[Web-SIG] some questions about the write callable

[Phillip J. Eby]

At 10:06 PM 10/15/2007 +0200, Manlio Perillo wrote:
>Hi.
>
>The only feature that remains to implement for nginx mod_wsgi is the
>write callable.
>
>The WSGI spec says:
>"""In other words, before write() returns, it must guarantee that the
>passed-in string was either completely sent to the client, or that it is
>buffered for transmission while the application proceeds onward."""
>
>
>With Nginx it can happen that the passed-in string cannot be completely
>sent to the client, since the socket can returns an EAGAIN.

In which case, your write() implementation will need to loop until 
all the data hits the OS-level buffers.


>In this case Nginx will buffer the data and it will send the buffer to
>the client when the socket is ready.

Note that the two choices are:

1. data is completely sent to the client
2. data is held in a buffer *such that transmission will continue 
while the app runs*

Buffering the data but not sending it while the application continues 
executing, is not a conformant option.


>I don't understand why WSGI explicitly says '*must not* delay', instead
>of a 'should not delay'.

Because the only reason for having write() or iteration blocks (vs 
sending a single giant string) is to support interleaving the client 
communication and some other computation, communication, or I/O.

Delay would negate the point of having the ability to stream in the 
first place.


>As far as I can understand, WSGI does not explicitly forbids an
>application to call the write callable from a separate thread.
>This means that, in theory, this is allowed.

In theory, yes.  In practice, we intended to document some 
thread-affinity restrictions, and I do not believe that anybody is 
trying to call write() from another thread.


>If a new WSGI 1.1 spec is going to be released, I hope that it will be
>more friendly with asynchronous servers without threads support.

Well, I hope that the *documentation* will be more friendly for 
implementing gateways for such servers.  It's doubtful that the 
actual execution model would change much.