Cloud Native Computing Foundation KubeCon + CloudNativeCon Europe 2020 - Virtual, 1 Sep 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Sponsor Demo: SUSE - Power AI Using Kubeflow / Deliver Applications Faster using Kubernetes

Description

Join SUSE Developers Marek Counts and Andrew Gracey for step by step demos about how to power AI using Kubeflow on SUSE Caas Platform, and how to deliver applications faster using SUSE Cloud Application Platform on Kubernetes.

A

I'd like to show you cube flow on a cast cluster. I want to show you first here that we have cube. We can run cube ctl. We see that our kubernetes version is 117, for this is the latest supported as of right now with the casp um command line tool, and it was very easy to set up and install with the casp tool and now is being automatically kept up to date with the latest patches security patches.

A

So this is our it's a very small cluster you. If it's a production cluster, you should never run with just a single control plane, but this is for demonstration purposes only and now we can go ahead and check on our cube flow deployment. Kubeflow. Is that thing that ties all the kubernetes components together for us?

A

So this is our cube flow deployment and there's a lot of stuff and we're not going to dig into all of it, but I'm going to pull out some of the things that you need here, but you have stateful set replica, sets your deployments, your services and your pods. So just a few things that I'm going to pull out is um you have a ui that you can look at and we'll actually go over and look at the ui in a little bit more detail.

A

Then you have your hatib hyper parameter tuning. You have your metadata servers, you have your ml pipelines and you have your selden server and your tensorflow work uh workloads. um So this is the general layout of your um kubeflow deployment and you can see that it's leveraging the coupe, the operators, the kubernetes operators for your machine learning, workloads and those deployments.

A

So this is the deployment and it's running on a casp cluster. So let's go ahead and jump over to the ui. So here is the ui for kubeflow, and this is the central dashboard with recent notebooks and activity.

A

This is actually set up using ldap, so we actually have authentication enabled and working so that you can use it in a large team environment where you want to restrict access to certain things, as well as harness the ability to manage resources. So what team gets? How many resources can all be done through kubernetes and kubeflow? By empowering you to use the namespaces to restrict that so that your your teams don't accidentally step on each other's toes, so the very first thing that we have here in kubeflow is pipelines.

A

I do apologize for the slowness. This is again going across the ocean, so our pipelines here are the things that are going to allow you to create and here's some example pipelines these. These are pipelines that give you these reproducible experiments and you can see each one of these. The really cool thing about these pipelines is each one of these is its own container.

A

So if you look here, this is a container with its own input and its own outputs, and this is really awesome because it a lot. It's harnessing the power of containers to give you a reproducible training, and not only that is it, gives you details at each step, because each one of these- and when you see artifacts over here, you'll notice that each one of these containers produces its own output and that's the input for the next step.

A

So the input from or the output from this container is the input to this container if it required input, and so these artifacts are built each time a pipeline is run and you can schedule the pipelines to run nightly or when there's changes or whatnot, and you get the you get to build up these artifacts, and so you get better insight into your pipelines how they were running, why they were running what decisions they made and you get this really nice ui and then you can switch over and look at the source of what that pipeline is.

A

These are the experiments we haven't run any experiments, but your experiments will run your pipelines.

A

You can create different sets of experiments with different parameters for those pipelines, so you could run the same pipeline with different parameters to for testing or to tune it to a different style, and then you can see your artifacts that were produced now. The next thing is notebook servers, and this is one of those things where I said collaborations was an issue.

A

What kubeflow does here is it provides you a a way to create notebook, servers and share these and restrict access, so you can restrict them via the namespaces or whatnot, and it provides you a way to share your kubeflow.

A

Your your jupiter notebooks as a server and limit the resources and whatnot to them and share them and collaborate as a team, and so that's what this does and how it works, and you can just click here, create a new server when you're creating a new server. You can create it from a bunch of different images.

A

You can even use a custom image if you have a custom image that you want to use. These are just some of the default ones. You can set the resources for your image and you can even attach gpus to it so that it has resources to use.

A

The next thing is hatib. It is a hyper parameter tuning, so this takes this takes things that aren't learned and it tweaks them and will continue to run them and change them to to try to get the highest accuracy or whatever your goal is for that and run it that way, and so this allows you to do some automatic things that will give you a higher accuracy at the end of your run, uh automatically and yeah.

A

So that's hyper parameter tuning and it's all automatic and you can enable it in your um in your pipelines and then you have your artifact store with all of the artifacts based off of your builds.

A

So this is kubeflow and we've got it running and we've got it going on our kubeflow in our casp install and we currently have the instructions for how to use this here and you. It will also point you to the manifest repo that we're looking to contribute back up to the upstream um as soon as we are finished, uh working on them and updating them a little bit to support the base, cryo, install and so yeah. If you'd like to try to install this on your cluster, please feel free to follow those instructions.

A

If you run into any problems, please just feel free to open an issue, and so that I I know and we'll start looking into it and I'd love to hear your feedback.

B

If you spend much time around the cloud foundry community, you hear all about the wonders of cf push. There is good reason for that. There is good reason for this. What allows you to do is in a single command, abstract, away the building or running of applications in the single command it uploads. Your code builds containerizes and runs it. It also sets up any routing connects any services that it needs.

B

This simple graphic shows the steps needed to build a container and deploy it in a typical kubernetes environment.

B

While you can obviously automate much of this yourself, it's error prone and costly to maintain, to give you an idea of how cf push simplifies this process, I'm going to run through three examples of how to use it to deploy a very simple node application.

B

The first is just a simple cf push from the command line to use. Any of these. All you need is app manifest. Here are the contents of mine.

B

Basically just says: here's the application and his name is hello world node, so it can be as simple as just these two lines stating the name or it can be as complicated as needed. My app is a pretty basic node.js application, so this manifest can be simple again just to show you uh what is in this application.

B

The the only dependency that I have is express.

B

There aren't any extra dependencies that you need for cloud foundry from node, so let's run cf push and see what happens.

B

So first thing that it does is it sees the it pulls the name and sets up a route with that application um based on the application's name.

B

Then it goes ahead and uploads the files to to cloud foundry, and then it starts working on it pushes over to the build pack.

B

If you caught that it was I'll scroll up real fast, it automatically knew that we were using the node.js build pack, that's based on the the package.json, so buildpass give you a very easy way to um to simplify your deployment or your your building of your packages.

B

So once it goes ahead and deploys or sorry once it goes ahead and builds it, it will go ahead and start the deployment, and now it's just waiting for for that deployment to start as you see here, it went it started it's in the running state and it gave us a route to use. So I've already got this loaded up on this, so I reload. I can see that it's running so, let's let's say we want to change it. It would be really easy. Let's.

B

Source- and we add that in so all we need to do is push again again uploads the files.

B

It goes through all the same steps as soon as it gives us the complete I'll reload the page and we'll see what.

B

Changed uh this is because everything is running in the cloud. This is much easier than having to um figure out like weird ssh piping, to be able to run, especially if you're building a uh the part that you're building is kind of a middle section in in a larger ecosystem a lot of times. Those can be very hard to break apart and actually be able to have running on your own development laptop.

B

So you this gives you a way that you don't have to worry about that. You don't have to worry about the parts that you can't really mock up or run in a local development environment, so that finished. So, let's go ahead and reload. You can see that that change took.

B

So the next is through stratos, so stratos is a ui that is built for a cloud foundry and kubernetes, as you see actually I'll back up. As you see it sees that our new app is there um for the sake of this demo, I'm going to go ahead and delete it just so that we've got a clean, a clean way to deploy from.

B

Okay, so that's deleted, so to create a new application.

B

We would go up here. My application is hosted in gitlab right now. We pick a organization and space that we want this to deploy into and with gitlab and github. One of the really interesting things is the um the integrations that we have, so I can just type in my namespace and it'll actually pull up which projects I have available.

B

These are note loaded in ahead of time.

B

So I can also come in and select the git commit.

B

So I made a get mistake earlier, so here's me fixing that good mistake uh here are all the options these can be set up in the manifest file or they can be created. They can be edited here, so I can come in and give it a different application name. If I wanted to I'm going to leave them as the defaults for right now,.

B

And you can see that it's doing basically the same thing it does. It uses all the same processes it uses the same, build pack really the benefit to using stratos is that you get a little bit more of a pull system.

B

So it gives you some control from your administration to be able to pull uh new versions of of applications as you want, uh rather than having your developers uh push uh new packages all the time.

B

This really, the benefits really depend on on how you you work as a team internally and also the application that you're you're developing.

B

You can see that this is coming up and yep now we're running, so I can go to the application summary you can see that deployed. It's online we've got one of one running. um You can do this even after you, but I'm going to go ahead and spin up two instances. You can see that it starts here. Stratos is a fantastic tool for managing your cloud. Foundry.

B