========================================
 Design proposal for mod:`IPython.core`
========================================

Currently mod:`IPython.core` is not well suited for use in GUI applications.
The purpose of this document is to describe a design that will resolve this
limitation.

Process and thread model
========================

The design described here is based on a two process model. These two processes
are:

1. The IPython engine/kernel. This process contains the user's namespace and 
   is responsible for executing user code. If user code uses
   :mod:`enthought.traits` or uses a GUI toolkit to perform plotting, the GUI
   event loop will run in this process.

2. The GUI application. The user facing GUI application will run in a second
   process that communicates directly with the IPython engine using
   asynchronous messaging. The GUI application will not execute any user code.
   The canonical example of a GUI application that talks to the IPython 
   engine, would be a GUI based IPython terminal. However, the GUI application 
   could provide a more sophisticated interface such as a notebook.

We now describe the threading model of the IPython engine. Two threads will be
used to implement the IPython engine: a main thread that executes user code
and a networking thread that communicates with the outside world. This
specific design is required by a number of different factors.

First, The IPython engine must run the GUI event loop if the user wants to
perform interactive plotting. Because of the design of most GUIs, this means
that the user code (which will make GUI calls) must live in the main thread.

Second, networking code in the engine (Twisted or otherwise) must be able to
communicate with the outside world while user code runs. An example would be
if user code does the following::

    import time
    for i in range(10):
        print i
        time.sleep(2)

We would like to result of each ``print i`` to be seen by the GUI application
before the entire code block completes. We call this asynchronous printing.
For this to be possible, the networking code has to be able to be able to
communicate the current value of ``sys.stdout`` to the GUI application while
user code is run. Another example is using :mod:`IPython.kernel.client` in
user code to perform a parallel computation by talking to an IPython
controller and a set of engines (these engines are separate from the one we
are discussing here). This module requires the Twisted event loop to be run in
a different thread than user code.

For the GUI application, threads are optional. However, the GUI application
does need to be able to perform network communications asynchronously (without
blocking the GUI itself). With this in mind, there are two options:

* Use Twisted (or another non-blocking socket library) in the same thread as
  the GUI event loop.

* Don't use Twisted, but instead run networking code in the GUI application
  using blocking sockets in threads. This would require the usage of polling
  and queues to manage the networking in the GUI application.

Thus, for the GUI application, there is a choice between non-blocking sockets
(Twisted) or threads.

Asynchronous messaging
======================

The GUI application will use asynchronous message queues to communicate with
the networking thread of the engine. Because this communication will typically
happen over localhost, a simple, one way, network protocol like XML-RPC or
JSON-RPC can be used to implement this messaging. These options will also make
it easy to implement the required networking in the GUI application using the
standard library. In applications where secure communications are required,
Twisted and Foolscap will probably be the best way to go for now, but HTTP is
also an option.

There is some flexibility as to where the message queues are located. One
option is that we could create a third process (like the IPython controller)
that only manages the message queues. This is attractive, but does require
an additional process.

Using this communication channel, the GUI application and kernel/engine will
be able to send messages back and forth. For the most part, these messages
will have a request/reply form, but it will be possible for the kernel/engine
to send multiple replies for a single request.

The GUI application will use these messages to control the engine/kernel.
Examples of the types of things that will be possible are:

* Pass code (as a string) to be executed by the engine in the user's namespace
  as a string.

* Get the current value of stdout and stderr.

* Get the ``repr`` of an object returned (Out []:).

* Pass a string to the engine to be completed when the GUI application
  receives a tab completion event.

* Get a list of all variable names in the user's namespace.

The in memory format of a message should be a Python dictionary, as this
will be easy to serialize using virtually any network protocol. The 
message dict should only contain basic types, such as strings, floats,
ints, lists, tuples and other dicts.

Each message will have a unique id and will probably be determined by the
messaging system and returned when something is queued in the message
system. This unique id will be used to pair replies with requests.

Each message should have a header of key value pairs that can be introspected
by the message system and a body, or payload, that is opaque. The queues
themselves will be purpose agnostic, so the purpose of the message will have
to be encoded in the message itself. While we are getting started, we
probably don't need to distinguish between the header and body.

Here are some examples::

    m1 = dict(
        method='execute',
        id=24,  # added by the message system
        parent=None # not a reply,
        source_code='a=my_func()'
    )

This single message could generate a number of reply messages::
    
    m2 = dict(
        method='stdout'
        id=25,  # my id, added by the message system
        parent_id=24,  # The message id of the request
        value='This was printed by my_func()'
    )

    m3 = dict(
        method='stdout'
        id=26,  # my id, added by the message system
        parent_id=24,  # The message id of the request
        value='This too was printed by my_func() at a later time.'
    )

    m4 = dict(
        method='execute_finished',
        id=27,
        parent_id=24
        # not sure what else should come back with this message,
        # but we will need a way for the GUI app to tell that an execute
        # is done.
    )

We should probably use flags for the method and other purposes:

EXECUTE='0'
EXECUTE_REPLY='1'

This will keep out network traffic down and enable us to easily change the
actual value that is sent.

Engine details
==============

As discussed above, the engine will consist of two threads: a main thread and
a networking thread. These two threads will communicate using a pair of
queues: one for data and requests passing to the main thread (the main
thread's "input queue") and another for data and requests passing out of the
main thread (the main thread's "output queue"). Both threads will have an
event loop that will enqueue elements on one queue and dequeue elements on the
other queue.

The event loop of the main thread will be of a different nature depending on
if the user wants to perform interactive plotting. If they do want to perform
interactive plotting, the main threads event loop will simply be the GUI event
loop. In that case, GUI timers will be used to monitor the main threads input
queue. When elements appear on that queue, the main thread will respond
appropriately. For example, if the queue contains an element that consists of
user code to execute, the main thread will call the appropriate method of its
IPython instance. If the user does not want to perform interactive plotting,
the main thread will have a simpler event loop that will simply block on the
input queue. When something appears on that queue, the main thread will awake
and handle the request.

The event loop of the networking thread will typically be the Twisted event
loop. While it is possible to implement the engine's networking without using
Twisted, at this point, Twisted provides the best solution. Note that the GUI
application does not need to use Twisted in this case. The Twisted event loop
will contain an XML-RPC or JSON-RPC server that takes requests over the
network and handles those requests by enqueing elements on the main thread's
input queue or dequeing elements on the main thread's output queue.

Because of the asynchronous nature of the network communication, a single
input and output queue will be used to handle the interaction with the main
thread. It is also possible to use multiple queues to isolate the different
types of requests, but our feeling is that this is more complicated than it
needs to be.

One of the main issues is how stdout/stderr will be handled. Our idea is to
replace sys.stdout/sys.stderr by custom classes that will immediately write
data to the main thread's output queue when user code writes to these streams
(by doing print). Once on the main thread's output queue, the networking
thread will make the data available to the GUI application over the network.

One unavoidable limitation in this design is that if user code does a print
and then enters non-GIL-releasing extension code, the networking thread will
go silent until the GIL is again released. During this time, the networking
thread will not be able to process the GUI application's requests of the
engine. Thus, the values of stdout/stderr will be unavailable during this
time. This goes beyond stdout/stderr, however. Anytime the main thread is
holding the GIL, the networking thread will go silent and be unable to handle
requests.

GUI Application details
=======================

The GUI application will also have two threads. While this is not a strict
requirement, it probably makes sense and is a good place to start. The main
thread will be the GUI tread. The other thread will be a networking thread and
will handle the messages that are sent to and from the engine process.

Like the engine, we will use two queues to control the flow of messages
between the main thread and networking thread. One of these queues will be
used for messages sent from the GUI application to the engine. When the GUI
application needs to send a message to the engine, it will simply enque the
appropriate message on this queue. The networking thread will watch this queue
and forward messages to the engine using an appropriate network protocol.

The other queue will be used for incoming messages from the engine. The
networking thread will poll for incoming messages from the engine. When it
receives any message, it will simply put that message on this other queue. The
GUI application will periodically see if there are any messages on this queue
and if there are it will handle them.

The GUI application must be prepared to handle any incoming message at any
time. Due to a variety of reasons, the one or more reply messages associated
with a request, may appear at any time in the future and possible in different
orders. It is also possible that a reply might not appear. An example of this
would be a request for a tab completion event. If the engine is busy, it won't
be possible to fulfill the request for a while. While the tab completion
request will eventually be handled, the GUI application has to be prepared to
abandon waiting for the reply if the user moves on or a certain timeout
expires.

Prototype details
=================

With this design, it should be possible to develop a relatively complete GUI
application, while using a mock engine. This prototype should use the two
process design described above, but instead of making actual network calls,
the network thread of the GUI application should have an object that fakes the
network traffic. This mock object will consume messages off of one queue,
pause for a short while (to model network and other latencies) and then place
reply messages on the other queue.

This simple design will allow us to determine exactly what the message types
and formats should be as well as how the GUI application should interact with
the two message queues. Note, it is not required that the mock object actually
be able to execute Python code or actually complete strings in the users
namespace. All of these things can simply be faked. This will also help us to
understand what the interface needs to look like that handles the network
traffic. This will also help us to understand the design of the engine better.

The GUI application should be developed using IPython's component, application
and configuration system. It may take some work to see what the best way of
integrating these things with PyQt are.

After this stage is done, we can move onto creating a real IPython engine for
the GUI application to communicate with. This will likely be more work that
the GUI application itself, but having a working GUI application will make it
*much* easier to design and implement the engine.

We also might want to introduce a third process into the mix. Basically, this
would be a central messaging hub that both the engine and GUI application
would use to send and retrieve messages. This is not required, but it might be
a really good idea.

Also, I have some ideas on the best way to handle notebook saving and
persistence.

Refactoring of IPython.core
===========================

We need to go through IPython.core and describe what specifically needs to be
done.