I am the author of Zasper.
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.
It's probably an unrelated post (apologies in advance) but I wanted to shoutout to the Marimo (https://marimo.io), it's the only Jupyter alternative that really got me excited, it's like Streamlit and Jupyter had a kid (and the kid took the best genes from both).
>> marimo notebooks are pure Python and stored as .py files
That sounds like a solid improvement. I’m going to give this a test drive. I feel like modularity is one of the hardest aspects of Jupyter notebooks in a team environment.
I’d be interested to hear if anyone has cracked a workflow with notebooks for larger teams. Notebooks are easy for solo or very small teams, and the literate programming style benefits still apply in larger teams but there’s a lot of friction: “hey just %run this shared notebook with a bunch of useful utilities in it - oops yeah it tries to write some files because of some stuff unrelated to your use case in there (that’s essential to my use case)”
My current best that I know of is to keep “calculation” (pure) code in a .py and just the “action“ (side-effectful) code in the notebook. Then as far as physically possible, keep the data outside of notebook (usually a database or csv’s). That helps avoid the main time sink pitfalls (resolving git conflicts, versioning, testing etc) but it doesn’t solve for example tooling you might want to run - maybe mypy against that action code - sure you can use nbqa but… interested to learn better approaches.
I actually think the problem you are describing is actually sometimes helpful from a design perspective, if you can be conscientious enough to periodically review your notebooks and figure out what is the actual useful code which should be properly integrated into the codebase vs what is the “one-off” / non-modular code. Like you mentioned, calculation vs side-effects is one way to help you decide but not the only. There’s definitely no single answer. The key is to just periodically figure out what ought to be refactored into library code, which notebooks should just be straight up deleted (hopefully as many as possible - you can always get them back in your git history if needed!), and so on.
I'm mostly in the camp that notebooks aren't that great for software development, they thrive as an "excel for coders" of sorts, but take a look at nbdev from fast.ai.
The literate programming aspect is very nice and I wish it was explored more.
marimo is really cool, albeit "pure python" is only true insofar as the diff is concerned. other than that, it's an unconnected group of functions that need the marimo runtime to stitch together.
would be cool if marimo could "unroll" the compute graph into a standalone python script that doesn't need the marimo library
it’s already possible to do this: marimo export script nb.py
Pure-python also helps to work with existing tools out of the box: formatting, linting, pytest, importing notebooks as modules, composition, PEP 723 inline metadata
If you are going to do that, you could stick with Jupyter + nbconvert.
I rarely use notebooks directly anymore unless I require the output to be stored. Do most everything in VSCode with interactive .py files. Gets you the same notebook-y experience + all of the Python tooling.
marimo notebooks are actually DAGs as cells, reusable out-of-the-box in three diffferent ways: as interactive computing notebooks in a reactive environment with no hidden state, as Python scripts, as data apps. So Jupyter + nbconvert (or perhaps you meant jupytext) is not a replacement for marimo.
What’s the advantage of this? It isn’t obvious to me that reducing memory usage and CPU of an empty/idle kernel is all that meaningful if the actual Python code in your notebook uses far more resources. It’s also not obvious to me how Go’s better threading helps, either, if all the computational bits are in Python anyway.
I have one nit with JupyterLab. When I press Ctrl+F, it takes ~0.4 seconds for the search box to open, and sometimes the first keystroke doesn't register when I type something into that search box.
"Zasper ... provides ... exceptional speed".
If they can just make input latency indistinguishable from vim, that's a very worthwhile value add.
It is quite beneficial for people who aren't writing python. And for them managing jupyterlab installations is a bit of pain.
I would like to use this with xeus kernel for sql (which is also native) and if this reduces the resource consumption of that setup significantly, its a big plus for me.
The README says the savings is ~75 MB. In most notebook workflows you’re at most running a couple at once. Saving <1% of my system memory doesn’t let me do anything I couldn’t do before. This also isn’t going to add concurrency/parallelism to your SQL unless xeus has some special magic that this is somehow able to exploit.
I was primarily talking in the context of shared server deployments for teams
uvx --from jupyter helps with that significantly.
Yes, the problem with such projects is that the must be very clear benefits for users (rather than developers) to attract a critical mass. At work we had Apache Zeppelin running on the servers alongside with Jupyther. In practice almost nobody used it (probably because almost nobody else used it, so if yourun into any issues you're on your own), so it was quietly shelved after a few years.
Honest question: what is not modern about JupyterLab? I know JupyterLab has existed for a long time, but continuous development has kept it modern.
My take:
- The UI is over bloated and bugged, sometimes things scroll, sometimes they don't, sometimes you have to refresh the page. You cannot easily change the UI as lots of CSS parts have hard coded fixed sizes.
- The settings are all over the place, from py files in ~/.jupyter to ini files to auto generated command line parameters.
- The overall architecture is monolithic and hard to break down, jupyter proxy is a good example of the hacks you have to go to to reuse parts of jupyter
- The front end technology (Lumino) is ad hoc and cannot be reused, I had to write my own react components basically reimplementing the whole protocol, come on its 2025.
- The whole automation around nbconvert is error prone and fragile
This is mixing quite a few different things (backend, frontend, auxiliary CLI utilities).
No time to write a lengthy reply here, but I think it's worth separating legitimate like-for-like comparison with a wider feeling on the ecosystem.
The need to start the server is really annoying. Especially when you have notebooks in multiple places, or multiple virtual envs.
This is why I moved to working with Jupyter notebooks in VS Code, there is no server to manually start.
Vscode will start the server for you, in practice. This is great if you just want to get going. It gives you a bit less flexibility though, if you want to do something fancy.
Vscode can also connect to existing servers. This can be very useful. For instance, you can put a ton of data and CPU in a server and work with vscode on a small laptop. If network latency is low enough, this works great.
While it looks like a great effort was put into this, an alternative has to support the same platforms, languages and related tooling, not run only on macOS, partial support on Linux, and IPython.
Then all the performance improvements by using Go, are taken away by using Electron.
For a fully fledged web app that all the major code notebooks tend to be, Electron makes a lot of sense. The bundled webviews built into OSes tend to be weak and outdated compared to the Chromium build that comes with Electron.
It's why Jupyter fits pretty well into VSCode/VSCodium.
> 5. Rendering your app
> Electron uses Chromium under the hood so your user sees the same on Windows, Linux and macOS. Tauri on the other hand uses the system webview: Edge Webview2 (Chromium) on Windows, WebKitGTK on Linux and WebKit on macOS. Now here comes the bad part, if you are a web developer you know that Safari (Based on WebKit) is always behind a step from every web browser. Just check out Can I Use. There is always a bug that you are not seeing from Chrome, only your dear Safari users. The same issues exists in Tauri, and you can't do anything against it, you have include polyfills. The winner has to be Electron here.
Then why can't Chrome/Edge provide a up-to-date library for everyone like mshtml.dll?
Well, there we have the whole issue that Web has literally become ChromeOS for all practical purposes of who is driving it, and browser market share.
It is called Webview2, because the first MSHTML.dll replacement was based on the original updated Edge engine, which Microsoft dropped for their own Chrome fork.
So either one cares to use portable Web development practices, or whatever Chrome does, with the side effect to increase its market share even further.
I wish the author considered https://wails.io/ for UI. Why all the effort with Go and ended up with Electron?
Because they use code mirror to build their ide
You can use CodeMirror with Wails I think, it's still web technologies but with a thinner layer compared to electron.
Honest question: what's the advantage of this over the Jupyter notebook support in VSCode? (which I use daily)
Or for that matter Emacs
How's the maturity compared to Livebook?
This looks pretty nice - this is specifically replacing the JupyterLab frontend and keeping the connections to Jupyter kernels - there shouldn't be any theoretical reason that it couldn't support Javascript or other language kernels, although I guess the project has only been tested with IPython kernels.
Would be interested to see where this goes.
Can I disconnect and reconnect from a running frontend (close and reopen a tab) and not lose any output?
I'll look later if this is allowed but I would love love an rstudio like interface in Jupyter. Being able to control enter to run a block of code (not line) in the accompanying repl is huge for iterating and building new things
As an example I love jupyterlab's "open console for notebook" but can't find a way of sending copied text to it, or switching focus with a keyboard shortcut
It's a big reason I can't do vscode Jupiters implementation
How does this compare to Marimo?
Doesn't seem to fix the invisible state problem that Marimo fixes.
From my POV is not fixing, is another way of work. I don't like what Marimo does, because for that I have scripts.
If I'm loading files from S3, I'm being charge for it. If Marimo re-executes this cell to maintain the state, it will charge me double. I don't need that. I'm able to organize my code, and know how it is being run.
I use mostly the script workflow, but for exploration Marimo is more convenient. It got also to-disk memoization recently. Kinda best of both worlds for exploration (although I'm not a huge fan of editing code in browser). In comparison to the JupyterLab hidden state spaghetti it's a fix.
With proper structuring of the blocks, Marimo will not re-execute the cell. Also memoization in script based workflows is still somewhat clunky on Python even with something like Snakemake.
I do find Marimo's approach, "global" variables tracked between blocks, less than ideal, but it's the best out there.
That demo gif is horrible.
You are not showcasing anything, but looping low resolution screenshots with special effects.
Good looking project I will check this out for sure
What's the point of this? Only benefit seems to be decoupling frontend in react. Nobody complaints about Jupyter performance. You can just build frontend and keep Jupyter as it is, it is already concurrent enough for multiple users use cases.
Install instructions seem incomplete.
> ... A Modern and Efficient Alternative to JupyterLab ...
This is not meant as criticism, just perspective. It's a classic development sequence:
* A team creates a powerful, small-footprint, REPL environment.
* Over time people ask for more features and languages.
* The developers agree to all such requests.
* The environment inevitably becomes more difficult to install and maintain.
* A new development team offers a smaller, more efficient REPL environment.
* Over time ... wash, rinse, repeat.
Hey -- just saying. Zasper might be clearly better and replace everything, in a process that mimics biological evolution. Can't leave without an XKCD reference: https://xkcd.com/927/
Again, not meant as criticism -- not at all.
> JupyterLab kernel
There is no such thing. There are Jupyter kernels. JupyterLab is just one of many UIs that speak the Jupyter protocol. Other examples include the original Jupyter notebook editor, VSCode Jupyter extension, and now Zasper.
I'm pretty sure Sage was always intended as a project that integrates the world, never "small footprint".
>> JupyterLab kernel
> There is no such thing.
A Web search reveals that the alternate term "Jupyter kernel," appears equally often. The terms are interchangeable.
> I'm pretty sure Sage was always intended as a project that integrates the world, never "small footprint".
A large install became true eventually, but it began as a small Python-based install, about 120 KB. Then people asked for extensions, and William Stein said "Yes".
No.
Thanks for explaining that background!
Only on HN!
Sagemath offers a different purpose which is scientific computing in order to compete with Mathematica and MATLAB. It offered a good interactive notebook interface which went on till about 2016, and later on was migrated to using the jupyter backend. It currently isn't well supported in Windows which is what you might have meant by the complexity. However it works pretty well with linux systems.
> Sagemath offers a different purpose which is scientific computing in order to compete with Mathematica and MATLAB.
Yes, that was its goal, when Python wasn't as evolved as it is now. More recently I've come to rely on Python libraries like sympy for symbolic processing. For these kinds of results Sage relies on a rather old environment called Maxima, and I think current sympy does pretty much everything that Maxima does. And as time passes Python libraries are beginning to provide some of the numerical processing originally provided by MATLAB (but more slowly).
> It currently isn't well supported in Windows which is what you might have meant by the complexity.
Actually I was thinking of JupyterLab itself. As time passes I find it more difficult to get it installed without library conflicts. But that can be said about many Python-based projects in modern times, which is why a Python virtual environment is becoming more the rule than the exception, in particular with GPU-reliant chatbots and imaging apps, to avoid the seemingly inevitable library version conflicts.
If memory serves, Sage now installs on Windows by creating a Linux VM to support it.
Yes, since sage 10.4 onwards, native installation on windows is not supported.
[dead]
[flagged]
What are you suppose to mean?
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.
That's stellar sportmanship right there.
Not that jupyter's team needed even more respect from the community but damn.
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 last paragraph let me think your normal is particularly collaborative lol.
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!!!
[1] https://github.com/jupyter/nbformat
[2] https://jupyter-client.readthedocs.io/en/latest/messaging.ht...
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.
Hello and thank you for making this!
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.
[0]: https://github.com/emacs-jupyter/jupyter
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.
Have you tried https://github.com/emacs-jupyter/jupyter Eg as org-mode code blocks.
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.
See https://github.com/emacs-jupyter/jupyter
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
> 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.
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.
Did somebody say eMacs? I dunno, I think VI integration could be more important.
I mentioned vim as well and generally proposed something that would be editor agnostic. Shoo! back to your cave.
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?
Have you tried the Jupyter desktop app? It's more self-contained.
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.
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.
Does it have a Racket kernel yet? I love using Racket for notebooks in Jupyter, but the UI is just too slow.
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 :)
On a quick glance, it seems it's possible to run this as a service similar to JupyterLab, right?
I'd be keen to offer it as an alternative to Jupyter on my little GPU platform experiment.
How was your experience working with 0mq?
It currently hard-codes launching ipykernel, right?
I mean, I appreciate the effort, but my average notebook uses gb to tb of ram and vram. At that scale having mb is...