Cloud Native Computing Foundation Kubernetes Community Days (KCD) Africa 2022, 24 Jul 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Product Recommendation system on Kubernetes using Argo Workflows by Osinachi Chukwujama

Description

DataOps is the process of automating the delivery of data from where it is produced to where it is analyzed. It mainly consists of three processes: Extract, Transform, and Load. These processes are automated using pipelines. A pipeline consists of extract, load, and transformation tasks that are connected using a scripting language, oftentimes Python. You can set up a pipeline using a tool like Apache Airflow, but you will have to trade off not using the awesomeness of Kubernetes. When you want to utilize the autoscaling advantage of Kubernetes, you may consider using Argo Workflows for your pipelines. Argo Workflows is a tool that allows you to set up pipelines that run on Kubernetes. You can set up a pipeline that extracts insight from IoT time-series data using components like Spark and Minio, deployed on Kubernetes. This talk is meant to educate Data Engineers and DevOps engineers looking to set up data pipelines on Kubernetes. It’ll go over a normal setup of a product recommendation system using Argo Workflows for setting up the pipeline, Minio for storing blob files, and Apache Spark on Kubernetes for running the analysis tasks.

---
KCD Africa 2022 is the 2nd iteration of the Kubernetes Community Days Africa, a CNCF-powered free community event. Visit https://kcdafrica.com for more information.

A

All right so um we're close to break time, and uh our next speaker is um right here. We, the next section, is going to be focused on product recommendation ecosystem on kubernetes, using aggro workflow, and this is going to be taken by osinatchi and um osinashi. Is a software engineer and technical writer building within the african tech ecosystem.

A

But now she enjoys building solutions on web 3 and also extending the web um extending the web using web assembly when ostinati isn't coding, he enjoys reading books good to have you here.

B

um Okay yeah. Thank you. Thank you, anytime, hello, everyone. um So I just I'll just get started immediately right.

C

Yes, yes, you can, can you try sharing your screen, let's see if it works? Okay,.

B

All right, my conditional sentence, um one minute.

C

We'll be going on a very short break while sinatra figures out the permission, issues.

C

A

A

A

Oh hey I'll snatch: it are you ready to go so. Are you good to go.

B

Yeah yeah thanks my screen all.

D

Right, let's get your screen. Yes, you can see that now.

B

B

I'm doing some background music. Is it okay.

A

B

Think you're good to go: we're ready all right so um hi everyone! um Today, I'm talking about product recommendation system on kubernetes, using algo, workflows, right and yeah just a bit about me. I love working with cloud native tools that pertain to data engineering and ai right. So let's get to product recommendation and you.

B

Maybe, when you're using a website and you're scrolling.

C

B

And you keep seeing things being recommended to your videos, maybe tick tock, I mean you keep seeing a new video being shown on your screen right. So there's an algorithm working working at the back to make this happen right. So it's we we have recommended stems and a lot of places I think most products, most modern products have one form of recommendation system on or another because because of the sheer amount of data present in this product, I'm generated by users with these big companies that have these products want to make.

B

They want to make the most money right, so they would want to recommend the product that user would likely, buy or use our likely watch, but something that the user wants. Personalizing experiences, basically all right so say here like on conga conga, is like a market.

B

It's like an e-commerce platform that is present in africa. I guess so, and you want to purchase a modem or my fight right. So if you scroll to the but the down you'd see similar products right, so the different algorithms are used to achieve this product recommendation.

B

We could do it strictly based on visual similarity, like you look at this one. If you look at this one, it's kind of similar to the one shown at the that's the first two products recommended right. So if we use, if we strictly and we base our recommendation strictly on virtual similarity, might not get the best results right, but usually they use a combination of a combination of of algorithms or combinational methods to achieve product recommendations.

B

I will won't be focusing on the query recommendation in this system in this talk right, so this talks about more about the pipeline that makes product recommendation possible right. So why kubernetes? Why not just do this product recommendation on a virtual machine somewhere so training and serving is hard at skill right, straining and selling machine learning models is hard at scale and I think that's the whole point of kubernetes to be able to scale your um containers or what to scale your yeah, your containers. Basically, when you need more more compute right.

B

So if you're running a large e-commerce business and say you're improving your product recommendation model every day, I you would want to keep running the model, maybe every day by 6, 8, 6 p.m. I don't know, and some days you might have more data coming into your app and others you have less so you want managers to control all this yeah. So what's this argo workflows thing so I'll go up. Close is a kubernetes and it is native pipeline orchestration.

B

So what's that, so it's all about pipelines right so, um but you, you see, pipelines a lot when we deal with real world applications and deployments and stuff like that, so even the normal way we develop applications, it's using a kind of ci cd way, even if you're not using fully using ci cd or githubs right. So if you commit some code, the code goes through several processes of building and testing until it's ready to be served to customers or the end users right. So it's a similar thing for machine learning pipelines.

B

So you your data, you want to get your data from just it from just data to a machine learning model that can be used in real world applications and it's all structured in a pipeline format.

B

Let's see it starts from you preparing your data, no matter where the data store realize the database or file storage, right, training, training, a machine learning model on that data, um testing the model packaging it deploying it and serving it up the whole the whole.

B

It's like a very similar pipeline to how you do it code, but in this case, machine learning models so um normally when you're developing a machine learning model right if you've worked with machine learning, if you work with data, scientists, you'd see them using um chiptan, notebooks or similar notebooks notebooks on collab cargo notebooks everywhere.

B

Basically, so these notebooks will help them develop the do their data cleaning and the whole thing that they do for making data of the ready and making their model fits the data all right, but now we're talking about them, moving away from these notebooks and moving to containers.

B

But it's kind of still it's just it's similar right, but in this case work with their notebooks and commits they are not looking to see the github repo and we run a pipeline that can take their code containerize it and put it in a pipeline right. So let me go back a bit, so I will work close. It helps you think in terms of containers.

B

You can use aggro clues for quite a number of things, not just machine learning pipelines. You can use it for data processing for cicd and a few other things, and so for machine learning pipelines. You'd like to process your data first extract features depending on the.

D

B

Learning process here isn't sure right. So you don't like all the steps that you do in, like the normal machine learning flow. You want to break it down into containers right and connect those containers together and if you do connect them together, you'll be able to and you'll be able to get your machine learning mode there already from start to end right. So um this is how you normally use other workflows. So you'd like you, deploy the aggro workflows on kubernetes, it's like a custom resource, it's like when it says custom resource right.

B

You deploy it on kubernetes and you write your workflow in yaml. You can also write it in python. They compile it down to jano. Then you submit your work to its kubernetes and clan is starts the workflow right, so we're going to see that in a bit. So like, let's go back to our product recommendation, I mentioned the steps earlier right, but let's just show you very simple, very simple model here right, so we this is something called a um it's just a pipeline right.

B

Some people call it dag some people, call it dag um direct, a cyclic graph. It's basically a graph, but this one is straight. Sometimes it's branches, so we have the stats. Then we have some pre-processing and we have training. Then we have maybe other things like testing that goes on in the register step right. So let me just do it of the demo and just start presenters.

B

I might just see audio. Am I still live confirmed.

B

Okay, just let's continue alright, so um here I have yeah. I have a pipeline pipelining ammo that contains um a workflow, it's an example workflow, and so I'm going to just quickly explain this so um I'll go close like I said earlier, it's uh sitting in yam, so it consists of something called templates and these templates are like right. So the main thing in algo workflows is templates and templates are like functions, and these functions are the bes at those steps in your pipeline.

B

So we have a pre-process. We have. First, we have a product, recommender function, it's like the entry points right and it's defined here then we have a preprocess step, the training step and you register step all right. So let's go to the entry point in the entry point we have something called a dag.

B

The dog is like it directed a cyclic graph right. So it's it's. Basically just the pipeline right and the this dog has tasks, and we have she tasks, they pre-process the train and to register. So I guess the main thing, um then other things are just like helping this one.

B

So this task um the the code or the instructions for it. That's the template is defined here and for the training. We also have the templates here and the register here right.

B

So I'm just going to go into a kubernetes cluster that I set up on azure. I I'm going to ssh into the vm my setup and I'll start meeting. You.

B

Oh, my goodness,.

B

So one minute you've been starting.

B

So otherworldly also has a.

B

Dashboard right here as a dashboard, I'm going to open it up in a bit. Just anything will load up in a few seconds.

B

B

Okay started right, so um I've really deployed ago workflows on this kubernetes environment. That's this mini tube plus stuff. I'm just going to get the code.

B

I deployed it on the namespace ago.

B

All right, so we we don't have it running, we don't have the server running. Yet it's actually we started.

B

So I think in a few minutes it's going to start fully that has to check you have to check regularly right so when he does that we can submit our workflow to it and to run our workflow and we do all the processing right. So I'm just going to go back to the code and explain it a bit more.

B

So when I say we do everything in terms of containers all right, so it means you're going to define all your processes in in containers. So I have here.

B

I have here the uh jupiter notes book that I did that try. I got help from my friend, so we did the machine learning um the whole data processing and model kitchen here right. So what I did was to copy the code from here and transfer it to these folders all right, so these folders contain um contain the files that will be packaged into containers.

B

So each of the folder has its own docker file right. So you'd set up an automated pipeline that would build the containers that would be containers for each of the folders or each of the steps right. So you good the containers then push it to your registry. Whatever registry, you want to push it or whatever registry you're using and when you do push it, they the workflow them. The workflow templates here is going to pull it's going to pull it. So here I push them to the car, push them to docker hub right.

B

So that's why you have this image here: you're, going to pull these images from the hook and run and run the containers? Basically, let's just check with a minute you've expected yeah. So it has it's running now, so I'm just going to go here and I'm going to s I'm going to um ssh into the server or connect the server to my local.

B

Here meant to, I meant to meant to put forward it's through the exports for workers.

B

So, for some weird reason only runs on https.

B

All right so this this um workflow I submitted earlier right so but normally when you're submitting a workflow you just maybe copy and paste or use an automated process. So let me open the different angles here.

B

I'm going to associate into this order.

B

Right so if we copy this code and like create the workflow, so let's.

B

Then we submitted, we just use qctl.

B

On the namespace.

B

You can watch, we can watch how the workflow runs right.

B

Let's copy this.

B

B

So it's running now and we're running into watch mode right basically to.

B

Run the first press and run the other products, so you can actually check the dashboard right. It's running the best pulling the images pre-processing the data and running the training mode and run the training container.

B

After that, one stream is going to run the register container.

B

Of code as the collision.

B

So I didn't actually connect the front-end interface right, but I found something very similar um done by a pd, and so the idea is like um we have these products and when we do click on a product, it's it's run through the model through the train model. This is our pipeline, which is basically for training right. We just like set up a pipeline for new friends when you click on the model. It's it's runs. It sends the image through it sends the image through.

B

The trained model, then gives you recommendations, and you like display the recommendations right. So that's like the basic idea of running such a pipeline on kubernetes and I'll. Go up feels like it's like it's like a very basic. It's like the basic pipeline.

B

It's like a very bad pipeline, medium on kubernetes, there's something even like him built on ago office. Basically, and that's queue flow right. So q3 is easily like more of an end-to-end way of running machine learning, pipelines or machine learning processes on kubernetes right, but it's still built on other workflows. So having knowing algorithms can be like a plus right, you can apply the knowledge you're using in say machine learning pipelines to do ci cd on kubernetes. So that's it basically right!

B

So um that's my talk. I guess we'll have go for questions now.

A

Awesome, thank you so much for that section. If you have any questions, okay, you can drop it in the chat and well it's available to actually take your answers.

A

All right, it seems like we don't have any questions for now in the meantime.

A

Think of any questions you can always refer back to the social media handle um please drop it in the youtube channel, so we can always reach out to you yeah. You can always refer to um or synapses to show me the handle and um ask your questions on there, and you can always refer back to this recording if you're just jumping in.

A

I think this is uh the last session before break hooray who's. So excited yes, so thank you so much asanashi for that um awesome section. It was good to have you here.