Filecoin Compute Over Data Working Group, 25 Apr 2023

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Compute Over Data Working Group Session #16 (Taubyte)

Description

Sami from Taubyte joins us to tell us about the architecture and use cases powering the Taubyte decentralized compute network.

https://taubyte.com/

A

Right, okay, we're live hello! Everyone listening to recording, welcome to the computer data working group. I think this is meeting number 16 for us, so we're chugging right along and today we're fortunate to have Sammy from Talbot joining us to tell us about Cowboy for the community they're building the platform they're building, how we can take advantage and start using Talbot for some decentralized compute needs uh quick advertisement. We have Cod Summit coming up May 9th through 10th in Boston Massachusetts.

A

So please, if you have not go to Cod summit.io and I'll post the link here in the chat as well we'd love to have you join, give a talk if you've got something interesting to tell us about sponsorships opportunities are still available. So if you're a builder in the space- and you want folks to find out about your project- we'd love to have you on as a sponsor as well. So please join us for the event.

A

It's going to be a good group of folks that are building in the community with that I will turn it over to Sammy. Thank you for joining us, Sammy and I'll. Let you run with it when you're ready.

B

Yeah sounds good. um Thank you, uh Wes, for uh invite inviting me to uh this. uh This talk, um I love the opportunity to share what we're building and what we've been built in for some time now, which you know is uh basically a decentralized cloud computing, so I'm, the founder of Tau, White and um teleport is kind of like um a cousin of Pacquiao.

B

To some extent, like bacliao is more like a batch processing, uh decentralized Computing, where a Tau bite we, you know like focus on uh serverless right now, so like real time, decentralized, Computing um more more, like cloud computing, because um when, when you you look at what made internet successful, um it's actually cloud computing right, so uh most of like every not most, but every application that everybody uses is almost powered by the cloud right or cloud computing. That includes web3 apps uh as well. The problem that is today is uh cloud. Computing.

B

Implementations are centralized right, so cloud computing itself is just a model. It doesn't have to be centralized. It could can also be decentralized, and this is kind of like the path that we uh we have take.

B

You know like let's implement or re-implement cloud computing in a decentralized way, and this is important, uh because cloud is uh an amazing tool or cloud computing is an amazing tool to enable Innovation like when, when you have, when you have a tool where you can uh run all-purpose Computing in a a decentralized way that enables uh a lot of things that also reduces cost. So you can build faster. You can build truly decentralized um applications and solutions, but you can also avoid um relying on let's say, expensive, oracles, for example.

B

So you can build your own oracles instead of having to rely of what's existing and kind of like have to bear all these fees or rely on on one feature. You know protocols and stuff like that, so I'm going to share a deck here, just to go over like how it works high level and then um I'll show you some things that I've built quickly for this.

C

B

C

Go to present mode.

A

You can hear us: okay,.

C

A

um Sometimes I'll just go on a video mute, see if that helps.

B

Okay: um okay, okay, I'll! Do this.

C

C

C

C

B

Can you see my screen.

A

Coming across it's starting to.

C

A

Oh here we go got it: okay, decentralized cloud computing looks good.

B

Okay, awesome, so um I talked about some of the uh like some of the slides I already talked about so I'm, just gonna move straight to um what we've built um and how uh what kind of like uh layers we have and what kind of Technologies we we use.

B

So um if you can see the diagram here um and they're like on the uh lower uh like layer there, you can see that we're using um things from ipfs, so only P2P, dag, PhD and stuff like that, so it's uh it's basically um ipfs uh compatible and then, on top of that, we implemented Cloud features as uh peer-to-peer protocols. So um registry is a peer-to-peer protocol.

B

um Ci CD is implemented by two products, vocals and so on, and then, on top of that we have the execution and orchestration layer which has the taobite web assembly virtual machine, which is an all-purpose webassembly, virtual machine that powers the functions or decentralized, serverless functions and smart Ops, or what I also call serverless Ops, which are basically serverless functions that are executed before provisioning and can alter provisioning. So, for example, you can say Do not execute a specific function unless the the node that's running the function has a GPU or is taking it.

B

You can also alter it in different other ways like leveraging 5G apis. Do you have a specific slice or provision the slice before provisioning, the function or the storage or whatnot?

B

Then, on top of that, so it's a cloud native, serverless interface. That's get operated, so you alter your project in the state of the network, pushing to git and- um and you know, you push your code to get and you or you config and then that's built and pushed into the network by the CI CD protocol, and on top of that, we've built tools.

B

So a web console uh that I'll show uh here shortly: uh Tau CLI, which is a command line, and then we have other tools that we've built too, like Dreamland, which is running like a little uh tabbyte Cloud Network locally in your laptops, you can try things locally before push into uh Network um and then talk order which is kind of like a tricorder, but you know for um for tall white and you can explore the network.

B

Troubleshoot, you know protocols stuff like that from a deployment perspective, so Talbot itself is not like a service provider. We are building a software that organizations can deploy to create clouds in our Visions. We see millions of interoperable clouds. That people is, um some of them will be private clouds, and some of them will be public.

B

The public wants, some of them will be, you know, uh powered by tokens or crypto, and some of them will be uh regular, Fiat based networks, but all these networks will be decentralized and and um uh uh uh permanentless and make developers can leverage. You know multiple of these networks at the same time as well from a deployment's perspective, so we created like uh tools to deploy it easily. So um deploying a network is config or joining a network is a very simple yaml config then run in a command line. We call Sport Drive.

B

uh If, in the audience there are some trickies, you recognize what that is, and you just do. Sport Drive displays you give it the name of the network and you have a new network deployed in a few minutes or you can join a network with your notes uh like that. As well, so why did we start with serverless?

B

So serverless is very powerful because we can treat it as like just like content addressing so um it's it's like combined with content to dress in now our our serverless functions or uh location, independent and they're also immutable and cannot inhibit. All the nice features that um a file on ipfs has like immutability verifiability and so on. The duplication.

B

um And then we use webassembly because it's portable, it's interoperable, so you can use multiple languages like every module could be a different language and they can talk with with each other.

B

um um Like, for example, uh you can write, um you can write a ZK proof with circum the witness compiles the web assembly and then your function could be a rust or go and they can communicate with that, uh or we also did like a lit protocol integration. Their code is is rust, but you can call it from either go or rust um and then the cold start, which is a serverless term like how long it takes for the function to be available like uh provision in this container or provisioning.

B

The web assembly runtime with the modules and everything it's way shorter, with webassembly and with predictive provisioning, can be uh it kind of like virtually zero, um and then there is possibilities to verify, build and executions that we're working on uh and exploring as well. uh So this is like what a function looks like it's very simple code, and um this is just a very simple function that takes an HTTP event and return pong.

B

uh It like go on Rust, and it goes uh well with uh the diagrams I'm gonna go through so kind of like explain how a request on topic works and that way you have like an idea how it's different from a centralized uh system.

B

um So let's say we have a client it's uh trying to get a specific endpoint, so example.com ping.

B

The way it works is uh the first request is a DNS request, always and the resolution happens through a protocol instead of a static, DNS server um or clouds layer or whatever centralized DNS server uh is generally used, so that goes to a node run in what we call the Seer protocol, which is connected, Network knows what's going around and it will grab the nodes that have capacity their IP addresses and return that to the client um to to be serviced.

B

So the client is established, an HTTP connection. It says, like okay, give me slash Bing. The node that has been selected uh by the client will execute the the serverless function and return pong right, so just just kind of like high level uh in details uh a little bit more things happen.

B

So when the when the uh client send that, like HTTP request, the node might not know what example.com Ping is, so it's going to connect to nodes that run the Tau byte name system and it's it's a it's a registry protocol and it will say like what is this? uh Can you tell me? What exactly is this and TNS will return informations about what that is in this case? It's a d function and it will return things like the the ID of the D function.

B

Parameters like how much memory um uh how long it will run like the time to live and one of the things that's in there, too is the what's a module CID and so the the node will grab the CID from all the nodes. If it doesn't have it, we we use another protocol to do pinning and we don't call it pinning. We call it actually hoarding so node stash like cids to make sure they have a specific number of replicas. uh It's a better way.

B

We think to have a cloud native, a way to define replica. So you can say new configuring, Network I want three replicas.

B

You know modules and the neck will do its best to keep a minimum of three replicas. If you, if you say like and then when you create a storage um like resource, you can say like hey for my storage bucket or distributed storage bucket, you can say like keep. You know, five replicas and, uh and that can differ from one project to the other and hoarder will make sure that is the case in the network.

B

B

That's doing like opinion, or anything like that, um so once the module is retrieved, uh dependency will be parsed and loaded. This dependency could be in the same project cids as well, so other DNS requests will happen. uh It could be just other modules in ipfs or they could be in HTTP. We have a plan to have like a kind of decentralized repos, for we don't have it yet. But for now you can just grab modules from local local project or you know, make an ipfs, uh ipf, scid, HTTP and so on.

B

So these will be loaded, the SDK will be loaded and then the function will be executed and that's what return Returns the uh the punk response and just to resume all. Like uh recap, all of that, like DNS requests for example.com uh sear, does that it's Dynamic. It looks what's going on on the peer-to-peer, Network returns that and then uh the second request is uh this.

B

Next request is HTTP to one of the nodes, the node scrubs, the function executes it and Returns the response, and you might think this is all slow, but it's actually very fast. There is cash and obviously happening as well. So uh we uh did some benchmarking compared to AWS Lambda, it's uh eight to ten times faster, so basically web3 faster than web 2 just like or what it should be.

B

um We built tools, so the tools I was talking about. I'm gonna, walk you through some screenshots and then I'll walk you through the real thing, like a little demo here so web console it's a very easy tool to use uh to build your project. So it's just console.tabai.com.

B

um Cli, it's a it's a command line, that's very easy to use. We made it so um it kind of like reads like English like Tau new function, creates a function or tell new projects, creates a project um and then Dreamland is running. You know a TalkBack Network locally. It goes with audible, uh which is like the graphical interface.

B

You can see your notes, you can add, notes, kill, nodes, run, fixtures and stuff like that, um and then we have more tools that I talked about, so sport Drive talk order and Q, um and then this is something I I wanted to show. uh Actually.

B

So this is a um ipfs Gateway or a serverless ipfs Gateway that that I've built um doing ipfs thing it's roughly like 40 lines of code, so you can see there that it's getting like an HTTP, an HTTP event grabbing the CID from the the query and then parsing the CID, uh creating on ipfs uh kind of like a shadow node, because, like it's just gonna like making an interface to webassembly, there is only one ipfs node in the node, which is the node itself.

B

um And then it's going to open the file. Give us a reader and we're just gonna like take that like that reader and send that over to http and the client will get it and I did this pretty quickly, roughly like 15 minutes and for today today's presentation I even added some image processing. So let me uh let me stop sharing this screen and then I'm going to share another screen with.

B

The code here in a second.

C

B

Okay, so this is uh like the the same, um the same endpoint, um and this is like the same endpoint processing, the CID so basically grabbing the CID. Sorry, that's uh gots up demos uh not being happy at me. Okay, so grabbing the CID from ipfs, doing some processing and then send it that uh sending that to the client. So let me show you what it looks like so go into uh log out here. Just to show you like what uh the web console looks like.

B

All right, so uh web console is just a web interface uh that is actually going to clone your repos instead of uh anything else. So uh there are a couple networks here, so there is the old sandbox. That's that's deprecated. There is the new Sandbox you can. If you have Dreamland running it will show Dreamland here or you can connect it to your own Uh custom Network as well. So I'm gonna log in here with my GitHub gonna, go fast because I already have that uh going and then I'm gonna open this project.

B

If you have a project elsewhere, you can just import it just registering the repos and that specific Network that the Network's not really like the like. There is no login or whatnot really in the network. Your source of Truth is really uh your repos, so this is the project that I have right. Now, uh that's the ipfs project, so you can see here. uh These are the locations of the nodes in this network um and then I'm going to go to functions and I basically have two functions. The simple one that you've seen before.

B

Is this one? So it's just get get the CID and then um returns. We make it slightly bigger and then basically uh forwards. You know uh the content to the client, the one that's doing, processing. Let's look at this one, um it's exactly the same function! So it's uh it's HTTP.

B

um It takes uh 10 seconds to run TTL, but it will be, cannot stop if it takes more than 10 seconds uh 10 Megs of ram. It's a get request. It is connected to a generated domain um and then it's on slash build.

B

um This is the uh yaml config, that's also on the repo and then the code is right here. So I'll show you the what's different. So get is basically the other get. So it goes to ipfs grabs the file Returns the read closer and then build because I'm using the build Library just called it build. So it's going to get the reader using get and then process it or make an image. So image.d code. Note that I also imported jpeg here.

B

Otherwise you won't decode jpeg and, and then um it's gonna invert, resize and rotate and then convert it to a PNG and send it over to the client with the right here right. So super simple, I I like made this in in a few minutes. You can also- and then once you made modifications you just hit like push here, and this will kind of like walk you through like uh what was modified.

B

Then uh like did you modify code or config or whatnot, so um I'm gonna do actually something here: I'm gonna, create or host like a um a um a website here quickly, just to kind of showcase. How? How easy is that so I'm gonna call it uh card.

C

B

Simple, so I'm Gonna Leave it to generate here um so the uh the interface here will will generate like a Reaper for me uh for domain. I'm gonna uh use uh the generated domain here and we're gonna host it on slash, which is uh where most websites go and I'm gonna hit. uh Okay here and it's gonna ask me to select a template: uh I mean you can do whatever you want, I'm going to go with view here and hit generate, um so this will generate a website. It's a separate repo!

B

um That's um that's been added to the project here. So now in my project, I have a repo or a website. I need to push that config. First, to tell the network that uh now uh I have like a website and uh I'm gonna push that. So this is just config and say like added website and hit finish. So this is going to push to GitHub like there is nothing like really um a special here happening if I go to config repo.

B

You can see my commit here added website, and this is the yaml file I added. If I go back to this interface and go to builds, you will see the build here that should show up here. Okay, so it's a config build. It was successful. So if nothing happened yet because the website still needs to be uh uh built, so I'm gonna go to my website open the repo of my website and modify something here. Just so I trigger a build.

B

Let me modify actually the.

B

C

The Source here.

C

B

I'm going to commit so this is going to trigger a build. So if I go back here, go to builds, you can see here that it's it's triggered the build of the website. It's a magic build. Obviously, uh if you go back here or to any of your functions, you can see like there is a tabbyte uh folder that has like a config yaml that describes how it should be built.

B

So this build we're using like node.js uh container and it's running the build script here, which is just doing like yarn, install audit, fixed and any on build and then move in whatever was built to slash out.

B

um Let's see if the build is done, it's it's successful. You can download your website here. You can see the logs of the build as well right here and the locks are also a Cid in the network. Obviously, um and then let's go to the website here and click on this little like thunder icon, which will open the website.

B

Oh good demos, okay, uh so this is our new website, so welcome to code cloud and this hosted on a generated domain here. uh The network will also get like a certificate uh from let's encrypt dynamically and and reply with this super easy uh same thing with functions. uh I can create a function here uh very easy. If I go to create a function, you can just do select template and let's do Rust for a change. We're gonna do a ping pong with rust.

B

Yaml describes the function, Cloud native, like I, said, and then the code here is our code. We're going to do code.cloud, loves rest uh uh and then hit enter same thing by the way. So you can see this is like the the the build script to uh build this module um and then we're gonna attach it to a generated domain and give it a little path here. So we're gonna, call it rust and hit enter and push everything, so it compiles. So we have like config, which is basically um like defining the function itself.

B

We have the code, that's right here and obviously you can also create separate repos for code that that's, what's called libraries here in the interface uh uh on the bus code finish. So this is going to push uh to uh GitHub. So if I go to uh project detail and go to my code repo, you can see here added rust code and that's my that's my diff here.

B

um Moving back to build you'll see that there is a build happening.

B

There is a build happening right here. uh Let me see um this one, so there is a build happening. uh It should feed on in a few seconds here.

B

And then we can execute our function.

B

Obviously it's rust so it takes longer to build than than go.

B

Like you'll know that Ross is not like or go actually was designed to build fast.

B

It's gonna take some time here before it builds all right boom. So successful you can see here like all our functions, you can download the wasum file of any of them um and then, if I go back to my function- and this is my ping pong in Rust and just go there.

B

um Okay, let me refresh this okay, so card.cloud loves rust, so that was executed by a node in the network and your function is live just like that um very easy, very quick! You can explore your config. If you go to config production, you can see what TNS is seeing for a specific project and then everything I did here can be done in command line as well.

B

So if I have like a a little bit of time here, I don't know if I do I could show you how to how all of that can be done uh in Tau. Do I have time West, or should we reserve that for questions.

A

We do have time, yes, we've got at least another 15 minutes.

B

Okay, should we do questions or yeah.

A

Let's, let's do questions as well and actually just to kick it off one question for you: it's it's always interesting to see folks in the community which types of Technologies they prefer to start with. um You know some people will start with batch Technologies. Some people are more kubernetes Centric or they want to implement something like ec2 and obviously serverless functions which which direction you've taken with Talbot is probably the most popular.

A

But could you just share more about when you were getting started and you're looking at different architectures how you got into the thinking of settling on this architecture? What was the driver for you? Were there any trade-offs? You were thinking about.

B

Yeah yeah yeah, so.

B

um uh The the uh the reason we uh we went for the uh decentralized architecture first uh was because we wanted uh mainly um to uh enable uh cloud computing for web3, but we also wanted to uh kind of like do in that enable horizontal scaling, which is very important for uh Edge Computing, um uh like how do you?

B

um How do you scale horizontally uh through different um through, like you know, a large number of servers that are uh or devices that are in different locations and also most of the time owned by different uh peoples and organizations? And this is why we cannot Excel on on this architecture and um lead. P2P and ipfs was like a technology that made sense to leverage to build on top um and then lip P2P also offered a lot of nice Primitives like streams.

B

So we could, you know, uh build peer-to-peer, Communications um and then uh web assembly, like from the um um kind of like execution side of it, webassembly, made sense, uh because um it was something that's very interoperable you compile once and then you can execute it on different architectures, especially you know on your scaling horizontally. You don't know if the node is going to be an Intel, an arm or something else. So you want that code to run all over the place. We wanted to enable interoperability between languages, but also webassembly is easy to verify.

B

So it's easy to verify the code that made it like immutable and you know like a VM, smart contracts or webassembly right so, um and you know like uh you cannot, you cannot modify the code uh without modifying, in our case the CID, and we can verify that, like notes can verify before execution and then there are a lot of things happening in webassembly, so it's easier to secure the context, it's easy to create, create like plugins and and make development easy for developers like provide powerful tools and sdks, and then there there are things happening like uh verified.

B

You know verified, builds and verified executions that we are looking into also integrating with top bytes so um uh with.

B

Of course, there is also a trade-in like if you want to like trade in speed for uh verified execution, and uh you know uh um that you could you could you could like um compile your awesome code to generate a ZK proof and, um and, and it's gonna be like uh the modified was on file that at the end will generate the ZK proof and that ZK proof can be used to prove that the execution was correct right, but obviously that's going to be slow, but there is also always a a trade uh trade-off there, um uh but but webassembly definitely a powerful tool um and I mean we believe that most software is moving the like toward being serverless.

B

So that's what why we went the serverless route.

A

Fantastic Gmail I'll give you some space. If you have any questions as well, no I got one more for Sammy.

C

A

uh One last question: for you: in terms of use cases that would take advantage of Tau bite. You know, would you like to see folks building websites on Talbot? Would you like to see them doing decentralized, Finance uh things or what sort of scope of use cases would you like to see over the next few months.

B

Yeah yeah um they're, like from from a developer perspective, like uh would love to see like people um you know uh hosting like front ends and and building um uh or kind of like moving that back-end logic. That's you know in general, in AWS, Lambda and other places uh to uh to be decentralized, uh also building like uh instead of relying on uh rigid and expensive oracles, maybe building their own oracles um uh and basically accelerating development in general.

B

From a deployment perspective uh like we have tools to easily deploy networks and uh we would love to see uh networks being deployed and created. uh You know whether it's a network that's uh publicly available, so people can use it, and you know uh you want to like monetize um your infrastructure or you can also use Top by to create a protocol. Let's say you're, creating an AI protocol or creating like a gaming protocol and want developers to have more powerful tools.

B

Not just you know, uh API calls, so they have like T functions and storage and key value stores and direct access to your Primitives through webassembly, which makes it you know way more powerful and way more versatile to build interest in things.

B

Exactly yeah and then and then just kind of like, if you don't mind, uh answering like a question here, so um we have a website for documentation, it's Tau dot, how t a? U dot, h, o w, and then um we have a Discord uh server, uh uh Discord dot, GG, slash uh tall byte. If you need any help uh with anything.

A

Very good all right! Well Stephen, do you want to spend any more time showing any other things to wrap up or or do you want to go ahead and call it a call it a wrapper today.

B

Yeah yeah yeah I need to go downtown Austin to uh join the uh um econ day uh downtown so uh yeah I mean thank you for the opportunity. uh There are some videos uh on TaoTao and then um we're like we're happy to um uh to help and answer questions on Discord um at any time.

B

um uh So yeah yeah, please please try it uh give us feedback uh break it. If, if you can break it too that'll be awesome, so you can fix it, um but yeah. Thanks for the opportunity, yeah.

A

So, thank you so much for presenting today Sammy looking forward to seeing you at the city, Summit and uh good luck at the conference this week.

B

Yeah, thank you. Bye, take care bye-bye.