Show HN: WebRPC – a simple alternative to REST and SOAP | Hacker News Reader

Show HN: WebRPC – a simple alternative to REST and SOAP

github.com

19 pointsgk_brown13 comments

While this may be simpler than REST and cleaner than SOAP, I find REST to be far more elegant.

>An HTTP 200 is returned on successful completion, and HTTP 500 is returned in the case of an error (i.e. an exception). Note that exceptions are intended to represent unexpected failures, not application-specific errors. No other HTTP status codes are supported.

Now you are just using HTTP as your transport layer, you can very well make it customised to your particular needs rather than defining a spec. This might result in easier client code, but why not use Thrift if that is what you need.

I haven't used Thrift myself, but I just took a quick look at the docs. My initial reaction is that Thrift seems to be a bit heavier. WebRPC doesn't require any intermediate interface definitions and doesn't create language-specific client proxies. Method invocations are all dynamic (although you could easily wrap the invocation API in a statically typed adapter if you wanted to).

WebRPC isn't meant to be a replacement for REST - just an alternative. Sometimes it's simply more convenient to think in terms of "methods" than "resources".

IDLs are awesomely useful and you can build all sorts of things on that abstraction, especially if you have a parser that produces a standard AST (I wrote a PegJS parser that takes the Thrift IDLs and creates a plain old javascript object AST that can be used for things like: (1) writing dynamic serializers/deserializers for binary data (uber/thriftrw), (2) validators (not yet written, but will both lint as well as check for backwards compatibility when IDLs are updated); (3) bash autocompletion for an RPC api via a curl-like command (soon to land in uber/tcurl))

At Uber, we started with Apache Thrift, but found many things about it limited and broken with dubious code quality for some languages. Furthermore we had more ambitious ideas like application layer sharding (uber/ringpop) and service discovery and routing (uber/hyperbahn).

We're still using the thrift protocols (uber/thriftrw) with our rpc library (uber/tchannel), but could have built on top of another IDL format like protobufs (v3).

For those curious about the RPC interface, just check out: https://tchannel.readthedocs.org/en/latest/

The interface supports all sorts of things

This is all over TCP right now and designed for performance and service to service communication, but there is nothing preventing the addition of HTTP (via XHR or WebSockets) as a transport for allowing web apps to speak with tchannel services. I contributed to the Apache Thrift implementation that allows this (https://github.com/apache/thrift/blob/master/lib/nodejs/lib/...). The implementation for tchannel would be similar, but rely on XHR/WS in the browser and the `http` module instead of the `net` module in NodeJS on the server. It's trivial to write a proxy frontend that relays TJSONProtocol on the frontend to a Thrift binary protocol like TBinaryProtocol or TCompactProtocol.

https://github.com/uber/hyperbahn https://github.com/uber/tchannel https://github.com/uber/ringpop-node https://github.com/uber/ringpop-go https://github.com/uber/idl https://github.com/uber/tcurl https://github.com/uber/tcap

etc.

That sounds cool. WebRPC isn't meant to be an "everything to everyone" solution. It's main advantage is simplicity and low overhead. It allows you to quickly and easily build a service tier that is consumable by multiple clients. But if your requirements scale beyond what WebRPC provides, Thrift sounds like a great option.

Thrift is such a broken mess I willingly opted instead to fork and update a Java project that exposes xml over http.

Thrift must have been written by people who though SOAP was too easy to use and that developer lives should be made harder.

Completely agree that the thrift stuff is a mess, which is why we only used the one interesting part, the IDL, for our system. These days, proto3 would be viable, but wasn't feature rich enough at the time for our needs when we were evaluating options.

So how does this compare to XML-RPC?

I see that it is using JSON instead of XML.

> Support currently exists for implementing web RPC services in Java, and consuming services in Java, Objective-C/Swift, or JavaScript.

While I may personally feel that JSON is a better format than XML, there are a implementations of XML-RPC for almost all languages and platforms which is a huge advantage.

WebRPC is a lot lighter than XML-RPC (or JSON-RPC). It's really just a thin layer on top of HTTP. Arguments are passed in the query string (or as form parameters, if POSTing), and the response value is returned directly as JSON. There's no body or envelope to parse. This makes WebRPC services really easy to consume by any client that supports HTTP and JSON (which is pretty much most platforms). WebRPC services are also really easy to test, since you can execute a method and get the results back simply by typing a URL in your browser.

We should not use HTTP for such things. It is against HTTPs intentions. Just use any other protocol.

If you are implementing a REST service, then I agree with you 100%. WebRPC is simply a different interpretation. In WebRPC, GET means "get the result of executing this method". POST basically means the same thing, but the arguments are passed in the message body (just like an HTML form).

The client libraries all use POST internally - GET is primarily useful for testing and debugging (so you can easily execute a method in the browser).

If you're transferring anything other than hypertext over http then you're a hypocrite.

Does it open a socket for every rpc?

Not sure. The client libraries don't work with sockets directly. The Java client uses HTTPURLConnection, the Objective-C/Swift client uses NSURLSession, and the JavaScript client uses XMLHttpRequest. You'd have to check the documentation for each of those to see what they might be doing under the hood, or what options you might have for configuring them. Same goes for the server, which just sits on top of a servlet container. You'd have to see the container docs to see what your options are.

Reply on news.ycombinator.com