Events Add an event Speakers Talks Collections
 
MLconf Online 2020
November 6, 2020, Online
MLconf Online 2020
Request Q&A
MLconf Online 2020
From the conference
MLconf Online 2020
Request Q&A
Video
DevOps for Data Science With Kubernetes
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Add to favorites
152
I like 0
I dislike 0
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
  • Description
  • Transcript
  • Discussion

About the talk

Kubernetes is today’s hottest way to deploy and manage applications in the cloud, but it also offers the essential foundation for repeatable and reliable machine learning workflows. In this session, I will demonstrate open source tools that build on Kubernetes to facilitate solving data science workflow challenges for practitioners, without forcing data scientists to care about the primitive details of their infrastructure.

You’ll leave this talk with an understanding of how Kuberenetes supports data scientists at each step of the machine learning workflow. You’ll be introduced to high-level tools that effortlessly provision custom research environments, publish reproducible notebooks, operationalize models and pipelines as services, and detect data drift automatically.

About speaker

Sophie Watson
Senior Data Scientist at Red Hat

Sophie is a Senior Data Scientist at Red Hat, where she develops tooling for reproducible machine learning workflows and helps customers build machine learning systems in the hybrid cloud. Sophie holds a PhD in Bayesian statistics.

View the profile
Share

Okay, next week. Like to welcome Sophie Watson. She's a senior data scientist at red hat, and she has a presentation devops for data science with kubernetes, and I hope I did not destroy the, the title of your present, a second selfie. I like Tiffany, and Thank you. Will welcome. You can see some alpacas and I'm going to go ahead and see you all the time anymore. Please just do. One interrupts me Crest, if something isn't looking quite right right now. Today. I want to talk about how they see

it from things that software Engineers have been using for quite a while now, as a platform to practices and principles. But can I look at some high-level tools and technologies that are going to help us get benefit from practicing without having a chance to get to? Thank you. And I today, if I'm unable to answer your question today, please, dude. Just wait to untwist that email. So I thought I told you about the machine learning books. Like, what do we say to find it?

She'll be using the same language so fresh. And so when I say that, what metric I'm going to use to define. How will, I know when I've done a good job? Collection or cleaning? And you might already be collecting the one. I have to stop trying to extract information about the day that we want to expose with a machine learning model, type of picture that says be trained on Two Notch. From that warranty deployment into a production Model status and intelligent

application and that you can interact with but obviously a lot of different levels of production. And I went to in production before I go ahead and do a monitoring, and validation, changing changing habits and potentially updated model number. How it works. So many stages. You find yourself, backtracking changing decisions. He's made. We implementing different models and so on. I saw this web site is simply exploratory. But some of these stations need to code that will eventually run in production as possible pipeline to the code. We are monitoring

and filing station is going to be more or less directly used in production engineering directions ABC News on real data. And production to go has an extra clothes important dates of property auction. I also informed my production scoring pipeline. Ultimately. I'm thinking that we want something we can put into production, the extra fee. So that just makes metrics about our model that we have an idea when it's going wrong. Don't go back to this week's flyer. I think that the books is representing. Each of these stages are in practice. Not really though. As

a data scientist. I want to be spending my time, training a lot of friends in your business, but it does look much more like fit the bulk of my time seems to be spent between these Pages or repeating stages. So for example, when I go to get my daycare and I want to start training my Moto E memory on my local machine. I want to connect it to some different date. It's always be a headache and I'm outside of that pipeline as well. The machine mist of largest Software System that use

machines to make everything so much more difficult. Now. It's a technical debt in machine Learning System. If you haven't read this paper. System too hard to maintain because I'm at the hold these other part that much more engineering effort than that machine, but it's the dependency between the machine money card and the rest of the widest system that can make this whole application, even more complicated. And he saw system, is really difficult to identify the focus

of the day, is going to be everything around that they sift science. How can we make everything around that day to science works, like what? Well, so that we have practitioners can focus on eyes a to plan but you never going to see what I can do to make it easier to Filthy systems to execute Plex. Sorry. I'm two and a half hours to get back to focusing on that date. All fighting fighting with models dependencies and see a demo or two and introduce them high-level tools and community that

you should be aware of if you want to read these benefit. Okay, so, and I'll just talk about keeping I see which will get it to you in a moment. We need to start with containment container is a Xanax process that uses name facing Alamitos quite as to provide lightweight isolation, simple, right? So, we cannot think of a container as an Icee contains everything that you need to run an application or a person. So we just in case is a container pot for it. Feels applications out to block the containers. But first we need

to build those containers themselves. So well this container images made up of immutable a it contained a face image. That provides any minimum essential functionality, any Library dependency? Alabama code for the application itself in the future. And it's unable to make comparisons between two models or keeping everything else and comfortable catching spring break. Start for example, application. Do each process is coming in cold and not have service and these are stateless. They communicate with each other, three well-defined interfaces and

costly Stoute net worth at the outside world and let them Aldi's application that we can sell them up on things. She has now come and eat when you're running stuff in real life. You'll find that pot of your application is a button that keeps blowing down the application of the hole and it stays, this is really easy. So you can just run multiple copies of that. Microsoft has that was slowing down the whole process. What time do you do? Model scoring to make those protection? If you are running a social network

running that as a 7 and 5 milliseconds, j-school want to social media post, but you need to be able to go more than 200 seconds. Stay in case you can just roll out multiple copies of your model to do the scoring that by preventing the bottleneck. Make it easier to tell if I develop a new model. Or when there's a media posts. I spy more legitimate. I don't want to go out to millions of social media uses instead bring down the whole system to upgrade a component,

simple ways to maintain, I set up the appointment. So I'm buying a declarative deployment configuration, which describes how your status is should be connected together. Other than giving a step-by-step deployment recipe for you. Anyway, you know, that card companies that give me pictures of people and fireman full keeping everything else constant. Okay, so that's some of the ways that keeping a cheese and capabilities to applications on machine Learning System and easier to manage in general.

Let's talk about some continuous integration on continuous deployment. We've been fixed on his answering machine learning in containers for a long time which isn't surprising given that way, we start to see an increasing number of obstacles that. I just had to get started with containers and details of the infrastructure. So he's a container recipe is a yamel file. He's a Litany of shells, what time is the care about the infrastructure said just to get that jobs done,

27 x 8, Future chief of me as the dates of scientists that can I see falls out of the fox. It supposed to be lazy. So I like to work on my own laptop, but obviously I cannot get access to that. They say I'm excited by the speed of my laptop keeps turning my machine and heavy and hard. Wax celebration can Brady help up here, but it doesn't do anything for model training in storage and private. Can create a kubernetes meter or specialized computer cluster or any other essential infrastructure services without

having to file a ticket with it without having to take a corporate card to the public cloud. That we could you spell environment one thing. But we could do you spell Reese Hitches. Another and they just like to do a lot of work and notebooks and and that she is unable to button and see because that is dependent on my environment. Since I had any available for order, I run until 1. We want to let you know, be able to get around this time. Be able to effectively reproduce the

work that we do and share it with others. And is a really cool device. That makes notebooks reproducible. You think shipping? I see. So it take to get from the tree full of Jupiter notebook. It felt these into containers for these containers in a Cuban, nessie's custody. And this is at public hosted service that you can go and try it out right now. I'm in your patience and sorry, just by putting a notebook and get repaired. You can share it with anyone in the world and have it be totally reproducible. Now if you're using find it, you'll or do you think I'm deaf out practices this idea that we

can all the time? Repeated pull deployments again, and again is so beneficial. You know what, your status is X card and services. By looking at the open shift is Red Hat to the Naches platform and answers to image belt. Provide a framework to make services that know how to turn code into status. It is basically the idea is that we have a build, an image that knows how to turn to get repository of something without Souls, tired or whatever container image, taken by NASA's Sabbath on kubernetes as a day to sign

test about so that I can use it without being a packaging asks, so I'm going to use it for my collie. Graham made this knows how to take a nap on the tree of notebook and turn them into and notebooks that contain image. I need real wax myself and I think the best is the image. That Graham ACNL talk to take a repo of notebooks and turn them into a reproducible. Container is my repo of notebooks that I want to pass it, and had some super secret password and it takes in order to get it up and running and

then little man, just building applications on Notebook. 750i McD's. So I can be implemented by one of these. For example, the cells in Project has serialized object. Auto Services. Except we'll see that these buildings can do something much more involved. What I'm going to show you now is a supposed to look at the model, creating. List of the image of the tree that contains jupyter notebook, post processes that notebook to identify all of its dependency. And it doesn't post process is The Notebook to extract the clothes because notebook for a mix of call

me some model. 88, some rest endpoint. If that model, so it can go back to the black. So I was thinking about notebooks for each of these two stages with running them repeatedly on ending up with a model deployment technique for That I think engineering part of my mortal and fake social media. So we're trying to identify things. So that was a model training. I'm just trying to get logistic regression model and you can see that it just looks like a normal jupyter. Notebook.

Age of Wireless is running to build us Immortal status from that and those notebooks that we packed in the get repaired in a list of the notebooks name cuz there's a lot of nobody. That is up and running. This is a pretty simple and Powerful way to go from a notebook to a Model status that you can interact with. I'm actually going to interact with it through just another Jeep with a notebook. So you can see you can make predictions and not passing them to that model, said it. So I could give that exposed status address to do an application to

another day shift scientist and they can have a mess around. See what needs to be done before we can put it into production and production. We need someone else to that a foreman. So we know that Jason can drift time. We don't always have a way to get the ground troops full of the beacons directly track. A bottle of predictive performance of the time could publish. Spell Frank sample. What does the distribution of predictions? Look like? And if one of these changes substantially over time, we

visit that stamp application pipeline demo on see this in action. We have here, get metrics puncture in it, and you can see all the suspended Tyson metrics that are exposed like normal, but the ones that were interested in today at these custom pipeline prediction and Turtle Tracks. Did they tell you how many projections of each type fuel model has made? And we can actually just use the inside, open chest, a native of Prometheus series, official eyes and tool to have

luck, and visualize when they change. So we just met a couple of experiments that you can see that, you know, the Constitutional protections and see what happens and predicted performance metrics to compare these models. What is my name? Okay, so in that way. So, I hope I've convinced you that you should be running your machine any but those on keeping a seed and a Twist in today can be use by date. If I intend to make magma, repeat. And stay a couple of projects. I want to tell you about about before I wrap one, is the day to have which is an open cell projects out of

the Red Hats cheap flights out of the cube fly project get wet with that. Keeps our team verify keeps out on open chest and car to be fixed is that she doesn't have a look at that. And let me think of the applications developed by the old have different priorities. RV beneficial to each of the Rosetta Stone and project managing streaming daytime and pipeline for defining, the machine learning pipeline. Stop with that. What happened today while we started by an apple sitting on the machine. Anyway, sorry, just to remind ourselves how important

repeatability is. And what the fairy is too. Can you put turn off from that? We that we stole? That containers can be used to isolate processes and orchestrate, manage and distribute those components and into Applications that made all containerized microservices image bills. That are providing every flexible way to benefit from the web about having to become one yourself and find it. That it's what she should definitely check out if your agency and its components on top of a

princess. Should you do with that? I'm going to stop and break. But there is there's a question here in the chat room. Are bender and openshift interchangeable as in. They perform the same type of function for DS workflow? Okay, the question. Sorry, find it. It's just doing that one thing which is the replaceable notebook Bells. You can use to take a hit repost that contains a notebook and repeatedly been up an environment in which the south side of the platform has capabilities to support a

much bigger Saturday. So it manages many more components and all Spanish website in open chest. I hope that ice is your question, but I'm having serious talk offline more about that. If I can be helpful.

Cackle comments for the website

Buy this talk

Access to the talk “DevOps for Data Science With Kubernetes”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free

Ticket

Get access to all videos “MLconf Online 2020”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Ticket

Interested in topic “Artificial Intelligence and Machine Learning”?

You might be interested in videos from this event

February 4 - 5, 2021
Online
26
104
ai, application, bot, chatbot, conversation, data, design, healthcare, ml

Similar talks

Somnath Banerjee
Head of Shopping Discovery at Pinterest
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Jake Shermeyer
Research Scientist at IQT CosmiQ Works
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Dan Gifford
Senior Data Scientist at Getty Images
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free

Buy this video

Video
Access to the talk “DevOps for Data Science With Kubernetes”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free

Conference Cast

With ConferenceCast.tv, you get access to our library of the world's best conference talks.

Conference Cast
949 conferences
37757 speakers
14408 hours of content