Subject: Re: [boost] [asio-users] [http] Formal review of Boost.Http
From: VinÃcius dos Santos Oliveira (vini.ipsmaker_at_[hidden])
Date: 2015-08-08 05:17:00
2015-08-08 0:10 GMT-03:00 Lee Clagett <forum_at_[hidden]>:
> FWIW it should be possible to do client and server side pipelining with
> this design right now. The reads and writes of the HTTP Socket concept
> should be (and currently are by `basic_socket`) handled independently.
> Since the notification for a write completion is based on the lower layer
> write completion, and NOT whether a HTTP response was received, multiple
> (pipelined) requests should be possible by a client. Similarly, this design
> should allow servers to handle pipelined requests in parrallel. Providing a
> framework for client or server pipelining will require more work, and
> should probably not be a part of the Socket concept. In fact -
The current design **does** support pipelining. Some HTTP clients disable
pipelining because some buggy servers will get confuse. This won't happen
with Boost.Http (there are tests to ensure pipelning is working).
What the current design does **not** support is handling pipelined requests
concurrently. When you starting read a request, the write_state will
change to finished until the fully request is received, so you can't get
multiple request while you don't issue the replies. This behaviour is
explained in the ServerSocket concept.
There are just too many ways to allow HTTP concurrent pipeline that would
be transparent for the user, but all of them are heavier. If I was to allow
handling multiple concurrent HTTP pipelined requests, I'd expose user
cooperation, so the abstraction doesn't become so heavy. The user would
need to be aware if the reply for some request can already be delivered or
not. Of course this means the socket will need to remember order and attach
some id to the request messages. It can be problematic in alternative
backends because the id could be of a different type and makes it difficult
to use the same message type with different HTTP backends. What do you
think? The design would just be more complex for not much gain. HTTP/2.0
allow real multiplexing and doesn't show this problem.
I think the ServerSocket concept should be removed entirely. The
> `http::basic_socket<...>` models both concepts, so theres lots of server
> specific code in it, which seems confusing to me (is this really a "basic"
basic_socket uses basic_ prefix just like basic_string.
I think standalone functions for typical client and server operations can
> be implemented as "composed" operations similar to `asio::async_write`. The
> documentation will have to be explicit on what the composed function is
> doing so that users know whether the Socket concept has outstanding reads
> or writes (and therefore cannot do certain operations until the handler is
> invoked). In fact, if there is a server operation that _cannot_ be done
> with the Socket concept, then the concept probably needs to be re-designed.
About "users know whether..." seems like a bad idea if I want to deliver
multiple backends. Some implementation details should just be abstracted
I don't understand what you mean by "if there is a server operation that
cannot be done with the Socket concept [...]". Boost.Http has two concepts,
ServerSocket and Socket. Socket concept will be useful to client-side HTTP
and that's the reason why there are two concepts.
 "The ServerSocket object MUST prevent the user from issuing new replies
-- VinÃcius dos Santos Oliveira https://about.me/vinipsmaker