[asio-users] strands and long polling

[JJ A. <jj....@gm...>]

Hi all,

We are using asio in a web application where server "push" is implemented
using long-polling. We are using io_service::run on a pool of threads (as
demonstrated in HTTP server example 3). As in the example we are associating
a strand with each inbound connection which it services.

We have recently run into an issue where when two requests arrive nearly
simultaneously it is possible that one can tie up the other and cause it to
block until the first one completes. We speculated that this may be due to
the use of strands, removed them, and the problem was no longer
reproducible. I subsequently noted that this issue was also previously
discussed on this list here:

http://comments.gmane.org/gmane.comp.lib.boost.asio.user/3945

Christopher suggests in this thread that the long-polling be moved off of
the strand using io_service.post.

While we can certainly do this, I was wondering whether it was even
necessary to use strands in the first place since the implementation only
has a single IO operation pending at a time (a series of calls to
async_read_some followed by an async_write). See:

http://www.boost.org/doc/libs/1_44_0/doc/html/boost_asio/example/http/server3/connection.cpp

It seems like we'd be OK without the strand since another aysnc_read_some
isn't posted until the previous one has completed. Perhaps the strand was
included in the example by way of illustration but isn't strictly necessary?

Of course, I may be missing the logic, purpose, and workings of strands,
threads, and io services entirely :-)  Would be curious as to what others
think the right approach is here.

Thanks,

J.J.