[Web-SIG] Emulating req.write() in WSGI
Graham Dumpleton
graham.dumpleton at gmail.com
Thu Jul 1 01:19:33 CEST 2010
On 30 June 2010 22:55, Aaron Fransen <aaron.fransen at gmail.com> wrote:
>
>
> On Wed, Jun 30, 2010 at 6:26 AM, Graham Dumpleton
> <graham.dumpleton at gmail.com> wrote:
>>
>> On 30 June 2010 21:35, Aaron Fransen <aaron.fransen at gmail.com> wrote:
>> >
>> >
>> > On Tue, Jun 29, 2010 at 6:17 PM, Graham Dumpleton
>> > <graham.dumpleton at gmail.com> wrote:
>> >>
>> >> On 30 June 2010 02:14, Aaron Fransen <aaron.fransen at gmail.com> wrote:
>> >> > Couple more things I've been able to discern.
>> >> >
>> >> > The first happened after I "fixed" the html code. Originally under
>> >> > mod_python, I guess I was cheating more than a little bit by sending
>> >> > <html></html> code blocks twice, once for the incremental notices,
>> >> > once
>> >> > for
>> >> > the final content. Once I changed the code to send a single properly
>> >> > parsed
>> >> > block, the entire document showed up as expected, however it still
>> >> > did
>> >> > not
>> >> > send any part of the html incrementally.
>> >> >
>> >> > Watching the line with Wireshark, all of the data was transmitted at
>> >> > the
>> >> > same time, so nothing was sent to the browser incrementally.
>> >> >
>> >> > (This is using the write() functionality, I haven't tried watching
>> >> > the
>> >> > line
>> >> > with yield yet.)
>> >>
>> >> Use a variation of WSGI middleware wrapper in:
>> >>
>> >>
>> >>
>> >> http://code.google.com/p/modwsgi/wiki/DebuggingTechniques#Tracking_Request_and_Response
>> >>
>> >> using it to 'print' returned data to Apache log and then tail Apache
>> >> error log to see when that data is output. Alternatively, change the
>> >> code there to output a time stamp against each chunk of data written
>> >> to the file recording the response content.
>> >>
>> >> This will show what data is returned by WSGI application, before
>> >> mod_wsgi truncates anything greater than content length specified,
>> >> plus also show whether it is your WSGI application which is delaying
>> >> output somehow, or whether Apache output filters are doing it.
>> >>
>> >> Graham
>> >
>> > I've actually tried a variation on this already using a built-in logging
>> > facility in the application that writes date/time values to an external
>> > log
>> > file with comments, and in the case of testing wsgi I actually included
>> > some
>> > time.sleep() statements to force a delay in the application.
>> >
>> > To give you an idea of the flow, here's essentially what's going on:
>> >
>> > def application(environ,start_response):
>> > mydict = {}
>> > mydict['environ']=environ
>> > mydict['startresponse'] = start_response
>> > # run program in another .py file that has been imported
>> > RunTest(mydict)
>> >
>> > Then in the other module you would have something like:
>> >
>> > def RunTest(mydict):
>> > status = '200 OK'
>> > response_headers = [('Content-type','text/html')]
>> > writeobj = detail['startresponse'](status,response_headers)
>> > writeobj('<html><body>Fetching sales for 2009...')
>> > time.sleep(2)
>> > writeobj('<br>Fetching sales for 2010...')
>> >
>> > ...then finally...
>> >
>> > writeobj('5000 results returned.</body></html>')
>> > return
>> >
>> > This is obviously a truncated (and fake) example, but it gives you an
>> > idea
>> > of the flow.
>>
>> Now go try the following two examples as illustrated instead.
>>
>> In both cases, do not use a web browser, instead telnet to the port of
>> the web server and enter HTTP GET directly. If you are not using
>> VirtualHost, use something like:
>>
>> telnet localhost 80
>> GET /stream-yield.wsgi HTTP/1.0
>>
>> If using a VirtualHost, use something like:
>>
>> telnet localhost 80
>> GET /stream-yield.wsgi HTTP/1.1
>> Host: tests.example.com
>>
>> Ensure additional blank line entered to indicate end of headers.
>>
>> First example uses yield.
>>
>> # stream-yield.wsgi
>>
>> import time
>>
>> def application(environ, start_response):
>> status = '200 OK'
>>
>> response_headers = [('Content-type', 'text/plain')]
>> start_response(status, response_headers)
>>
>> for i in range(10):
>> yield '%d\n' % i
>> time.sleep(1)
>>
>> Second example uses write:
>>
>> # stream-write.wsgi
>>
>> import time
>>
>> def application(environ, start_response):
>> status = '200 OK'
>>
>> response_headers = [('Content-type', 'text/plain')]
>> write = start_response(status, response_headers)
>>
>> for i in range(10):
>> write('%d\n' % i)
>> time.sleep(1)
>>
>> return []
>>
>> For me, using stock standard operating system supplied Apache on Mac
>> OS X, I see a line returned every second.
>>
>> If I use Safari as a web browser, in both cases the browser only shows
>> the response after all data has been written and the socket connection
>> closed. If I use Firefox however, they display as data comes in.
>>
>> This delay in display is thus possibly just the behaviour of a
>> specific browser delaying the display until the socket is closed.
>>
>> The example for multipart/x-mixed-replace which others mention is:
>>
>> import time
>>
>> def application(environ, start_response):
>> status = '200 OK'
>>
>> response_headers = [('Content-Type', 'multipart/x-mixed-replace;
>> boundary=xstringx')]
>> start_response(status, response_headers)
>>
>> yield '--xstrinx\n'
>>
>> for i in range(10):
>>
>> yield 'Content-type: text/plain\n'
>> yield '\n'
>> yield '%d\n' % i
>> yield '--xstringx\n'
>>
>> time.sleep(1)
>>
>> With telnet you will see the various sections, but with Safari again
>> only shows at end, although you will find that it only shows the data
>> line, ie., the number and not all the other stuff. So, understands
>> multipart format but doesn't support x-mixed-replace. It was always
>> the case that only certain browsers supported that mime type. In the
>> case of Firefox, it doesn't seem to like it at all and seems to give
>> up and not display anything, not even replacing the previously
>> displayed page contents.
>>
>> What this means is that you cant rely on browsers to handle multipart
>> mixed replace alone. If you were really going to use that format, you
>> really want to use JavaScript and AJAX stuff to process it. The same
>> applies for progressive display of plain text content when streamed
>> over time.
>>
>> In summary, you really want to be using some JavaScript/AJAX stuff on
>> browser side to get uniform behaviour on all the browsers.
>>
>> Graham
>
> I can see that this could potentially get very ugly very quickly.
>
> Using stock Apache on the current Ubuntu server, using yield produced a
> response error
What error? If you aren't going to debug it enough to even work out
what the error is in the browser or Apache error logs and post it here
for comment so can say what may be wrong on your system, then we cant
exactly help you much can we.
> and using write() (over the telnet interface) returned the 0
> only and disconnected. Similar behavior in Firefox.
All the scripts I provided you are conforming WSGI applications and
work on mod_wsgi. If you are having issues, then it is likely going to
be the way your Apache/Python is setup or how you configured mod_wsgi
to host the scripts. Again, because you are providing no details about
how you configured mod_wsgi we cant help you work out what is wrong
with your system.
> How odd that nobody's come up with a simple streaming/update schema (at
> least to my mind).
For response content they have and it can be made to work. Just
because you cant get it working or don't understand what we are saying
about the need to use a JavaScript/AJAX type client (eg. comet style)
to make use of it as opposed to trying to rely on browser
functionality that doesn't exist doesn't change that. Request content
streaming is a different matter as I will explain below but you
haven't even mentioned that as yet that I can see.
> It would have been nice to be able to provide some kind of in-stream
> feedback for long running jobs, but it looks like I'm going to have to
> abandon that approach. The only issue with either of the other solutions is
> that each subsequent request depends on data provided by the prior, so the
> amount of traffic going back & forth could potentially become a problem.
>
> Alternatively I could simply create a session database that saves the
> required objects then each subsequent request simply fetches the required
> one from the table and...
>
> Well, you can see why streaming seemed like such a simple solution! Back to
> the drawing board, as it were.
I'll try one last time to try and summarise a few issues for you,
although based on your attitude so far, I don't think it will change
your opinion or help your understanding.
1. Streaming of responses from a WSGI application works fine using
either yield or write(). If it doesn't work for a specific WSGI
hosting mechanism then that implementation may not be conforming to
WSGI requirements. Specifically, between a yield and/or write() it is
required that an implicit flush is performed. This should ensure that
the data is written to the HTTP client connection and/or ensure that
the return of such data to the client occurs in parallel to further
actions occurring in that request.
2. A WSGI middleware that caches response data can stuff this up. One
cant outright prohibit a WSGI middleware holding on to response data,
albeit that for each yield or write() technically it is supposed to
still pass on at least an empty string down the chain so as to allow
control to get back to the underlying WSGI implementation, which may
uses such windows to swap what request context it is operating on so
as to allow a measure of concurrency in situation where threads may
not be getting used.
3. Where a WSGI adapter on top of an existing web server is used, eg.
various options that exist with Apache and nginx, then an output
filter configured into the web server may also stuff this up. For
example, an output filter that compresses response data may buffer up
response data into large blocks before compressing them and returning
them.
4. Although response content can be streamed subject to above caveats,
streaming of request content is a totally different matter. First off,
WSGI requires that the request content have a Content-Length
specified. Thus technically a HTTP client cant leave out
Content-Length and instead used chunked request content. Further, the
way in which many web servers and WSGI servers are implemented would
prohibit streaming of request content anyway. This is because many
implementations, especially where proxying occurs, eg. cgi, fastcgi,
scgi, ajp, uwsgi, mod_proxy (??), and mod_wsgi daemon mode, expect
that the whole request content can be read in and written across the
proxy connection before any attempt is made to start reading any data
returned from the web application. The request content therefore
cannot be open ended in length because most implementations will never
switch from reading that content to expecting response from the
application. Thus it isn't possible to use WSGI as both way streaming
mechanism where some request content is written, some response content
returned and then the client sends more request content based on that
etc etc.
So what does this all mean. First up is that response content
streaming should be able to be made to work, however since request
content streaming isn't technically allowed with WSGI, if you need
that you are out of luck if you want to conform to WSGI specification.
Second, you can however with mod_wsgi embedded mode slightly step
outside of strict WSGI conformance and have request content streaming.
You are then bound to Apache/mod_wsgi, but whether you want to do that
is debatable for reasons below.
The bigger problem with both way streaming or long polling
applications which use the same HTTP request is that WSGI servers tend
to use processes and threads for concurrency. When you use this
mechanisms they will tie up a process or thread for the whole time.
Thus if you have lots of concurrent request you need huge numbers of
processes and/or threads, which just isn't usually practical because
of resource usage such as memory. For that reason, one would instead
on the server usually use special purpose web servers for these types
of applications and use HTTP directly and avoid WSGI, due to WSGI
blocking nature. Instead these servers would use an event driven
system model or other system which allows concurrency without
requiring a process or thread per application.
In short, this is what Comet and dedicated servers for that are about.
Allowing large numbers of concurrent long requests with minimal
resources. That they are dedicated systems also allows them to avoid
limitations in other high level web application interfaces such as
CGI, FASTCGI, SCGI, AJP etc which have an expectation that can read
whole request content before trying to deal with any response from a
web application that is handling the requests.
Anyway, hopefully that explains things better. You can do what you
want, you just need to select the correct tool for the job.
Graham
More information about the Web-SIG
mailing list