Handling overflows/underflows for integer numbers in ProtoJSON

[pananton13]

Hello! Looks like official ProtoJSON documentation is lying or may be I'm misunderstanding it (from here https://protobuf.dev/programming-guides/json/#out-of-range-numeric-values):

Out of range numeric values
When parsing a numeric value, if the number that is is parsed from the wire doesn’t fit in the corresponding type, the parser should coerce the value to the appropriate type. This has the same behavior as a simple cast in C++ or Java (for example, if a number larger than 2^32 is read as for an int32 field, it will be truncated to 32 bits).

However, looking into the code it looks like json parser will signal error instead of silently casting to a smaller-sized int https://github.com/protocolbuffers/protobuf/blob/main/src/google/protobuf/json/internal/parser.cc#L240 .

Can u please comment this?

[shaod2]

@esrauchg can you check if it's a documentation update or potentially a fix to json parser?

[esrauchg]

I double checked and it does look like our main implementations do appear to consistently reject out of range numbers (eg -1 for uint32), so this looks like a documentation bug not a 'real' bug to me.

[pananton13]

Yeah, I was surprised when read in docs that ProroJSON should silently cast overflowing values which looks very bug prone. Glad that actual implementation does not behave like this.

[esrauchg]

Binary wire format does (unfortunately from my point of view) have the spec behavior of silently discarding any extra bits on ints that you receive over the wire, which must be the origin of this paragraph.

Before I checked the actual implementations to confirm they reject, I wouldn't have been surprised either way if it had attempted to follow the binary wire format precedent or be 'better'.

[pananton13]

Hello again! Looks like I've found one more issue with ProtoJSON. Official spec tells that:

Exponent notation (such as 1e2) is accepted in both quoted and unquoted forms.

However, looks like it is not true, because I observe the following error when trying to use exponent notation (like '10e2' ) in a quoted form for int64 field:
"invalid JSON in proto_json.messages.Int64Message @ field1: int64, near 1:11 (offset 10): non-number characters in quoted number"

My protobuf version is 22.5.

[pananton13]

UPD.
Ok, i've found ParseFloatStringAsInt was added in newer versions of protobuf and so it is fixed)