The first four Val Town runtimes
The core of Val Town is running untrusted code. We take your TypeScript code and run it on a server, safely, quickly, and flexibly.
Code in Node.js has unfettered access to your system’s environment variables, filesystem, network, and much more. This access is routinely abused in supply-chain attacks. In response, the Node.js project is adding a permissions system that can allow & deny access to the filesystem and other resources, but it’s still experimental.
To safely run multiple vals, we need to sandbox and contain them:
- Vals need to be isolated from the system they’re running on: they shouldn’t be able to read from the filesystem or learn about environment variables.
- Vals should be isolated from each other: you shouldn’t be able to affect anyone else’s val’s behavior, or read anyone else’s information.
On the other hand, if you were to make every Val run in a full-fledged Docker container, it’d be tremendously slow. You’re booting up a whole Linux machine with all the bells and whistles just to evaluate a few lines of code. It’d be pretty secure, but pretty slow. You could create a pool of Docker containers, but the memory and devops overhead would be significant.
Val Town gives you the flexibility of using modules and APIs from a number of different ecoystems. If you’ve used Node.js, great - your NPM modules should mostly work. Thanks to our current system, a lot of DOM methods work too - you can parse query strings with URLSearchParams and use new APIs like web-standard streams.
Safety, speed, and flexibility are often in conflict with each other. Some approaches would be secure but limited, like QuickJS. Others would be flexible but bring complexity and performance questions, like Dockerizing every Val. We need to find a system that balances these concerns.
So, those are the properties we’re aiming for. What have we tried so far? Buckle up.
Runtime 1: Node.js + vm (protoype phase)
The very first version of Val Town used Node.js’s vm module. Schematically, running a val looked like this:
Now, the documentation for the vm module warns that “The node
The Node.js vm module is incredibly convenient to use, but it was dangerous and didn’t check our boxes:
- Infinite loops in user code will take down your server
- It’s easy to escape the sandbox and access sensitive environment variables
- You can’t install NPM modules within the sandbox
- Scripts that run in the Node.js VM sandbox run in script mode, not module
mode, so you can’t use
Runtime 2: Node.js + vm2
Due to the security problems with vm, we quickly switched to vm2, a now-deprecated NPM module that, unlike vm, aimed to be a security sandbox. VM2 still used vm “under the hood”, but it formalized all of the security workarounds that people used to secure it. It used Proxies to try and prevent people from escaping the sandbox. A familiar exploit in vm is something like this, taken from the vm2 readme:
vm2 prevented this attack by tweaking all of the objects exposed to the VM context.
The vm2 maintainer valiantly fought to make this strategy robust, but in the end it was impossible to add security to an unsecure construct, and the sandbox escapes kept coming. Eventually vm2 was deprecated by its maintainer and isolated-vm is the recommended replacement for Node.js users.
Runtime 3: Deno + Workers
At this point, we started to learn some lessons. In broad strokes:
- The Node.js vm module isn’t a good base to build on.
- It’d sure be nice to run vals as modules, not scripts.
Of course, we hadn’t learned these lessons all the way - but we were getting there.
At this point, I had become really interested in Deno. Ever since 2018, when Ryan Dahl gave his excellent presentation on Node.js regrets, the Deno project had prioritized the exact kind of sandboxing and security we were interested in. A lot had changed since 2018 - they had even ported the whole project from Go to Rust - but the permissions system was a consistent focus, and all signs pointed to it being very high-quality.
With Deno you can import modules and approve or disallow, one by one, their requests to access the filesystem or environment. It’s like going from code running as root by default, to a capabilities-based system: iOS asking you whether you want some app to see your location.
Better yet, Deno had a Workers API that mimics the Web Workers API in browsers: you can spin up separate threads, with their own permissions, within an existing Deno process. It’s super fast – much faster than spinning up a new process. I idolize how Cloudflare Workers works - spinning up a V8 isolate per execution instead of a whole process. This looked like the way to do it.
So we implemented a separate Deno server that only handled Val evaluation: giving us three servers total (Remix for the frontend, Express backend, Deno server). This elegantly provided isolation between Deno and Node.js, so Vals were much farther away from environment variables or anything sensitive.
All in all, this was an improvement, but we still had further to go.
- However, Workers triggered memory problems that made the evaluation server slower and slower over time until it restarted. We were spawning and killing a lot of workers, and understandably this wasn’t a typical usecase.
- Server isolation had a performance cost: for a val that powers a website, for each request we’d receive the request in the Express server, serialize it, send it to Deno, which would deserialize and operate on it, send the response to Express, which would then send it to the user.
Despite these issues, the switch to Deno was a big upgrade, and we were significantly less affected by anything that untrusted code could do.
Runtime 4: Node.js + Deno processes
The current iteration of our runtime learns from the lessons so far:
- Deno is a powerful tool for simple containment of functionality.
- Memory leaks are scary.
Right now, Vals run as temporary Deno subprocesses of a Node.js server. The subprocess’s code is tiny: it uses node-deno-vm, a terrific module from Kallyn Gowdy, which works by exposing a WebSocket connection between Deno and Node.js and wrapping it in a postMessage-style API. The “guest” code that wraps vals is also now public, though it’s still an implementation detail and could change at any time.
While this gives far better isolation between user code and system code than our previous vm and Worker-based strategies, it takes a little more time to start up, so we use a pool of pre-initialized processes to avoid that time penalty.
This strategy also gives the ability to have more fine-grained communication with Deno, because it’s based on WebSockets instead of a simple REST HTTP request cycle. We can send multiple messages to control vals while they’re running, and receive multiple messages from them as they run. This is key to how we just cut 100ms from typical val runtimes - more on that on a future post.
|Dynamic imports only
|Static + dynamic ES Modules
As you can tell, we’ve been iterating fast. v4 is not the last version of the Val Town runtime: we still have more ground to cover on all of the dimensions that we care about. It’s heartening to see others contribute to this discussion - Figma’s adoption of QuickJS, after using Realms was not secure enough, Amazon open-sourcing Firecracker one of the components of Lambda, and Deno sharing details of how their Deploy product works.
We want to learn from everyone. Vals should run instantly, securely, and everything should just work. Try out Val Town and see for yourself how an instantly-deployed, zero-configuration script can work.
Oh, and if you read this and you have the answer and the experience and are the one to unlock the next-gen runtime: we’re hiring.Edit this page