On 30 June 2012 02:19, Scott Ferguson <ferg_at_caucho.com> wrote:
> On 06/29/2012 04:03 PM, Danny Coward wrote:
>
> Hi Scott, all,
>
> Thanks for the feedback, pls see below:-
>
>
>
> OK. So we can certainly add streaming to process messages.
>
> But, are you suggesting using blocking Java i/o streams, like the servlet
> api, to represent that ? Something similar or equivalent to:-
>
> WebSocketListener -> public void onMessageAsInputStream(InputStream is)
> RemoteEndpoint -> public OutputStream startSendByOutputStream() ?
>
>
> Yes. (Also with the equivalent Reader/Writer of course.)
>
> It's necessary for the core use-case of serializing JSON, XML, proto-buf,
> hessian, etc, messages, as well as for custom protocols like STOMP over
> web-socket, or rewriting ZeroMQ over websocket. For example XMPP over
> WebSocket is already a draft proposal.
>
> Streams are also better for high-performance and for very large messages.
I'm generally stream sceptical. We are doing websockets so we can
scale better and going for a blocking API just does not feel like the
right thing to do to achieve that.
However, the use-cases Scott has outlined are compelling - for
compatibility with existing libraries having a stream is necessary to
avoid buffering entire messages.
However, that is not to say that we can provide a primary websocket
API that is not stream based and supports non blocking partial message
writes - and then on top of that we can layer a blocking output
stream. Ditto for receive and input.
> And possibly exploring the type of additions that would allow non-blocking
> IO based on the traditional i/o streams as well, such as the servlet expert
> group has been looking at for servlet 3.1 ?
>
>
> Possibly. I'd personally prefer nailing down the core blocking API first
> before discussing the more complicated non-blocking APIs. I'd like to match
> the servlet model as much as possible, while avoiding javax.servlet
> dependencies.
I see it the other way. Doing non-blocking APIs is hard and harder if
you have to be compatible with a blocking API that was conceived
without consideration for non-blocking (see the mess in servlet 3.1
JSR now!).
On the otherhand, if we have a good non-blocking API, creating
blocking derivatives is normally trivial (eg like providing stream
wrappers).
> Although at least one of the APIs allows this, most APIs seem to favor a
> type of asynchronous processing same as or equivalent to:-
>
> WebSocketListener-> public void onMessage(byte[] fragment, boolean isLast)
> RemoteEndpoint-> public void send(byte[] fragment, boolean isLast)
>
> What are people's thoughts on standardizing this kind of chunking API ?
Well firstly we'd have to standardize the language - chunk is kind of
taken by HTTP and fragment is kind of associated with frame.
I'm cool with having an API that sends and receives partial messages,
but we have to be clear that those partial messages are unrelated to
(or at least decoupled from) websocket frames. Ie a single websocket
frame might be delivered in multiple calls to onMessage(byte[]
fragment, boolean isLast), or multiple frames might be delivered in a
single call. Further a call to send(byte[] fragment, boolean isLast)
may result in 1 or more frames being sent.
[ yes I realise that I'm mixing up responses to Danny and Scott in the
same message - but my mail client is confused ]
> Well, let me unpack this because there are several intertwined issues that
> should be separated:
>
> 1) The WebSocket frame/fragment is not an application-visible concept
> (excluding extensions for the moment).
+1
> 2) Even for extensions, the IETF WebSocket Multiplexing group is finding
> that frame-based extensions are a problem, because they interact in
> difficult ways. There's a suggestion on one of their threads that perhaps
> only the mux extension itself even be aware at all about frames, and all
> other extensions work on messages.
Indeed - messing with the frame headers can cause all sorts of grief.
I think most extensions should be restricted to mutating payloads
(including fragmenting them).
> You could have "buffer" as in write(buffer, offset, length, isLast) just to
> make it absolutely clear that the sending buffer has nothing to do with
> websocket frames, but...
>
> 3) async sends are a problem because they imply queuing or at least large
> buffering, and raise the question of who owns the buffer, memory allocation,
> if there are extra copies just to handle the async, etc. These aren't
> appropriate for the low level API. (A higher async messaging/queuing on top
> of the low level API might be fine, but not the lowest level.)
I think NIO.2 got async sends pretty well right. Send the passed
buffers and call me back when the send is completed. Those buffers do
not need to be copied and internally an implementation can slice off
frame by frame from a huge passed buffer and control it's own memory
footprint.
NIO.2 got async reads a bit wrong as you need to allocated buffers
when you schedule the read. But in this context we are calling back
with the message (or partial message), so that is less of a problem.
> 4) the "isLast" kind of API (assuming blocking) is functionally equivalent
> to a stream, but someone would need to write a stream wrapper if they're
> serializing xml, json, etc. Which isn't really something that the API should
> force on an application.
I agree that isLast is not a nice API and does not entirely solve the problem.
> 5) async/isLast receives are truly messy if you're deserializing xml, json,
> etc, because you either need to buffer the entire message (!) before
> starting to deserialize, or create a complicated threaded producer/consumer
> model to create a stream wrapper. Again, this is fairly brutal to require of
> an application.
+1 Such applications should not attempt async serialization and
should have a stream abstraction at their disposal.
But that does not mean that we cannot expose an async API below the
stream abstraction.
> Although the base layer shouldn't be an async/chunked API, I'm all for an
> async/messaging layer written on top of the simple blocking streaming layer,
> if it's a general solution useful for a large class of applications.
I'm the other way around. Let's get async sorted and then blocking is easy.
Async will be hard and ugly no matter what, but less so if we don't
have pre-existing blocking API to work around.
--
Greg Wilkins <gregw_at_intalio.com>
www.webtide.com
Developer advice, services and support
from the Jetty & CometD experts.