add unknown field support

[pavelstudeny]

protobuf.js version: all, up to the current 6.8.0

proto2 preserve unknown fields for future serialization. This is also planned for proto3 in the near future: https://docs.google.com/document/d/1KMRX-G91Aa-Y2FkEaHeeviLRRNblgIahbsk4wA14gRk/edit#heading=h.w8dtggryroj4

I'm working on a merge request that will add unknown fields when decoding as

_unknownFields: { <id|wireType>: }

This should be doable by add read_type_bytes() to Reader and using it instead of skipType. The _unknownField can be later serialized again or discarded by a new method discardUnknownFields().

Tests:

var proto = "message Simple_v1 {\
  optional string name  = 1;\
  optional string value = 3;\
}\
message Simple_v2 {\
  optional string name  = 1;\
  optional int32  flags = 2;\
  optional string value = 3;\
}";

var msg = { inner: [{}, {}, {}] };

tape.test("unknown fields", function (test) {
    var root = protobuf.parse(proto).root,
        Simple_v1 = root.lookup("Simple_v1");
        Simple_v2 = root.lookup("Simple_v2");

    var s2 = Simple_v2.create({ name: "v2", flags: 2, value: "dummy" });
    var s1 = Simple_v1.decode(Simple_v2.encode(s2).finish());

    var restored = Simple_v2.decode(Simple_v1.encode(s1).finish());

    test.equal(s2.name, restored.name, "assert: even known fields are missing");

    test.equal(2, restored.flags, "are preserved by default");
    test.end();
});

tape.test("discarded unknown fields", function (test) {
    var root = protobuf.parse(proto).root,
        Simple_v1 = root.lookup("Simple_v1");
    Simple_v2 = root.lookup("Simple_v2");

    var s2 = Simple_v2.create({ name: "v2", flags: 2, value: "dummy" });
    var s1 = Simple_v1.decode(Simple_v2.encode(s2).finish());

    try {
        s1.discardUnknownFields();
    }
    catch (ex) {
        test.end("discardUnknownFields() exception: " + ex);
        return;
    }

    var restored = Simple_v2.decode(Simple_v1.encode(s1).finish());

    test.equal(undefined, restored.flags, "are removed from the message");
    test.end();
});

Does it sound OK?

[dcodeIO]

I have a feeling that this will come at quite a performance hit. For example, encoders serialize fields in ascending field id order, as of the spec. Making this work with unknown fields might be quite a challenge without hurting performance significantly.

We'd probably be better off ignoring order of unknown fields completely, simply appending uninterpreted buffer slices to the final buffer again. skipType could, for example, in addition to skipping over the data, just return the respective buffer slice. These slices would then be stored in an array without any id/wiretype association required, for the encoder to append.

[pavelstudeny]

Yes, implementations for other languages also place unknown fields at the end, so protobuf.js could follow the habit. This should also remove any performance impact, I hope.

May be there should anyway be a global option to ignore the unknown fields, in addition to an explicit discard?

Does it sound OK that the discardUnknownFields method would be on the decoded data instance (s1/s2 above) or should it rather stay at the type descriptor (Simple_v1/Simple_v2)?

[dcodeIO]

May be there should anyway be a global option to ignore the unknown fields, in addition to an explicit discard?

Yeah, why not.

Does it sound OK that the discardUnknownFields method would be on the decoded data instance (s1/s2 above) or should it rather stay at the type descriptor (Simple_v1/Simple_v2)?

Instances don't have any methods on them by default (except .toJSON, which must be on the instance). But this could also be just a global helper function, as it basically just does a delete someMessage._someProperty.

[pavelstudeny]

How about #779?

I hope the CODECLIMATE_REPO_TOKEN is a problem on the test machine.

I had to remove the first variant of writeBytes in writer.js. It looked like a mere optimization, that was however never called and was failing on nodes 4-. It was also ignored by test coverage - that's why the existing tests couldn't find the problem.

[pavelstudeny]

Just wondering about a general opinion, whether the solution looks acceptable as it is or needs any changes?

[dcodeIO]

Sorry, haven't yet got around to check. Being a bit lazy here because it's not in the official implementation yet, but I appreciate your efforts!

[pavelstudeny]

Fair enough for proto3, if the upcoming specification it's going to define API calls for this feature. This, however, also fulfills the requirement for proto2, that is also handled by protobuf.js.

[niicojs]

I would need this also. Is your PR usable already?

[pavelstudeny]

It's complete and usable, although I'm not using it myself, as dcodeIO may decide to modify it or come up with a different solution.

[robin-anil]

I would love to +1 this. Currently, the parse fails with unknown fields in the byte stream. this is a hallmark feature of protobufs in other languages that allows developers to push different applications without breaking.

[chakmingli]

It's been a while. Will this change be accepted?

[pavelstudeny]

Protobuf 3 preserves unknown fields by default since release 3.5.0 from 2017-11-13:

Unknown fields are now preserved in proto3 for most of the language implementations for proto3 by default.

This should be a reason to accept pull request #779 or point out what shoul be improved.

[jcgertig]

BUMP this is very needed