AWS Compute for Data Science: EC2 vs. SageMaker vs. Lambda Explained

Key Takeaways

Hello everyone. I welcome you all as a part of this particular video lecture where we are going to explore compute for data science. And this video lecture is based out on EC2 versus Sage Maker versus Lambda for data science task. Now before I'm going to get deep dive into this video lecture, let me put one statement in front of you and that is servers are everywhere. Means whatever task we are doing, whatever things we are performing, we will be needing server or we can say that we will be needing compute power. Now our journey is to provision the server where we are looking for more control to the flexibility where AWS or any cloud provider is going to manage the things for you and in this journey only we will be discussing EC2 SageMaker and Lambda. Now before you are going to talk about these services there are two pointers you always need to understand what and why means what is EC2 sage maker or lambda and second thread which you have to talk about that is why EC2 sage maker and lambda and if you want to understand in a more detailed manner you can also get an answer to one more question that is where EC2 sage maker and lambda now best place to start about any AWS service is the AWS management console or the AWS documentation. So I'm assuming that till the time all of you would be having an active AWS account and if you all have an active AWS account. So without any further delay let's move towards your AWS account. So I'm navigating to my AWS account. This is a landing page everybody and you know how can you search for a particular service in the AWS search bar. So either you can type the name of the service like I'm typing here and when you are typing the name you will be getting a very short description about a service let's say virtual servers in the cloud [snorts] means this service is helping you out in order to provide guest machines or virtual machines in the cloud and the best part is it will also be giving you some top features in a similar manner you can search for lambda here also you can see run code without thinking about servers means we will be putting Lambda into serverless categories and finally you can search for Sage Maker. Sage maker is the center for data analytics and AI. Also, if you want to put the service into specific category, you can explore the categories from here. Let's say you can explore the category called analytics. In a similar manner, you can explore the category called compute. So, Sage Maker comes under analytics category. From the definition I have get an idea into. In a similar manner, if you will be going to compute, you will be seeing EC2 and lambda. Okay. Now, this was a short description about Lambda, EC2 and SageMaker. Let me give you a quick overview one more time. So, EC2 stands for elastic compute cloud which gives you virtual servers you manage completely. So here I have put one pointer in front of you that is management is on you. When management is on you it means control is with you. Okay. So here our choice is control. Our choice is management means you can choose the instance type. You can decide what is a specification of your virtual machine you need. You can choose the operating system. You can choose the libraries or other prerequisite. So EC2 is best for full control custom environments longunning or high performance compute workload. Now remember these keywords the highlevel keywords which I have presented in front of you and if you are aware about these keywords you will be able to relate that for this particular use case this service is a best fit. In a similar manner when we talk about lambda a simple definition for lambda you can give lambda provides you eventdriven serverless compute option. Now serverless means that you don't have to gain visibility into underlying infrastructure or you don't have to worry about underlying infrastructure. Just focus on your code and AWS is going to take care about rest of the things for you. Means you can say lambda provides you serverless functions that run code in response to events with short execution time. So what lambda is best for? So lambda is best for lightweight eventdriven data processing or inference. And the last service I will be talking about that is Sage Maker. So SageMaker is a fully managed service for building, training and deploying machine learning models. So Sage Maker is providing you a best fit for end to end ML workflows where you are talking about data preparation, training, deployment with less operations overhead. So what journey we are talking about here? So we're talking about our journey starting from control to the less management to the more flexibility means here we are talking about these services in terms of what these services are best fit for. Okay. Now whenever you are planning to use any service you always need to understand certain pointers. Let's say if I'm talking about compute model and cost. So in terms of compute control if I'm just taking this as a feature in EC2 you will be having a full control means you can choose the things based on your need based on your choice but always remember when you're talking about full control the management would be more at your side. Okay. So accordingly you have to decide accordingly you have to make a choice. In SageMaker as it is a managed service you can choose the instance type but AWS is going to handle the setup and scaling and for lambda there is no servers means AWS will be allocating the runtime automatically underlying infrastructure would be there when I'm using a word no servers it means you don't have visibility into the underlying infrastructure like how AWS is provisioning the resources for you where the actual execution will happen. Okay, always understand the billing. That is a very important part because even if you are from a development background, okay, at a certain point of time, you should be having a little bit understanding on pricing. Let's say if you're a part of a call with your stakeholders, then you need to present that why you are using this particular service. So if I talk about a billing model, EC2 follows a billing model called pay per second. Means while instance is running, you have to pay the charges per second. If I talk about Sage Maker, Sage Maker follows a billing model where you are paying for a time used for training inference and also the idle notebooks cost money. For Lambda, you are paying per request and compute time. In a similar manner if I'm talking about scalability so scalability in case of EC2 can be manual or you can also scale your infra via autoscaling group. Sage maker provides you built-in autoscaling for training and inference means you can choose the instance type. You have that much of flexibility but setup will be taken care by AWS for you and also it will be the same thing with the scaling part. Lambda provides you automatic scaling for event trigger. Okay. So you have got a basic understanding on these services like what are the use cases where these services are best fit for. Now understand the use cases for data science. If I'm just putting a subject in front of you, we have to understand the use cases from that subject perspective. Let's say if I am talking about different options in our data journey. The first option where we are talking about data exploration and notebook work. Okay. If you have to explore the best AWS options or best AWS services for this particular use case, you can say that use a combination of Sage Maker Studio on EC2 plus Jupiter. Okay. Means Sage Maker gives manage notebook. EC2 gives you flexibility. If you're talking about feature engineering or ETL, extract transform load. Lambda you can use for eventdriven small jobs. EC2 and SageMaker you can use for processing jobs. If you are talking about model training stage, there you can make use of Sage Maker training jobs or EC2 custom. Means Sage Maker handles the distributed training hyperparameter tuning and spot instances management. If you're talking about the next stage which is model deployment, you can make use of SageMaker endpoint, Lambda or EC2. In a similar manner for batch predictions, you can make use of Sage Maker batch transform or EC2. And finally if you are having some eventdriven ML means let's say if you're talking about realtime classification lambda is a best choice for you. So this is something where we are talking about the use cases for data science. Don't worry soon I will be presenting some example scenarios in front of you. Okay. But before that the next thing which we are going to talk about that is maintenance and management. Okay. This is a very important part although we have already got a very good understanding with respect to control and flexibility but let's understand different aspect for example if I'm talking about infra management everybody in case of EC2 you manage everything manage everything means on top of virtual machines in case of sage maker AWS manage almost everything And in case of lambda you don't have to worry about underlying infrastructure at all. Second thing if I talk about that is environment setup that is the next aspect which you have to understand. So in case of EC2 environment setup is manual in case of sage maker you have pre-built environments like you have tensorflow you can have py talk okay you can have skarn in case of lambda it's lightweight means you can bring your code plus dependencies if I talk about monitoring and logging which is a very important aspect for any particular service so in case of EC2 you have cloudatch but you need to do a manual setup. In case of SageMaker, you have integration with Cloudatch and also you have a feature of SageMaker called SageMaker Studio. In Lambda, you can make use of the Cloudatch logs means you can do the monitoring automatically with the help of Cloudatch logs. Okay. Now, let me give you some example scenario to test your understanding. Okay. And when I'm talking about testing your understanding, I will be checking that how can you relate that for this example scenario which service is a best fit. So my first example scenario for all of you is you need a GPU instance to train a large deep learning model with custom setup. Okay, I hope you have listened to the question very carefully. I can again put that question in front of you. You need a GPU instance to train a large deep learning model with the custom setup. Now question itself is pointing to certain keywords. We are looking for a custom setup. We are looking for a GPU based instance. So answer to this scenario would be easy to everybody. Okay. In a similar manner, let me put few more scenarios in front of you. Like second one is you want to quickly train and deploy a model without managing servers. So here we are talking about deploying and training a model and even we don't want to manage the servers. We're looking for flexibility. So we know Sage Maker is a best choice for this kind of use cases. Okay, you want to trigger predictions when a new file lands in S3. So here we are talking about eventdriven use cases. Always try to understand the keywords from the scenario itself which can give you a right choice for a service. So as I told you in my this particular scenario we're talking about eventdriven. So answer would be lambda. Okay. Next we have you're building a data science pipeline that includes data prep-processing training and deployment. Okay, so here we are talking about building a pipeline. So you can make use of Sage Maker pipeline. Okay, that is a feature of Sage Maker only. Next, if I'm talking about you need fine grained control and custom networking whenever we're talking about customization. So here customization is related to control. If you will be having control then you would be able to do the customization. So here answer would be EC2. Okay. So this was a quick introduction where I wanted to give you an overview on EC2, SageMaker and Lambda. And I wanted to make a quick comparison that where EC2 is a best fit, where Lambda is a best fit and where Sage Maker is a best fit. Okay. So EC2 is best for custom machine learning environment. Sage Maker is best for end to end machine learning pipelines and Lambda is a good choice for eventdriven or realtime influence. Okay. And we have compared all these services on different factors. Let's say if I'm talking about manage, EC2 is not managed. On the other side, Lambda and Sage Makers are managed. If I'm talking about longunning jobs, EC2 and Sage Maker are good fit for longunning jobs where Lambda is not a good fit for longunning jobs due to the limitation of Lambda of 15 minutes execution time. Okay. Training support in EC2 the training support would be custom in Sage Maker it is builtin as SageMaker is designed for this for Lambda it is limited. If I talk about inference means API serving. Okay. EC2 supports this. Sage maker again supports this. Lambda is a good choice if you are dealing with short jobs. Scalability again in EC2 it is manual. If you have to make it automatically you have to use autoscaling. Sage maker and lambda provides you auto scalability. Cost efficiency. EC2 is good if you are optimizing it. Sage maker is medium and lambda is excellent for short jobs. Okay. So recommendation is beginners and managed machine learning workflow make use of SageMaker. Advanced custom environment or cost optimization make use of EC2. Eventdriven lightweight machine learning or ETL make use of Lambda. So this is something I wanted to help you out as a part of this particular video lecture. Soon we are going to talk about more on this particular track and we will be seeing some hands-on exercises.