Alchemy Runtime Design: Layers, Bindings, and Durable Objects

Añadido: 2026-06-01

[00:00:00]Here we get into what I would call the runtime stuff.

[00:00:04]So beyond infrastructure as code, Alchemy also allows you to blend runtime logic, which is the implementation of your servers and your jobs with your infrastructure.

[00:00:14]And you get it all in one single type safe program.

[00:00:17]So if it type checks, you know all your infrastructure is correct.

[00:00:21]And what I'll show here is that we follow this very simple pattern.

[00:00:24]And it should look exactly like layer.effect.

[00:00:28]You got this layer.effect tag.

[00:00:30]And then you got your gen.

[00:00:32]And then you would pull your dependencies, right?

[00:00:39]And then you return your service implementation, right?

[00:00:44]Like get your dependencies, return the implementation.

[00:00:47]Alchemy is totally built around this pattern.

[00:00:49]I love this pattern.

[00:00:50]This is all you need to build anything, pretty much.

[00:00:52]So this is basically a layer for a Cloudflare worker.

[00:00:56]And in this first step, what we call this is initialization.

[00:01:00]This is where resources and bindings go.

[00:01:05]So you declare your infrastructure dependencies like you would any other sort of effect dependency.

[00:01:12]So here I need a bind to a bucket.

[00:01:14]So I call Cloudflare RT bucket bind, and I get back a client.

[00:01:17]So now I have a way to interact with my bucket.

[00:01:19]And Alchemy is going to go automatically wire up the infrastructure.

[00:01:23]So if I run deploy, you'll notice that the plan now has a worker, which is a new resource.

[00:01:29]But that worker has one binding, which is the bucket.

[00:01:32]So you see this in the plan.

[00:01:33]That's all automated.

[00:01:35]So AI really can't screw this up.

[00:01:37]It needs to interact with the bucket, so it adds the bind.

[00:01:41]That automatically figures out the infrastructure.

[00:01:43]And then it uses it.

[00:01:44]It's type safe all the way through.

[00:01:46]And I have this really just contrived example that you can see here.

[00:01:50]I'm just bringing in the HTTP server request.

[00:01:53]So it's all integrated with effect platform.

[00:01:56]So the way you implement HTTP is you just give us an HTTP effect.

[00:02:00]That's an effect that has one requirement, which is the server request.

[00:02:05]And then we integrate with stream.

[00:02:06]So we give you a nice effect native interface over Cloudflare, over AWS.

[00:02:10]Everything gets a purpose built effect native implementation.

[00:02:13]So here we can just pass that stream in with the content and it works.

[00:02:19]So deploy that.

[00:02:22]It'll go through, create the worker, upload it.

[00:02:26]And we should get-- I won't go to it yet, but we get our worker.

[00:02:30]I don't want to put something in R2 right now.

[00:02:32]So what I really want to show is durable objects, which I think are the coolest feature of Cloudflare.

[00:02:39]So I'll say let's add a challenge.

[00:02:41]Don't mention it too many times, otherwise Sunil will materialize himself.

[00:02:47]Yeah.

[00:02:49]That support that uses hibernateable WebSockets.

[00:02:54]So our docs are all optimized for markdown.

[00:02:57]So when AI is fetching these docs, it's getting the markdown.

[00:03:00]It can crawl the docs.

[00:03:01]Everything is sort of laid out in a step-by-step workflow that AI can just pull from.

[00:03:07]So you're getting this sort of interactive experience with any coding agent.

[00:03:10]So what we'll see here is it's going to create a durable object, the same pattern of export default class.

[00:03:19]We've got where we initialize our resources.

[00:03:23]And then we've got an effect that we return.

[00:03:26]This will run per durable object instance.

[00:03:30]So for those who don't know, a durable object is effectively a JavaScript isolate in Cloudflare's network.

[00:03:39]And you can have millions of them.

[00:03:40]And they're guaranteed to be singletons.

[00:03:42]And you can have long-lived WebSockets on them.

[00:03:45]They may have backed like 10 gigabytes of storage with SQLite.

[00:03:48]Super powerful primitive.

[00:03:49]You can have swarms of these small little isolates in the Cloud all over the world and just managing WebSockets and managing little bits of state.

[00:03:57]This is an effect-native way to work with durable objects.

[00:04:00]So we don't have any resources.

[00:04:02]So we'll just leave this empty.

[00:04:03]But here, you can see that when the durable object instantiate, we'll run this closure.

[00:04:10]So this is going to run per durable object instance.

[00:04:13]And it's going to set up-- basically restore the hibernated WebSockets.

[00:04:17]And then we'll return our fetch handler.

[00:04:19]So returning the implementation of the durable object, this is going to just simply upgrade to a WebSocket.

[00:04:27]And I'll also implement the WebSocket message and WebSocket close event handlers.

[00:04:32]And it's really unfortunate.

[00:04:33]I can't use EffectSocket here because they're hibernatable.

[00:04:37]It'd be really cool if this could just be EffectSocket.

[00:04:39]But for now, you have to implement it in this decoupled way.

[00:04:42]You can't pipe off the socket.