Launch HN: Moonglow (YC S24) – Serverless Jupyter Notebooks

[tmychow]

Hi Hacker News! We’re Leila and Trevor from Moonglow (https://moonglow.ai). We let you run local Jupyter notebooks on remote cloud GPUs. Here’s a quick demo video: https://www.youtube.com/watch?v=Bf-xTsDT5FQ. If you want to try it out directly, there are instructions below.

With Moonglow, you can start and stop pre-configured remote cloud machines within VSCode, and it makes those servers appear to VSCode like normal Jupyter kernels that you can connect your notebook to.

We built this because we learned from talking to data scientists and ML researchers that scaling up experiments is hard. Most researchers like to start in a Jupyter notebook, but they rapidly hit a wall when they need to scale up to more powerful compute resources. To do so, they need to spin up a remote machine and then also start a Jupyter server on it so they can access it from their laptop. To avoid wasting compute resources, they might end up setting this up and tearing it down multiple times a day.

When Trevor used to do ML research at Stanford, he faced this exact problem: often, he needed to move between cloud providers to find GPU availability. This meant he was constantly clicking through various cloud compute UIs, as well as copying both notebook files and data across different providers over and over again.

Our goal with Moonglow is to make it easy to transfer your dev environment from your local machine to your cloud GPU. If you’ve used Google Colab, you’ve seen how easy it is to switch from a CPU to a GPU - we want to bring that experience to VSCode and Cursor.

If you’re curious, here’s some background on how it works. You can model a local Jupyter server as actually having three parts: a frontend (also known as a notebook), a server and an underlying backend kernel. The frontend is where you enter code into your notebook, the kernel is what actually executes it, and the server in the middle is responsible for spinning up and restarting kernels. Moonglow is a rewrite of this middle server part: where an ordinary Jupyter server would just start and stop kernels, we’ve added extra orchestration that provisions a machine from your cloud, starts a kernel on it, then sets up a tunnel between you and that kernel.

In the demo video (https://www.youtube.com/watch?v=Bf-xTsDT5FQ), you can see Trevor demonstrate how he uses Moonglow to train a ResNet to 94% accuracy on the CIFAR-10 classification benchmark in 1m16s of wall clock time. (In fact, it only takes 5 seconds of H100 time; the rest of it is all setup.)

On privacy: we tunnel your code and notebook output through our servers. We don’t store or log this data if you are bringing your own compute provider. However, we do monitor it if you are using compute provided by us, to make sure that what you are running doesn’t break our compute vendor’s terms of service.

We currently aren’t charging individuals for Moonglow. When we do, we plan to price individuals a reasonable amount per-seat, and we have a business plan for teams with more requirements.

Right now, we support Runpod and AWS. We’ll add support for GCP and Azure soon, too. (If you’d like to use us to connect to your own AWS account, please email me at [e-mail].)

For today’s launch on HN only, you can get a free API key at https://app.moonglow.ai/hn-launch. You don’t need to sign in and you don’t need to bring your own compute; we’ll let you run it on servers we provide. This API key will give you enough credit to run Moonglow with an A40 for an hour.

If you're signed in, you won't be able to see the free credits page, but your account will have automatically had free credits added to it.

We’re still very early, and there are a lot of features we’d still like to add, but we’d love to get your feedback on this. We look forward to your comments!

[randomcatuser]

Cool! Optimal UX!

What do you think of this compared with running a Jupyter server on Modal? (I think Modal is slightly harder, ie, you run a terminal command, but curious!) https://modal.com/docs/guide/notebooks

[tmychow]

It seems like there are two ways of using Jupyter with Modal. One is adding Modal decorators to specific functions within a notebook, and for that use case, I think Modal is fantastic (if you are using GPUs on a function-by-function basis)!

On the other hand, if you are trying to run an entire notebook on a remote machine by starting a Jupyter server with Modal, then the workflow with Modal is not that different from other clouds (e.g. you can start an EC2 instance and run a Jupyter server there). For that, Moonglow still makes it easier by letting you stay in your IDE and avoid juggling Jupyter server URLs.

Also, you might need to use a specific cloud e.g. if you have cloud credits, sensitive data that needs to stay on that cloud or just expensive egress fees. One of Moonglow's strengths is that you can do your work in that cloud, rather than having to move stuff around.

[AnotherGoodName]

We can literally transpile and run everything client side these days. You can run a recompiled quake 3 inside your browser. Why is a hosted notebook anything other than a static html+js+css site that runs behind a cdn (effectively free to host)?

I suspect that’s a matter of time right?

[dinobones]

I too share the dream of transpiling everything/WebAssembly/everything can run everywhere, but we're pretty far.

For a lot of ML/AI workloads and tasks, Python is just a binding for underlying C/C++.

It's already a nightmare to try to reproduce any ML/AI paper, pip breaks 3 times, incompatible peer deps, some obscure library emits an obscure CLANG error that means I need to brew install some libwhatever, etc...

I don't think the WebAssembly toolchain is quite ready for plug and play "pip install" time yet. I hope it eventually will be though.

[bblcla]

A lot of the people we've talked to who get the most value out of remote compute are doing really intensive stuff - they need server-level resources far beyond what you can find on a consumer laptop!

Hopefully someday you'll have 8 H100s on your Macbook, but I think we're still a long way away from that.

[AnotherGoodName]

In fact a quick google: https://github.com/jtpio/jupyterlite

Yay I really can have serverless notebooks! Not just an easy to manage server environment but literally a static html file that can be passed around and runs the full notebook environment. It’s weird it was ever done any other way.

[ukd1]

GPU time

[williamstein]

How do you deal with the filesystem? Eg do you make the local file system visible to the remote kernel somehow?

[bblcla]

We don't right now, but it's something a lot of people have asked for, so we're rolling out a file sync feature next week. (It will basically be a nice wrapper over rsync.)

[sidcool]

Congrats on launching. How does it compare to Google colab?

[bblcla]

Thanks!

The big difference is that Google Colab runs in your web browser, whereas Moonglow lets you connect to compute in the VSCode/Cursor notebook interface. We've found a lot of people really like the code-completion in VSCode/Cursor and want to be able to access it while writing notebook code.

Colab only lets you connect to compute provided by Google. For instance, even Colab Pro doesn't offer H100s, whereas you can get that pretty easily on Runpod.

[yanniszark]

Great work! Was wondering if you deal with transferring the python environment remotely. Usually a large part of the difficulty is dealing with dependencies.

[bblcla]

We make sure the remote containers have CUDA/Pytorch/Numpy/Matplotlib set up if you're using a GPU-based machine. It's actually far easier for me to run ML-based code through Moonglow now than on my Macbook - it's really nice to start with a clean environment every time, instead of having to deal with dependency hell.

We don't yet transfer the python environment on the self-serve options, though for customers on AWS we'll help them create and maintain images with the packages they need.

I do have some ideas for making it easy to transfer environments over - it would probably involve letting people specify a requirements.txt and some apt dependencies and then automatically creating/deploying containers around that. Your idea of actually just detecting what's installed locally is pretty neat too, though.

[ayakang31415]

I really like this concept. I SSH to my university HPC to submit Python script for ML related work (sbatch script.py), and sometimes I edit a script with VIM. Now I can use Jupyter Lab on HPC with port forwarding, but it is not as convenient as just running Jupyter lab locally. Does your software have some sort of command line features within Jupyter Notebook that can be run on HPC?

[jerpint]

I recommend VScode for that use case (other editors probably support this too) where you can ssh to your cluster, access all your files in the tree, and run Jupyter notebooks from VScode directly using HPC resources, very nice workflow

[joouha]

Toy could try using euporie [1] to run Jupyter notebooks in the terminal

[1] https://github.com/joouha/euporie

[fluxode]

Two questions:

1. How is this different from Syncthing and similar solutions? Syncthing is free, open-source, cloud agnostic and easy to use to accomplish what seems to be the same task as Moonglow. 2. What is serverless about this? It is not clear from the pitch above.

[bblcla]

1. It looks like Syncthing is an application that lets you sync files? Moonglow runs one layer below - we call cloud apis for you, to set up and connect you to a remote Jupyter kernel.

2. I think the serverless here is actually pretty literal - you don't have to think about or spin up a Jupyter server. We normally describe it as 'run local jupyter notebooks on your own cloud compute,' which I think might be a little more clear.

[aresant]

It's inevitable that access to the actual GPU compute winds up as an API layer vs captive behind proprietary systems and interfaces as the market matures

This is brilliant and "obvious" in a good way along those lines, congrats on the launch!

[tmychow]

Thanks! When I was doing ML research, every moment that my GPU setup was top of mind was a point of frustration, so hopefully we're moving the dial a bit towards making compute an abstraction you don't have to worry about.

[nobarpgp]

Congratulations on today's launch. Going to BYOC and test out the API on an A100.

Curious, are the SSH keys stored on Moonglow's internal servers?

[bblcla]

Excited to hear that!

There's no SSH keys to store - we start a tunnel from the remote machine and connect to that.

[daft_pink]

Will this work with zeds repl system that I believe uses Jupyter kernel spec?

I really like their Jupyter repl format because it separates the cells with python comments so it’s much easier to deploy your code when you are done versus a notebook.

[bblcla]

Good question! I'm not too familiar with Zed, but here's my high-level guess from reading the website: we don't currently integrate with Zed, but we probably could if it supports remote kernels (the docs I found at https://zed.dev/docs/repl weren't specific about it).

One nice thing about our VSCode extension is that it's not just a remote kernel - our extension also lets you see what kernels you have and other details, so we'd need to write something like it for Zed. We probably wouldn't do this unless there's a lot of demand.

By the way, VSCode also supports the # %% repl spec and Moonglow does work with that (though we haven't optimized for it as much as we've optimized for the notebooks).

[CuriouslyC]

Good idea and fun name (even if it doesn't tell me anything about the product). I don't see your path to a billion dollar market like YC usually expects to see though, do you have a plan there?

[bblcla]

Thanks!

One thing I've found while working in the ML space is that it seems like ML researchers have to deal with a lot of systems cruft. I think that in the limit, ML researchers basically only care having about a few things set up well:

- secrets and environment management

- making sure their dependencies are installed

- efficient access to their data

- quick access to their code

- using expensive compute efficiently

But to get all this set up for their research they need to wade through a ton of documentation about git, bash, docker containers, mountpoints, availability zones, cluster management and other low-level systems topics.

I think there's space for something like Replit or Vercel for ML researchers, and Moonglow is a (very early!) attempt at creating something like it.

[saurabhchalke]

The website UI looks really nice. I was curious if there is any library/framework/template for building websites with this "Anthropic" like UI?

[fabmilo]

What's the product? a visual code extension with a custom ipython kernel?

[whinvik]

Feels a bit like what Databricks already does with Dbx etc.

[bblcla]

I'm not super familiar with dbx (though its docs at https://docs.databricks.com/en/archive/dev-tools/dbx/dbx.htm... suggest it's deprecated).

However, looking at its replacement here (https://docs.databricks.com/en/dev-tools/bundles/index.html) - I think we're trying to solve the same problems at different levels. My guess is Databricks is the right solution for big teams that need well-defined staging/prod/dev environment. We're targeting smaller teams that might be doing more of their own devops or are still at the 'using a bash script to run notebooks remotely' stage.

[dcreater]

Is this similar to brev.dev?

[bblcla]

Not exactly. Brev.dev helps you set up a Jupyter server, but you still need to set up your connection to it yourself. It's also mostly designed to run on the browser.

Moonglow abstracts over this, so you don't need to think about the server connection details at all. We're aiming for an experience where it feels like you've moved your notebook from local to cloud compute while staying in your editor.

[xra_11]

Congrats on the launch!

[tmychow]

Thank you!

[mistrial9]

"Don't fork the Notebook format" --Fernando

[nickagliano]

I saw this comment on the marimo post too. Is there a source of Fernando saying this?