Thrift NiceJSON FAQ
TJSONProtocol is a JSON protocol in name only:
- 1
-
a Thrift IDL type like
struct Bar {
1: required i32 a ,
2: required string b,
}
with a=32, b="foo" produces a JSON like {"1":32,"2":"foo"}. To
wit, the keys are field-ids, not field-names. This makes the
JSON effectively unreadable and unwritable (to humans). And
completely useless for interoperating with any other JSON-capable
system (b/c none of them use such a field-naming convention).
- 2
-
The serialization of all containers (map/list/set) is different from idiomatic JSON. It include type-information, and element-counts.
- 3
-
It doesn’t support any variation in the white-spacing of the JSON.
- 4
-
It doesn’t produce correct JSON when applied to Thrift types like
map[list<i32>]string.
Thrift already knows how to produce a Thrift object that contains all
the information in the .thrift IDL file. It uses this to
communicate with "plugins". This Thrift object’s type is described in
plugin.thrift in the Thrift source-distribution. Let’s call this a
"type library" (typelib).
For serialization, Thrift can marshal to a binary potocol
(TBinaryProtocol). To produce the equivalent JSON, we merely take
the result of binary-protocol marshaling, and demarshal that using the
typelib to guide us (called "interpretive marshaling") and build a
JSON object. This is all done in C++, and for other languages, we
jsut convert that JSON into a string.
The inverse process (deserializing JSON into a Thrift object) is similar: we start with a C++ JSON object, and, using type-library information, serialize it in the binary protocol format. Finally, that binary protocol message is handled to Thrift code that deserializes to the Thrift object.
This method has intrinsic performance limitations, but it has the signal advantage that it can work with any Thrift data-types. That’s quite difficult to do with any method that didn’t use typelib information for many different reasons, the most important of which is that in all Thrift protocols, containers (maps/sets/lists) carry with them bits of meta-information (e.g. # of elements) that one would never want to have present in any JSON serialization of a container.