The unique feature of Zasper is that the Jupyter kernel handling is built with Go coroutines and is far superior to how it's done by JupyterLab in Python.
Zasper uses one fourth of RAM and one fourth of CPU used by Jupterlab. While Jupyterlab uses around 104.8 MB of RAM and 0.8 CPUs, Zasper uses 26.7 MB of RAM and 0.2 CPUs.
Other features like Search are slow because they are not refined.
I am building it alone fulltime and this is just the first draft. Improvements will come for sure in the near future.
IPython maintainer and Jupyter dev (even if I barely touch frontend stuff these days). Happy to see diversity, keep up the good work and happy new year. Feel free to open issues upstream if you find lack of documentation or issue with protocol. You can also try to reach to jupyter media strategy team, maybe they'll be open to have a blog post about this on blog.jupyter.org
I’m not adding a lot to the conversation, but it’s not often you run into someone who contributes to creating a tool so fundamental to your daily life, career, growth as a researcher etc, so let me just take the opportunity to say: thank you and the rest of your team for creating such an amazing interactive tool.
Thanks @carreau. I think the documentation is amazing! Zasper is built on the great work and documentation from Jupyter team. I will reach out to Jupyter media strategy team.
I think that's fairly normal, having alternative frontends can only be beneficial to the community. I know it also look like there is a single Jupyter team, but the project is quite large, there are a lot of constraints and disagreements internally and there is not way to accomodate all users in the default jupyter install. Alternative are always welcome ; at least if they don't fragment the ecosystem by being not backward compatible with the default.
Also to be fair I'm also one of the Jupyter dev that agree with many points of OP, and would have pulled it into a different direction; but regardldess I will still support people wanting to go in a different direction than mine.
The Jupyter community maintains a public spec of the notebook file format [1], the kernel protocol [2], etc. I have been involved with many alternative Jupyter clients, and having these specs combined with a friendly and welcoming community is incredibly helpful!!!
Vscode and vscode.dev support wasm container runtimes now; so the Python kernel runs in WASM runs in a WASM container runs in vscode FWIU.
Vscode supports polyglot notebooks that run multiple kernels, like "vatlab/sos-notebook"
and "minrk/allthekernels". Defining how to share variables between kernels is the more unsolved part AFAIU. E.g. Arrow has bindings for zero-copy sharing in multiple languages.
Cocalc, Zeppelin, Marimo notebook, Data Bricks, Google Colaboratory (Colab tools), and VSCode have different takes on notebooks with I/O in JSON.
There is no CDATA in HTML5; so HTML within an HTML based notebook format would need to escape encode binary data in cell output, too. But the notebook format is not a packaging format. So, for reproducibility of (polyglot) notebooks there must also be a requirements.txt or an environment.yml to indicate the version+platform of each dependency in Python and other languages.
repo2docker (and repo2podman) build containers by installing packages according to the first requirements .txt or environment.yml it finds according to REES Reproducible Execution Environment Standard. repo2docker includes a recent version of jupyterlab in the container.
JupyterLab does not default to HTTPS with an LetsEncrypt self-signed cert but probably should, because Jupyter is a shell that can run commands as the user that owns the Jupyter kernel process.
MoSH is another way to run a web-based remote terminal. Jupyter terminal is not built on MoSH Mobile Shell.
Cocalc's Time Slider tracks revisions to all files in a project; including latex manuscripts (for ArXiV), which - with Computer Modern fonts and two columns - are the typical output of scholarly collaboration on a ScholarlyArticle.
Docker Desktop and Podman Desktop are GUIs for running containers on Windows, Mac, and Linux.
containers become out of date quickly.
If programmer or non-programmer notebook authors do not keep versions specified in a requirements.txt upgraded, what will notify other users that they are installing old versions of software?
Are there CVEs in any of the software listed in the SBOM for a container?
There should be tests to run after upgrading notebook and notebook server dependencies.
> Any new package format must support cryptographic signatures and ideally WoT identity
Any new package format for jupyter must support multiple languages, because polyglot notebooks may require multiple jupyter kernels.
Existing methods for packaging notebooks as containers and/or as WASM: jupyter-docker-stacks, repo2docker / repo2podman, jupyterlite, container2wasm
You can sign and upload a container image built with repo2docker to any OCI image registry like Docker, Quay, GitHub, GitLab, Gitea; but because Jupyter runs a command execution shell on a TCP port, users should upgrade jupyter to limit the potential for remote exploitation of security vulnerabilities.
> Non programmers using notebooks are usually the least qualified to make them reproducible, so better just ship the whole thing.
Programs should teach idempotency, testing, isolation of sources of variance, and reproducibility.
What should the UI explain to the user?
If you want your code to be more likely to run in the future,
you need to add a "package" or a "package==version" string in a requirements.txt (or pyproject.toml, or an environment.yml) for each `import` statement in the code.
If you do not specify the exact versions with `package==version` or similar, when users try to install the requirements to run your notebook, they could get a newer or a different version of a package for a different operating system.
If you want to prevent MITM of package installs, you need to specify a hash for the package for this platform in the requirements.txt or similar; `package==version#sha256=adc123`.
If you want to further limit software supply chain compromise, you must check the cryptographic signatures on packages to install, and verify that you trust that key to sign that package. (This is challenging even for expert users.)
WASM containers that run jupyter but don't expose it on a TCP port may be less of a risk, but there is a performance penalty to WASM.
If you want users to be able to verify that your code runs and has the same output (is "reproducible"), you should include tests to run after upgrading notebook and notebook server dependencies.
The actual RAM issue is another one. Every Python kernel you start consumes around 100-150MB RAM. So unless you are starting different kernels using Zasper, the majority of RAM usage is still going to be the same.
Can I sway you to take this into a ... certain direction?
From my POV any browser based editor will be inferior to emacs (and to lesser extent vim) simply because it won't run my elisp code. While a fresh and snappier UI compared to eg jupyter would be nice, I would love to see something that integrates well with emacs out of the box.
So, perhaps it would be really nice if the backend+API was really polished as an end product itself in such a way that it could easily interface with other frontends, with remote attachment.
I could go on with my list of demands but I would be thrilled and amazed at my luck if even those two happen...
I'm curious what your thoughts on are on emacs-jupyter[0] which seems to integrate reasonably well with Org mode. I have some complaints about how it has to handle output blocks, but otherwise it seems like a great way for Emacs to act as a frontend to a Jupyter kernel.
I can't recall exactly right now (incidentally I started recording decisions like this to be able to answer questions like this - mainly for myself - but only recently).
To refresh my memory I just started it and tried using it with a julia kernel on a remote jupyter. To start, it wouldn't connect to https endpoint. Maybe because it's signed by a private CA? idk, but the mac trusts it for eg the browser and curl. Well anyway, let's forward the http port and try connecting to localhost.
Great, that works, and I'm offered some uuid as a "choice of kernel to connect to". I don't recall having one running before I connected, so it probably was started for me. How do I name it? Ah, there's `jupyter-server-kernel-list-name-kernel`, and now I'm recalling that whatever you name it as, will disappear if you quit emacs. Let's try.
Meanwhile, I import PlotlyJS and try to create a plot. I get complaints about WebIO (julia package that facilitates interaction with browser) like I do in jupyter (the package is old and doesn't work with current jupyter), except in the browser only the back communication (browser->kernel) is broken, for interactivity. Showing plots works. Anyway, PlotlyJS displays nothing. `Plots`, which renders to a png, somehow produces the axes but not data. Eventually I get PlotlyJS to display an image using explicit image mime types.
Still no interactivity - I would need node for that, to compile widget support for whatever reason - but it does display. I should retest widget support. Sending code to repl works, although at this point I'm used to seeing an overlay over variables that get set.
Ok. Close emacs, restart, go to session list (`jupyter-server-list-kernels`). Name has been cleared. I can reassociate the buffer to the kernel, but, if I have two, open kernels, how do I tell which buffers is associates with which kernel?
Overall it mostly works although there's room for polish. However, interactivity or any kind of bidirectional communication remains somewhat difficult.
There already is a library that can interface emacs with Juypter it is called ein. I think what you really want is a kernel that executes emacs code and if you did make that kernel it would probably work in any of these systems.
Yes, I'm aware of EIN. To start, it's been abandoned by it's author/maintainer as of April 2024 IIRC.
Further, I do not need a kernel to execute emacs code - I have one and it's called emacs. The point regarding executing elisp code was a cheeky way to state that I am not looking forward to finding replacement and/or porting of all the custom code - mine and others' - that my editor runs, and that no amount of "features" from a webui editor will ever replace that. Hence I also mentioned vim since over time it got customized for me as well and I wouldn't want to port that either. Nor the convenience of the terminal, which is what vim is for.
Putting that aside as with all respect and gratitude to the author, it was rather clunky in many respects - no interactive story, poor handling of sessions and remote kernels (have you tried to start one, disconnect and reconnect?), no integration with LSP, and lack of many many more features that /could/ be made.
I don't know how much use you make of jupyter kernels or mathematica notebooks or similar technologies, but in my case I explored the available landcape quite thoroughly and regularly revisit. I know what I'm looking for and EIN is/was not it.
[EDIT]
I just noticed you mentioned EIN but linked to emacs-jupyer. Used that as well, of course. Ill add a bit more detail to that in sibling
I'm guessing you already looked at org-mode code blocks which basically do the same thing as a juypter notebook without a web protocol, webUI and anything else if you wanted an experience that is easier to commit to a git repo and has a notion of cells which is the magic sauce for juypter (it was originally derived from ipython which is a command line interface). I am also a emacs user :)
Juypter has an interface and API built in. What Zasper is the reimplementation of the juypter protocol. You can see this at [1]. Juypter kernels are very different from Mathematica notebooks. Mathematica notebooks aren't related to juypter.
Juypter kernels encapsulate language runtimes so that they can be interfaced when called from a notebook.
> Mathematica notebooks aren't related to juypter.
I don't think that's fair. Rather, IPython, and later Jupyter, explicitly (successfully) sought to create a Mathematica-like notebook experience for Python.
I agree. The command line IPython by Fernando Perez was very inspired by Mathematica. He used Mathematica as a grad student and wanted a similar environment. In 2006-2007 Tom Boothby, Alex Clemesha and I wrote the first general used interactive web notebook called "The Sage Notebook", which became very popular with SageMath users over the years; the first version of Jupyter looked very similar to the Sage notebook. The Sage notebook was heavily inspired by everything around Google Wave and Google Docs (at the time), but definitely also by Mathematica's notebook. In particular, Alex Clemesha had recently been a physics undergrad and was a heavy Mathematica (and "Web Mathematica") user, and wanted a similar environment in a browser.
Thanks, I'm quite aware of org-mode. All my emacs config is in it, I have 1000s of LOC configuring /just/ org, I use it my computers and on my phones for any kind of information management, and I absolutely love it.
I think it can be very suitable eg when you are preparing a presentation, report, a paper or a repeatable analysis/process. Especially - as with most of those examples - if you want to interleave narrative and code/results. It is less suitable for doing exploratory analysis, for any kind of interactivity, for connecting to remote sessions (it's possible but clunky), for showing a chart that you can zoom into. For displaying a table with 10,000 rows, for displaying a large plot. Or multiple plots. For being able to zoom into a plot. It's not great at integrating with LSP and similar tools. Could be better at managing code blocks, though one could write additional helpers and bindings fairly easily.
And, finally, it is quite a pain in the ass to have the code stored in a document rather than as code since it does tie me down even to my beloved emacs. I develop most of my code as library code which I can directly import/run. During the development it is still helpful to see the results of running defined functions and to be able to interact with the dataset. I currently do have a solution and a workflow but the tools aren't ideal for it.
I want to be able to have my codebase run inside a docker container, to be able to `git pull` to update it on the remote without involving emacs on the remote end, without having duplicate versions of the code in the repo (ie one in the org document and one tangled) for me to manage, and I also want to be able to make a small change in vim and push it back without involving emacs.
> Juypter has an interface and API built in. What Zasper is the reimplementation of the juypter protocol. You can see this at [1].Juypter kernels are very different from Mathematica notebooks. Mathematica notebooks aren't related to juypter.
Thank you for the explanation. Up until this very moment I thought mathematica and jupyter were exactly the same. Just to make sure, when you say they are very different and unrelated, do you mean like matlab is unrelated to numpy+ecosystem, like how Honda cars are unrelated to Ford cars, or like how pandas is unrelated to excel?
It helps when you are actually familiar with the technologies before making any - especially contradictory - claims. Mathematica for all it's faults - primary of them being proprietary - has a quite finely polished product and jupyter notebook interface draws heavily from it. I'f I'm not mistaken it is the OG notebook interface, though I'm not making a strong claim here.
Mathematica also has an interface and an API built in. You can run mathematica (or is it "wolfram" these days?) code on a headless kernel, you can connect your notebook frontend to a remote kernel, and you can make your own completely independent UI using the APIs in the language. Alternatively, you can connect the notebook interface to a kernel in another language using J/Link MathLink or C/++Link APIs. Or you can embed the mathematica kernel into jupyter - an existing project/duct and run mathematica code in jupyter/Zasper/whatever. Or run it in their webui for the past .. decade at this point?
I'll give you the benefit of doubt and not assume that you are a trollbot but I sincerely don't understand your need to offer "first page of google" suggestions when you clearly don't use the technologies you're commenting on.
Congratulations on the launch! It's great to see alternatives to Jupyter. JupyterLab is an excellent, however creating editor for broad audience is challenging. I've found Jupyter difficult to use, especially for beginners. Managing kernels, Python environments, and installing new packages can be quite cumbersome. Are you planning to address these challenges in Zasper?
Yes, I tried Jupyter Desktop. It is fantastic, I like that you can double click on notebook file to open app. However, it might be a little to complicated for beginners, you need to setup Python and select kernels. That's too much.
I mean this without rancor or insult, but a lot of data scientists may use Python, but are definitely not Python programmers. They know the subset of Python necessary to process data, and literally not one bit more. They would have no idea how to create an iterator function, their own "with" handler, may not even know how to create a new subclass with a method. They just take data in, chew on it, and spit it out.
Again, not an insult intended to them. They have their job and they do it, and I don't know much about their world either, after all. And of course you can find some data scientists who also deeply know Python. My point is merely that modeling them all generically as "Python programmers" in your head can lead to a model that makes bad predictions, which I found in my brief stint in that world can include you building tools for them that expect more out of them than they have.
That's not to mention getting dependencies installed. I know a good amount about everything from the silicon up and it can still take some time to get to the point where I have a Python ML environment working. Debugging whichever vendor's barque build process, broken drivers, etc etc, not fun and not something we probably want every notebook user to spend time on.
Yeah, I will work on these problems and I already have solutions in mind. Just wanted to get the word out about the project first and see if the world actually needs something like Zasper.
I am really happy to see the welcoming response from the dev community.
I'm not directly involved with extending Jupyter Lab, but I'm involved with the results (and testing) of our extension on the daily basis. What I find very often to be the source of complaints is the error reporting. In particular, the kind of error reporting that just disappears from the screen after few seconds. If there's one singular feature of Jupyter Lab that I really want changed, it's this.
Just wanna say this is a really cool project, and I can't think of higher praise than me hoping I build something as cool as this some day! I've been meaning to learn Go for sometime now, and will be referring to Zasper for the future :)
The unique feature of Zasper is that the Jupyter kernel handling is built with Go coroutines and is far superior to how it's done by JupyterLab in Python.
Zasper uses one fourth of RAM and one fourth of CPU used by Jupterlab. While Jupyterlab uses around 104.8 MB of RAM and 0.8 CPUs, Zasper uses 26.7 MB of RAM and 0.2 CPUs.
Other features like Search are slow because they are not refined.
I am building it alone fulltime and this is just the first draft. Improvements will come for sure in the near future.
I hope you liked the first draft.