-
-
Notifications
You must be signed in to change notification settings - Fork 818
Description
Currently JsonParser
has method getNumberType()
, with semantics that are loose for many textual formats.
Basically formats like JSON do not have similar types as programming languages: so while we have separate NumberType
entries representing float
(32-bit binary FP), double
(64-bit "double"precision binary FP) and BigDecimal
(unlimited-precision, decimal FP), there is no efficient mechanism to actually produce correct NumberType
for floating-point values.
Because of this, basically all FP values claim to be of NumberType.DOUBLE
for such formats.
This can be problematic if values are converted first to double
, then to BigDecimal
, since former cannot accurately represent all decimal numbers.
However, binary formats often have specific storage representations that can provide this type information.
The problem comes when converting to Java types: both java.lang.Number
(or generally java.lang.Object
) and JsonNode
.
In this case we would ideally use either:
- Exact type if known (binary formats) OR
- Well-known type --
Double
ORBigDecimal
, based on configuration - In some edge cases (not-a-number aka
NaN
),Double
as that can represent such values.
(further complicating things, we also have secondary means of producing NaN
values: value overflow for double
(and theoretically, but not practically, float
) which can produce +INFINITY
)
Given all above confusion, I think we need a new method like getNumberTypeFP()
-- with matching enum NumberTypeFP
(to be able to express value UNKNOWN
, no need for integer types).
That will allow deserializers to know if "true number type", and base logic on that, specifically avoiding conversions in case of Binary formats and allowing their use for Textual formats (or in general formats without explicit type information for FP numbers).
EDIT: originally thought we'd need getNumberTypeExplicit()
, but since the need is specifically for floating-point numbers, let's call it getNumberTypeFP()
instead; no need for non-FP types. And can indicate semantics are for strict/explicit type.