Schema-full formats?

[colinator]

Would it be possible (and if so how difficult) to add 'schema-full' serialization formats? For instance protobuf, flatbuffers, cap'n proto, dds, etc.

Obviously it would require providing the schema, or maybe the compiled artifacts (as in protobuf for instance) at compile-time. And I realize it doesn't really map to this library. But something that could easily translate between, say, protobuf and flex buffers, or dds and json, would be really useful.

[liuzicheng1987]

I have taken a look at protobuf, flatbuffers and cap'n proto and concluded that there are issues with each of these formats that it make it difficult. I have even been in touch with Kenton Varda, the main author of cap'n proto, and he has confirmed that I would need significant amounts of reverse-engineering to make this work.

However, I think that Apache Avro (https://avro.apache.org/docs/1.12.0/) looks very promising and I will open an issue for it.

[tsurumi-yizhou]

The protobuf format isn't so complex. See struct_pb which implements a simple system to serialize into protobuf.

[liuzicheng1987]

Hi @colinator and @tsurumi-yizhou,

so, I've got some updates on this issue. I have devised a system on how we can use the reflection API for schemaful formats and the first format I have chosen to support is Apache Avro (https://avro.apache.org), just because its C API lends itself to doing something like this (https://avro.apache.org/docs/1.11.1/api/c/).

The next one I have in mind is cap'n proto, which actually has an API specifically for this purpose (https://capnproto.org/cxx.html#dynamic-reflection).

But I will certainly take a look at the library @tsurumi-yizhou suggested and how they approach this problem. If we could also support protobuf, that would be great.

[colinator]

This is good news - I've long wanted a c++ library that can serialize both to flex and flat buffers, for example - or any other format.

The neat thing about serialization formats such as flatbuffers (or cap'n proto) for example, is that they support "0-copy". Meaning, if your data contains a large thing (such as an image tensor), then no expensive memory allocation/copy is needed (at least at the user level) in order to read it. I'm concerned that reflect-cpp cannot support 0-copy, which negates one of the big benefits of some of the formats. See my issue #207. I'd be happy to chat more about this.