# What is MLOps? | MLOps Explained for Beginners | DevOps vs MLOps | Edureka Live

## Метаданные

- **Канал:** edureka!
- **YouTube:** https://www.youtube.com/watch?v=LMOKAaK06nY
- **Дата:** 06.05.2026
- **Длительность:** 35:41
- **Просмотры:** 694
- **Источник:** https://ekstraktznaniy.ru/video/50516

## Описание

🔥Integrated MS+PGP Program in Data Science & AI: https://www.edureka.co/dual-certification-programs/ms-data-science-pgp-gen-ai-ml-birchwood
🔥MLOps Certification Training Course : https://www.edureka.co/mlops-certification-training-course

In this Edureka video on “Understanding MLOps” , you will learn about the Machine Learning Operations , I aim to show you what are the most trending Machine Learning Operations and how it is going to change our world. What do you think about Mlops? Leave us a comment and let us know what you think of this video. We want to hear your thoughts.

📝Feel free to share your comments below.📝

𝐄𝐝𝐮𝐫𝐞𝐤𝐚 𝐎𝐧𝐥𝐢𝐧𝐞 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐚𝐧𝐝 𝐂𝐞𝐫𝐭𝐢𝐟𝐢𝐜𝐚𝐭𝐢𝐨𝐧𝐬

🔵 DevOps Online Training: http://bit.ly/3VkBRUT 
🌕 AWS Online Training: http://bit.ly/3ADYwDY
🔵 React Online Training: http://bit.ly/3Vc4yDw
🌕 Tableau Online Training: http://bit.ly/3guTe6J
🔵 Power BI Online Training: http://bit.ly/3VntjMY
🌕 Selenium Online Training: http://bit.ly/3EVDtis
🔵 PMP Online Training: http://bit.ly/3

## Транскрипт

### Segment 1 (00:00 - 05:00) []

Good evening all. MLOps basically covers those particular process where we start with post deployment activity, starting from deploying the model to to doing health checks, setting up parameter to some we can check model drift, model performance, function quality, overall trend based on different data sources, everything gets covered under M MLOps, all right? So, in this particular course or in this particular webinar, we will talk about you know, majorly the important points that generally uh we track as a part of this activity. In a different scenario, what happens is basically there would be multiple developers that work on a particular project, right? So, this is project is never a one-man show. So, as a part what we try to do, bring everyone together. What has been the develop ment done by, you know, different who are working on the project and how we can set up orchestrated pipelines, which will help us to take care of, you know, some pre-governance measures and mostly post-deployment steps. So, how we can organize organize and orchestrate all of them. In order to achieve this, we will use a number of tools, frameworks, and currently available for different solutions. Okay? We will start the exercises. All right? Generally, an of team is going to contain different resources, right? So, we'll have a guy who's running have things in the cloud knowledge, right? Someone who's doing right? A multi multiple resources working together in object. Particularly from we would need some someone who had IT of the which is what's in learning process, which with the engineer and which model What is the What are the different factor in modeling? Those information and along with that, they would also get to understand statistical decision points that we have considered. For example, gone with ROC curve, gone with confusion matrix. Did we go with F1 score, right? So, the logic is going for some machine learning, right? So, this will make the deployment when it comes to accuracy and depending learning model. Generally, we'd be using all the machine learning solutions that we will use. They will have the building solutions. So, accuracy building might include factors for new test R2 score, adjusted R2 score, plain variance, mean squared error, so on and so forth. This is going to include recall, right? All of this. Searching the model performance, we would mix with this become like how good the model. Third, make sure time the solution to hit us with coming. They generally which way solution or which way building it takes to solution ready or to deploy the entire solution a very less. The fourth is to make cost factor. The deployment environment, the system specification, RAM. There are different you know, like parameters which we generally consider whenever we go for model deployment server selection. Those have been con- have been considered wisely, and those are considered to the best of the environment or use case required. Okay? So, these are some of the things which is expected from an MLOps individual to formulate expected roles. Okay? I'll proceed forward. If you can see this diagram over here, in this particular case, in ML, basically we will use dev, we will use ops, and it will use primarily we use machine learning. So, when we talk about machine learning means machine learning will include the data science aspect where we are doing engineering around the data plus the modeling side, which is basically the machine learning models available. Dev could be product where we are integrating it, or it could be the page where we are ingesting it as a feature or as a solution, and ops is going to be the post-deployment artifacts. So, MLOps is going to integrate all three together, which is for the data science machine learning by the dev, okay? And followed by the operational side, which is the post- deployment artifacts. Right?

### Segment 2 (05:00 - 10:00) [5:00]

Right? This three combine and together forms M- MLOps. What we are going to to see next step is how MLOps or machine is different from DevOps. Okay? These are some of the differences that we have highlighted over here. Focus perspective. When we talk about MLOps, we are going to create this ML system. Okay? We will not development backups. Ops operated in good solutions. IT could be built for a back-end based application, right? Both of them can be governed very easily through DevOps. And they in DevOps, it can use machine learning, it can also do without machine learning as well. Okay? The machine learning part of DevOps is known as MLOps. All right? In terms of domain knowledge, which is point two, in order to understand MLOps, you need to know data science, machine learning, and you also need to know on what is the operational side of machine learning models, okay, which would be from cloud perspective, like AWS may have different tech stack, Azure may have different, we may have different, right? AWS we're using, you know, different environment. Likewise, all the clouds operate in a similar nature, but the functional names are different. In DevOps side, you need the application knowledge around the infrastructure management, okay? For example, whenever you're doing DevOps, if you want to want CI/CD pipeline, do Docker. Most specified RAM is 4 GB. You're 1 GB, in that case, you will not be able to do with it. So, you need to have around the which is the data capture and just say what is the minimum in the solution. The third difference is along the infrastructure lines, okay? When we expand DevOps to include ML feature, it becomes MLOps, as I clearly mentioned on the focus, which is the point one itself, that extending DevOps to handle ML based infra is basically making it as MLOps, okay? When it comes to infrastructure DevOps, it is going to focus on only those platforms which is required for delivering the application. Now, let me give you an example. See, MLOps and DevOps have certain framework. For example, DevOps has we have Git, it can integrate the code from Git, it can then connect it through Jenkins job, and then it can push to any environment, all right? So, these are the platform is basically the environment, and we are using different sources in order to connect and bring different data. When it comes to ML Ops, basically what it does, it expands this DevOps framework, but it will only include those tools and artifacts which have been considered for building the ML app. Okay? Fourth is data management. Whenever we are using data, we need ML Ops there so that we can orchestrate the data in a proper manner. For example, what do I mean using that? Let's say this is your solution. Okay? This is the solution where you have written the code. In this particular code, there are different pre-applications. Let's say this is S3 bucket or Azure blob storage. This is the pre-processing code. This is the ML code. Let's say data. This is pre-processing. Okay? ML. Then we are doing prediction. And then finally output. So, what do we need? We can integrate the entire So, whenever you're telling your data management, we should know at which point the data needs to tested. Okay? What are the data the health checks that needs As a part of prediction, what testing tools that we need? Okay? Everything is treated as a part of ML Ops. Okay? So, we can or we can using which external data that Similarly, along the DevOps, here in this case, it does not know the data you have orchestrated. It can It has basically has the amenities. So, it would be in a position to and fetch it. At different position to ML Ops, we will need to train it depending on different situations. So, we might have to train, retrain, we might have to do hyperparameter tuning in order to handle different scenarios. But when it comes to DevOps, it basically does not require any sort of model training and model development. It generally works on multiple tools and multiple ingestion mechanism through which the application is comes to monitoring and feedback. Generally, ML models are equipped with certain tools through which we can monitor and check if the model is doing fine or not. Along DevOps, we have a different set of tools through which we can check the server performance, the job performance, how much time it is taking to process the data, validate the information, right? All of

### Segment 3 (10:00 - 15:00) [10:00]

those can be done using different set of tools. So, to sum it up, MLOps mainly focuses along the data science machine learning side of things. DevOps focuses on anything apart from ML. When it comes to domain knowledge, here you need to understand data science machine learning. This needs only infrastructure management related knowledge through which you would be able to set up RAM and do different exercises. Okay? When it comes to data management, here MLOps is a very important part because through MLOps you can ingest data or you can connect data at different points. For example, for at prediction point you need test data set. At ingesting part you need training data set. In pre-processing thing you need, you know, some logic that we have written separately. So, data management is very important component from MLOps perspective. Those data sources is very much required. Whereas in DevOps, it is not required. When it comes to model development, we need to do model training, hyperparameter tuning, model testing, different things. But in DevOps, it does not require any of those. When it comes to monitoring and feedback, we use different set of tools. Whenever it comes to DevOps related, we use we are getting same result. For DevOps, we will use different set of tools. Okay? Now, let's try to understand how the MLOps maturity is working over here. Generally, there is multiple processes. What is the first process? Data scientists is going to handle a manual workflow. Okay? After that, what do you see? There is a section for managing and version pipelines, where you're going to set up a particular version, you know, data source, connect it to another particular source, sub source through which the uh you know, processing is happening. The third is automating ML pipelines. Okay? The fourth is continuous integration, fifth is monitoring and feedback. Now, let me explain how this is going to happen. Okay, so when you say data scientist handle workflow, which means data scientist has some predefined conditions or predefined logic which has been given by the develop or as a part of the requirement and using that we kind of set up the manual process or like the data scientist has initiated the manual code generation. Okay, now once the code is generated, the next thing that we are going to do is set up the code version and set up a pipeline. For example, pipeline is going to indicate what is the first check that needs to be validated followed by second, followed by third and when can we say it is completed. So, we will build this pipeline one by one. After that, what did I mention? Automating ML pipelines, which is a single click framework. Okay, so here the moment you click on a particular framework, what it is going to do? So, frameworks are of different types. We can use Flask, we can use Streamlit, we can use Gradio. These are different frameworks, but in terms of deployment or in terms of ML Ops, we will mostly use Flask. In some cases, Django can also be used, but Flask is most preferred. Okay. So, using automated ML pipeline, we are going to create a script, a main script that is going to combine all the sub scripts logics and bring them as one. Okay, so this is going to automate the ML pipeline. After this, what we'll do is we'll set up continuous integration. Once the pipeline is created, automated pipeline is running, we need to integrate it to our own application, correct? So, in that case, training, testing, deployment, right? Or you know, validation test, all of this runs that we will do based on this ML pipeline, it is going to be ingested to the application. And finally, once the pipeline starting from data science code of, you know, reading data, processing data, building pipeline and training model, checking numbers. So, what we'll do is finally monitoring and feedback. So, whenever the model is running on real-time data, we will try to see if the model is performing well, giving us the right prediction or not, okay? Or if some not performing well on unseen data, okay? So, likewise, we can do those parameters set up checks. For example, checks like model drift, checks like processing time of model, checks like training time of model, how much time the job is taking to make end-to-end completion. So, different health-related parameters can be set up so that we can analyze the data from the model and go ahead with our analysis. After this, we're going to talk about some of the challenges in ML Ops. So, when we talk about ML Ops challenges, the first thing that happens in data quality. So, if the quality of the data is good, like any other machine learning process, okay? If the input data is of good quality, the output data is automatically going to be of a similar quality. But, the input data, if it is garbage, then the output data is also going to be garbage, right? So, this is something which we generally follow. So, here we have to make sure that the data is thoroughly processed. The data is, you know, in sync with what the use case is. We have the data distribution done across multiple samples properly, so that we can do the model training properly, okay? So, this data quality is going to connect us to the model training, all right? So, in model training also, we get some challenges, all right? Similarly, let's say data quality is okay, but in model training, we do not have good number of use case examples. What happens is, let's say we want to test if a particular customer is going to do bank fraud or not, okay? Now, you have a data set of 1 million records where only one customer or five

### Segment 4 (15:00 - 20:00) [15:00]

customers has done has not repaid the loan or has done a fraud, okay? Now, would that data set be good for model training? Of course, not. We need a good balance of data samples where customers have done fraud, have customers have not made payment to that of, you know, who has made payment. Then only the model training gets done in a proper manner. So, what happens is model training we have to make sure that the data quality and the data variability is maintained. If not maintained, then model training will not be as successful as we plan it to be. In terms of security, we need to have sufficient mechanisms within the code so that our solution is not accessible to anyone and everyone. We can set up some access related roles which can be done as a part of access management through IAM console as well. Like who can use this application, who are the rightful persons who has access to this. So, likewise they can check, right? So, through that we can maintain some security related checks. In terms of culture, what we can additionally do is to make sure that the data that we are solving in is you know, done is analyzed properly or is belonging to the right set of audience, right? Or is maintaining the consistency that is required from the use case perspective. When we go towards model deployment, there are many challenges. Which cloud to choose? Shall we do it in local host? What is the trade-off? What is the cost if we choose a 4GB RAM instead of 2GB? Or instead of 8GB RAM I choose 16GB? If I choose a GPU instead of CPU, right? There are many questions which basically ends up confusing many people because we have to factor in multiple costs, multiple things. So, all of those gets designed, taken care as a part of model deployment related process. Thank you for your time. — Mhm.

### Segment 5 (20:00 - 25:00) [20:00]

— Mhm.

### Segment 6 (25:00 - 30:00) [25:00]

— Mhm.

### Segment 7 (30:00 - 35:00) [30:00]

— Mhm.