Skip to main content

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] [List Home]
Re: [ease-dev] Jupyter kernels

Hi Martin,

not sure if we talk about the right launch order here. From my point of view eclipse should never publish its launch to jupyter, because eclipse is there first. Each Eclipse instance launches its own jupyter core (if necessary) and then launches jupyter kernels for each opened notebook. As we do open notebooks from eclipse we know when its time to launch the core and kernels. Eclipse could provide calls, batch files or whatever is needed to launch jupyter stuff. As a first step it is perfectly ok if this relies on some locally installed, external python engine with jupyer extensions installed.

does this answer your question?

Christian

On 05/31/2016 10:55 PM, Martin Klösch wrote:

Hi,

Jupyter basically consists of 3 parts. The core, clients, and kernels.

In the typical scenario users want to create a notebook. This is done using a call to the Jupyter core (the Jupyter executable). Clients can then connect and display the notebook page (typically via a webbrowser).
To execute code from the notebook, the core needs to launch a kernel. You can think of a kernel as an interpreter doing REPL over sockets. The kernel opens a list of sockets chosen by the core for different types of commands. The information is given to the kernel using a launch file (my term) as a parameter to the kernel executable. We cannot influence this behavior without modifying Jupyter (or at least I haven't found a way yet).

Tobias already has some code to launch the Jupyter core and connect with clients, so this is not the problem.

My current issue is the launch of EASE as a kernel. As described, the typical scenario is to launch an interpreter in a new process. Since we actually want to connect to a running eclipse instance rather than launching a new one, we need to find a solution. My current strategy is to start a simple command line tool that passes the kernel launch file to a running eclipse. The command line tool can be freely configured and will receive the launch file as a parameter.

I might have been unclear on my plans in the last mail. I implemented a command line tool that receives the launch file and the socket information for the running eclipse instance as parameters. It simply tries to read the file, connect to eclipse and send the configuration over the socket.
In the future eclipse should try to publish its startup to Jupyter and choose a port for each instance. We then can have a kernel for each running instance and can choose which one we want to use from the client. For this I will have a look at your code and hopefully can reuse some (the port choosing mechanism).

For now, to be able to progress with the actual kernel implementation, I chose a port to be used and skipped the launch registration part. The basic concept stays the same but I can focus on the rest of the code.

Hope this sheds some light,
Best regards,
Martin

Am 31.05.2016 14:32 schrieb "Jonah Graham" <jonah@xxxxxxxxxxxxxxxx>:
> Does anyone of you know of a better solution to this problem?

As Christian mentioned I think some more context may help to address this.

You can solve the issue of allocating ports though. Depending on who
is doing which part of the launching may affect this. For example when
launching Python from EASE I need unique ports in each direction. I
allocate first a port in Java, then launch Python, have Python
allocate the return direction port number and then Python calls back
into Java to notify it of the port.

Have a look of the current draft implementation of the Python engine
(using Py4J for comms):
https://git.eclipse.org/r/#/c/73149/7/plugins/org.eclipse.ease.lang.python.py4j/src/org/eclipse/ease/lang/python/py4j/internal/Py4jScriptEngine.java
In that you should be able to see that:
1. setupEngine starts the Java server side (setting java and python
ports to 0, meaning automatically allocate free port),
2. launches the Python process, passing the actually allocated port
3. waits for Python to notify it of the port in the other direction
(Python does this by calling back into Java method
pythonStartupComplete)

To allocate an unused port, use 0 when creating your socket, see
http://stackoverflow.com/questions/2675362/how-to-find-an-available-port
for a good example.

HTH,
Jonah
~~~
Jonah Graham
Kichwa Coders Ltd.
www.kichwacoders.com


On 29 May 2016 at 22:55, Martin Klösch <martin.kloesch@xxxxxxxxx> wrote:
> Hi,
>
> I've been working on launcher strategies for the EASE Jupyter kernels this
> week.
>
> The flow for launching Jupyter kernels is basically the opposite of what we
> want. To launch a new kernel, an executable is called with command line
> parameters, the most important of which specify the ports that will be used
> for communication. We would like to tell Jupyter which ports we are using as
> to have more control and be more customizable from within eclipse.
>
> I currently do not see an easy solution to this problem, so I decided to
> simply create a command line tool that will be called by Jupyter and
> dispatches the parameters over a socket to a running eclipse instance. This
> strategy has some downsides, the most obvious being that with this simple
> solution we can only have one eclipse instance running at a time.
>
> I would still like to keep this method for now as it allows me to continue
> with the actual implementation of the kernels. In the future it would be
> possible to dynamically add running eclipse instance to Jupyter using the
> same method but with different dispatching ports for each. This code would
> not be platform independent as Jupyter has different locations for its
> kernel information on each OS.
>
> Does anyone of you know of a better solution to this problem?
>
> Besides working on this topic, Tobias from the Science working group has
> given me his code regarding Jupyter clients in eclipse. I have been checking
> the code and it looks as if I could reuse some of it for the kernels as well
> (mostly zmq messaging). You can check out his code at:
> https://github.com/openanalytics/japyter
> https://github.com/openanalytics/jupyter-console
>
> Thanks again to him and the guys from the Science working group for the
> great collaboration. I think this project can help both teams and get EASE
> and eclipse even further...
>
> Best regards,
> Martin
>
>
> _______________________________________________
> ease-dev mailing list
> ease-dev@xxxxxxxxxxx
> To change your delivery options, retrieve your password, or unsubscribe from
> this list, visit
> https://dev.eclipse.org/mailman/listinfo/ease-dev
_______________________________________________
ease-dev mailing list
ease-dev@xxxxxxxxxxx
To change your delivery options, retrieve your password, or unsubscribe from this list, visit
https://dev.eclipse.org/mailman/listinfo/ease-dev


_______________________________________________
ease-dev mailing list
ease-dev@xxxxxxxxxxx
To change your delivery options, retrieve your password, or unsubscribe from this list, visit
https://dev.eclipse.org/mailman/listinfo/ease-dev


Back to the top