24 00:26

Description

This stems from the review of CALCITE-6052

Which one is 8 bytes and which one is 4 bytes?

The intent seems to be that DOUBLE and FLOAT are synonyms, both using 8 bytes, (which is very weird for Java users), and REAL is 4 bytes.

But an audit of the code shows that:

In AggregateNode.maxMinClass:

    case FLOAT:
      return max ? MaxFloat.class : MinFloat.class;
    case DOUBLE:
    case REAL:
      return max ? MaxDouble.class : MinDouble.class;

In VisitorDataContext:

     case DOUBLE:
        return Pair.of(index, rexLiteral.getValueAs(Double.class));
      case REAL:
        return Pair.of(index, rexLiteral.getValueAs(Float.class));

(no case for FLOAT)

In RelMdSize:

   case FLOAT:
    case REAL:
   ....
      return 4d;

in RelDataTypeFactoryImpl:

    case REAL:
      return createSqlType(SqlTypeName.DECIMAL, 14, 7);
    case FLOAT:
      return createSqlType(SqlTypeName.DECIMAL, 14, 7);
    case DOUBLE:
      // the default max precision is 19, so this is actually DECIMAL(19, 15)
      // but derived system can override the max precision/scale.
      return createSqlType(SqlTypeName.DECIMAL, 30, 15);

The reference.md itself seems to be wrong:

| REAL, FLOAT | 4 byte floating point     | 6 decimal digits precision.  
| DOUBLE      | 8 byte floating point     | 15 decimal digits precision.

and there are many more I haven't even checked!

Issue Links

• links to: GitHub Pull Request #3492