Support jrpcv1 clients

[roytman]

Hello, and thanks for awesome package.

I'd like to ask about a small extension.

If the server is started with the AllowV1 flag, it accepts jrpcv1 requests (requests without the jsonrpc version string), however, its responses include the version string, and as a result, v1 clients don't accept the responses.

I have an extension that allows sending responses according to corresponding requests. If AllowV1 is true and a request doesn't include the version string, its response will not include it either.

Server-initiated messages (Callbacks and Notifications) require a slightly different approach. One of the possible implementations is to store the client version in the Server struct.

What do you think about this extension?

[creachadair]

What you're describing is possible, but I'm not convinced it's a good fit for this module. I added the version tolerance settings based on the advice of https://www.jsonrpc.org/specification#compatibility, but there are several other differences between the two protocols beyond just the version marker.

Making this module support JSON-RPC 1.0 properly would need several other changes—the 1.0 standard models both ends of a connection as full peers, for example. Also, a strict 1.0 implementation requires object formats that 2.0 forbids, e.g., a successful response in 1.0 looks like {"id":1, "result":…, "error":null}, where as 2.0 does not permit both "result" and "error" to be set. I point this out not to defend strictness—but simply to say that if an existing client is already strict enough to reject responses because of the "jsonrpc":"2.0" field in the reply, it is likely to have other issues even if that is removed.

If the eventual goal is to swap an existing service from 1.0 to 2.0, it's probably better to set up separate endpoints to allow clients to be rolled over incrementally. Having said that, however, I'm interested to hear more about what you have in mind.

[roytman]

Thanks for your response.
I went again over JSON-RPC specs, and there is an additional difference between the versions: the error response format.
jrpcv1 doesn't define its structure, where jrpcv2 has a strict definition.
Even current implementation doesn't correctly parse jrpcv1 errors.

All mentioned differences do not require significant design changes, maybe you can consider adding them.

I try to reimplement jrpcv1 server, which should be able to send notifications to clients, which is not supported by the standard Golang libraries. I cannot add new endpoints to clients, because they are legacy code and out of my control.

[creachadair]

I went again over JSON-RPC specs, and there is an additional difference between the versions: the error response format.

Yes, definitely—my list of example differences wasn't comprehensive.

I try to reimplement jrpcv1 server, which should be able to send notifications to clients, which is not supported by the standard Golang libraries. I cannot add new endpoints to clients, because they are legacy code and out of my control.

I see—and from your comment above, it sounds like these legacy clients are sensitive about the format of request and response messages from the server.

Based on this discussion, I don't think 1.0 support is a good fit for this module. Having to plumb version sensitivity through the request flow is not technically that difficult, but it's a lot of little changes and I don't see a clear benefit. Indeed, perhaps even the existing AllowV1 switches might have been a mistake, since they don't go far enough to meaningfully work.

For your needs, I think you might be better off with a fork to make an entirely JSON-RPC 1.0 implementation. That would not require any extra plumbing—just drop the version markers from the wire format, and update the encoding/decoding logic to be more forgiving about errors. It wouldn't be 2.0 compliant anymore, but it sounds like for your needs that isn't important. The main caveat is that the server isn't a true "peer" in the 1.0 sense.

[roytman]

indeed meantime I work with a fork.
Thanks, I'll close this issue.