On Tue, 4 Sep 2007 12:45:17 -0600, Brian Granger <ellisonbg.net@gmail.com> wrote:
This is a usage case that we run into regularly. We have a lot of PB based servers, but sometimes people want to use a client to them from an interactive Python or IPython prompt. At that point you have to run the reactor in a different thread.
Not true, in fact. See twisted/conch/stdio.py, for example. I don't know what the Twisted/IPython integration code ended up looking like (or if it was ever taken to a satisfactory "completion" point) but I don't think it is necessary to have multiple threads for that use case, either.
I have, but this doesn't solve the problem as I recall.
There are definitely some problems it doesn't solve. Maybe I'm not getting the problem you have in mind.
First, these users want to be able to use PB clients from vanilla python interpreters and IPython out of the box - not from within some other process doing the tricks that are in conch/stdio.py.
stdio.py doesn't do anything with multiple processes (as I think you see, based on what you said a couple sentences later, but just to be clear). Another point to clarify - there is a difference between having a blocking user interface and a blocking API. In fact, they're different things entirely - they don't relate to each other in any significant way. Keeping this in mind, there's no reason either the standard CPython REPL or IPython couldn't provide a blocking user interface to an asynchronous API, without resorting to multiple processes or threads. The implementation is asynchronous, the user interface blocks. Of course, actually implementing this for the CPython REPL poses some challenges. However, an API does exist which would allow it to poll Twisted at regular intervals in order to allow Twisted to take care of any event processing it needed to take care of. Whether this would yield a high- quality implementation or not, I'm not sure, but something should be possible. The reason I wrote stdio.py instead of doing this is that I'd rather have a REPL implemented in Python without having to jump through libreadline hoops. Plus I didn't know about the CPython polling feature at the time. ;) IPython could do this even more easily, since I think some IPython developers are actively interested in the feature, and there might be a chance of there being some changes made to IPython in order to facilitate this. I hope I have understood properly the use-case you have in mind and that the above makes some sense.
Am I not correct that threads are needed in this case? Second, what this module does is actually make the terminal asynchronous, which is exactly the opposite of what these users want.
stdio.py keeps everything asynchronous and throws Deferreds right in the user's face. But it does this because I'm a crazy guy and I love Deferreds. It's equally possible to implement something like stdio.py, without using threads or multiple processes, where the user experiences the interface blocking until a result is available each time he would have otherwise seen a Deferred.
They *want* a synchronous terminal and a blocking PB clients. I hope I am not coming across as angry about this - I'm not.
Not at all. :)
To me the lesson is that as long as everything is really asynchronous anything is possible within the constraints of this asynchronous universe. It's the multiple universe things that makes life complicated (asyn + syn).
Does the notion of dividing things up so that user interface concerns are dealt with separately from API concerns clear this up at all? Do you still think there's something here which can't be solved without threads? (To be sure, there is something - if you have a blocking API and you want to use it inside an event loop, then you may need to resort to threads, but I don't *think* that's the case you're interested in here - correct me if I'm wrong.)
But, I have never seen a method of putting a blocking API on top of such a client that was _really_ thread safe. I have seen a few approaches that seem to work, but that are clearly non-thread safe. I would love to see a robust solution for this problem though - one that could be fully trusted.
Why isn't a solution based on Twisted's thread-safe event-posting API (reactor.callFromThread) and the thread-safe event-posting API of whatever other thing is being integrated with "really" thread-safe? Sure, you can write programs which will deadlock as a result of application-level bugs where one thread is blocked on a request and the other thread needs some information from it before the request can be satisfied, but this is inherent to wanting to drive an asynchronous API synchronously. Aside from it, I don't think there are any problems with what I described above.
It probably is, but I have never seen an implementation that actually does this. Do you know of one? The ones that I have seen - even those which use callFromThread have not filled me with thoughts of robustness, stability and thread safety. I should look at this again though.
Ah, well, I don't have any that I can point too and claim are robust and bug free (or even bug lite). But hey, the *idea* is simple enough. >:)
Also, I should mention that in IPython this is not a show stopper for us, but it does mean that we currently don't use PB in all of the places we might otherwise.
Is this because you want to have blocking APIs in order to implement features in IPython or is it because you want to present an interface to users which behaves in a blocking manner? (Sorry if this is beginning to sound repetitive) Jean-Paul