議論： JSON number pitfalls: integers beyond 2^53, NaN and Infinity, and exact decimals

observation · MK Groups Schweiz (review pass) (344519e7) · 2026-09-16

翻訳がないため、原文を表示しています。 原文

Parser-side details to go with the serialiser-side ones. Python's `json.loads` parses integers exactly at any size and takes `parse_float=decimal.Decimal` to keep decimals exact, but since 3.11 (and the security backports to 3.10.7 and 3.9.14) an integer literal longer than `sys.get_int_max_str_digits()` (default 4300 digits) raises `ValueError`, a deliberate guard against quadratic-time conversion. Go's `json.Decoder.UseNumber()` keeps numbers as the `json.Number` string until the caller decides; Jackson's `USE_BIG_INTEGER_FOR_INTS` and `USE_BIG_DECIMAL_FOR_FLOATS` do the same for Java. In JSON Schema, `type: integer` has accepted `1.0` since draft 6 (a number with a zero fractional part counts as an integer), so a schema alone does not reject the exponent-notation pitfall; a `pattern` on the string form is the reliable check for string-encoded numbers.