Episode #96: Serverless and Machine Learning with Alexandra Abbas

April 12, 2021 • 44 minutes

On this episode, Jeremy chats with Alexandra Abbas about why we need MLOps, how teams productionize workflows and deploy models, the challenges for serverless machine learning use cases, and much more.

Watch this episode on YouTube:

About Alexa Abbas

Alexandra Abbas is a Google Cloud Certified Data Engineer & Architect and Apache Airflow Contributor. She currently works as a Machine Learning Engineer at Wise. She has experience with large-scale data science and engineering projects. She spends her time building data pipelines using Apache Airflow and Apache Beam and creating production-ready Machine Learning pipelines with Tensorflow.


Alexandra was a speaker at Serverless Days London 2019 and presented at the Tensorflow London meetup.


Personal links

Twitter: https://twitter.com/alexandraabbas
LinkedIn: https://www.linkedin.com/in/alexandraabbas
GitHub: https://github.com/alexandraabbas

datastack.tv's links
Web: https://datastack.tv
Twitter: https://twitter.com/datastacktv
YouTube: https://www.youtube.com/c/datastacktv
LinkedIn: https://www.linkedin.com/company/datastacktv
GitHub: https://github.com/datastacktv
Link to the Data Engineer Roadmap: https://github.com/datastacktv/data-engineer-roadmap


This episode is sponsored by CBT Nuggets: cbtnuggets.com/serverless and
Stackery:
https://www.stackery.io/

Watch this video on YouTube: https://youtu.be/SLJZPwfRLb8

Transcript

Jeremy: Hi, everyone. I'm Jeremy Daly, and this is Serverless Chats. Today I'm joined by Alexa Abbas. Hey, Alexa, thanks for joining me.

Alexa: Hey, everyone. Thanks for having me.

Jeremy: So you are a machine learning engineer at Wise and also the founder of datastack.tv. So I'd love it if you could tell the listeners a little bit about your background and what you do at Wise and what datastack.tv is all about.

Alexa: Yeah. So as you said, I'm a machine learning engineer at Wise. So Wise is an international money transfer service. We are aiming for very transparent fees and very low fees compared to banks. So at Wise, basically, designing, maintaining, and developing the machine learning platform, which serves data scientists and analysts, so they can train their models and deploy their models, easily.

Datastack.tv is, basically, it's a video service or a video platform for data engineers. So we create bite-sized videos, educational videos, for data engineers. We mostly cover open source topics, because we noticed that some of the open source tools in the data engineering world are quite underserved in terms of educational content. So we create videos about those.

Jeremy: Awesome. And then, what about your background?

Alexa: So I actually worked as a data engineer and machine learning engineer, so I've always been a data engineer or machine learning engineer in terms of roles. I also worked, for a small amount of time, I worked as a data scientist as well. In terms of education, I did a big data engineering Master's, but actually my Bachelor is economics, so quite a mix.

Jeremy: Well, it's always good to have a ton of experience and that diverse perspective. Well, listen, I'm super excited to have you here, because machine learning is one of those things where it probably is more of a buzzword, I think, to a lot of people where every startup puts it in their pitch deck, like, "Oh, we're doing machine learning and artificial intelligence ..." stuff like that. But I think it's important to understand, one, what exactly it is, because I think there's a huge confusion there in terms of what we think of as machine learning, and maybe we think it's more advanced than it is sometimes, as I think there's lower versions of machine learning that can be very helpful.

And obviously, this being a serverless podcast, I've heard you speak a number of times about the work that you've done with machine learning and some experiments you've done with serverless there. So I'd love to just pick your brain about that and just see if we can educate the users here on what exactly machine learning is, how people are using it, and where it fits in with serverless and some of the use cases and things like that. So first of all, I think one of the important things to start with anyways is this idea of MLOps. So can you explain what MLOps is?

Alexa: Yeah, sure. So really short, MLOps is DevOps for machine learning. So I guess the traditional software engineering projects, you have a streamlined process you can release, really often, really quickly, because you already have all these best practices that all these traditional software engineering projects implement. Machine learning, this is still in a quite early stage and MLOps is in a quite early stage. But what we try to do in MLOps is we try to streamline machine learning projects, as well as traditional software engineering projects are streamlined. So data scientists can train models really easily, and they can release models really frequently and really easily into production. So MLOps is all about streamlining the whole data science workflow, basically.

And I guess it's good to understand what the data science workflow is. So I talk a bit about that as well. So before actually starting any machine learning project, the first phase is an experimentation phase. It's a really iterative process when data scientists are looking at the data, they are trying to find features and they are also training many different models; they are doing architecture search, trying different architecture, trying different hyperparameter settings with those models. So it's a really iterative process of trying many models, many features.

And then by the end, they probably find a model that they like and that hit the benchmark that they were looking for, and then they are ready to release that model into production. And this usually looks like ... so sometimes they use shadow models, in the beginning, to check if the results are as expected in production as well, and then they actually release into production. So basically MLOps tries to create the infrastructure and the processes that streamline this whole process, the whole life cycle.

Jeremy: Right. So the question I have is, so if you're an ML engineer or you're working on these models and you're going through these iterations and stuff, so now you have this, you're ready to release it to production, so why do you need something like an MLOps pipeline? Why can't you just move that into production? Where's the barrier?

Alexa: Well, I guess ... I mean, to be honest, the thing is there shouldn't be a barrier. Right now, that's the whole goal of MLOps. They shouldn't feel that they need to do any manual model artifact copying or anything like that. They just, I don't know, press a button and they can release to production. So that's what MLOps is about really and we can version models, we can version the data, things like that. And we can create reproducible experiments. So I guess right now, I think many bits in this whole lifecycle is really manual, and that could be automated. For example, releasing to production, sometimes it's a manual thing. You just copy a model artifact to a production bucket or whatever. So sometimes we would like to automate all these things.

Jeremy: Which makes a lot of sense. So then, in terms of actually implementing this stuff, because we hear all the time about CI/CD. If we're talking about DevOps, we know that there's all these tools that are being built and services that are being launched that allow us to quickly move code through some process and get into production. So are there similar tools for deploying models and things like that?

Alexa: Well, I think this space is quite crowded. It's getting more and more crowded. I think there are many ... So there are the cloud providers, who are trying to create tools that help these processes, and there are also many third-party platforms that are trying to create the ML platform that everybody uses. So I think there is no go-to thing that everybody uses, so I think there is many tools that we can use.

Some examples, for example, TensorFlow is a really popular machine learning library, But TensorFlow, they created a package on top of TensorFlow, which is called TFX, TensorFlow Extended, which is exactly for streamlining this process and serving models easily, So I would say it TFX is a really good example. There is Kubeflow, which is a machine learning toolkit for Kubernetes. I think there are many custom implementations in-house in many companies, they create their own machine learning platforms, their own model serving API, things like that. And like the cloud providers on AWS, we have SageMaker. They are trying to cover many parts of the tech science lifecycle. And on Google Cloud, we have AI Platform, which is really similar to SageMaker.

Jeremy: Right. And what are you doing at Wise? Are you using one of those tools? Are you building something custom?

Alexa: Yeah, it's a mix actually. We have some custom bits. We have a custom API, serving API, for serving models. But for model training, we are using many things. We are using SageMaker, Notebooks. And we are also experimenting with SageMaker endpoints, which are actually serverless model serving endpoints. And we are also using EMR for model training and data preparation, so some Spark-based things, a bit more traditional type of model training. So it's quite a mix.

Jeremy: Right. Right. So I am not well-versed in machine learning. I know just enough to be dangerous. And so I think that what would be really interesting, at least for me, and hopefully be interesting to listeners as well, is just talk about some of these standard tools. So you mentioned things like TensorFlow and then Kubeflow, which I guess is that end-to-end piece of it, but if you're ... Just how do you start? How do you go from, I guess, building and training a model to then productizing it and getting that out? What's that whole workflow look like?

Alexa: So, actually, the data science workflow I mentioned, the first bit is that experimentation, which is really iterative, really free, so you just try to find a good model. And then, when you found a good model architecture and you know that you are going to receive new data, let's say, I don't know, I have a day, or whatever, I have a week, then you need to build out a retraining pipeline. And that is, I think, what the productionization of a model really means, that you can build a retraining pipeline, which can automatically pick up new data and then prepare that new data, retrain the model on that data, and release that model into production automatically. So I think that means productionization really.

Jeremy: Right. Yeah. And so by being able to build and train a model and then having that process where you're getting that feedback back in, is that something where you're just taking that data and assuming that that is right and fits in the model or is there an ongoing testing process? Is there supervised learning? I know that's a buzzword. I'm not even sure what it means. But those ... I mean, what types of things go into that retraining of the models? Is it something that is just automatic or is it something where you need constant, babysitting's probably the wrong word, but somebody to be monitoring that on a regular basis?

Alexa: So monitoring is definitely necessary, especially, I think when you trained your model and you shouldn't release automatically in production just because you've trained a new data. I mentioned this shadow model thing a bit. Usually, after you retrained the model and this retraining pipeline, then you release that model into shadow mode; and then you will serve that model in parallel to your actual product production model, and then you will check the results from your new model against your production model. And that's a manual thing, you need to ... or maybe you can automate it as well, actually. So if it performs like ... If it is comparable with your production model or if it's even better, then you will replace it.

And also, in terms of the data quality in the beginning, you should definitely monitor that. And I think that's quite custom, really depends on what kind of data you work with. So it's really important to test your data. I mean, there are many ... This space is also quite crowded. There are many tools that you can use to monitor your distribution of your data and see that the new data is actually corresponds to your already existing data set. So there are many bits that you can monitor in this whole retraining pipeline, and you should monitor.

Jeremy: Right. Yeah. And so, I think of some machine learning like use cases of like sentiment analysis, for example... looking at tweets or looking at customer service conversations and trying to rate those things. So when you say monitoring or running them against a shadow model, is that something where ... I mean, how do you gauge what's better, right? if you've got a shadow... I mean, what's the success metric there as to say X number were classified as positive versus negative sentiment? Is that something that requires human review or some sampling for you to kind of figure out the quality of the success of those models?

Alexa: Yeah. So actually, I think that really depends on the use case. For example, when you are trying to catch fraudsters, your false positive rate and true positive rate, these are really important. If your true positive rate is higher that means, oh, you are catching more fraudsters. But let's say your new model, with your model, also the false positive rate is higher, which means that you are catching more people who are actually not fraudsters, but you have more work because I guess that's a manual process to actually check those people. So I think it really depends on the use case.

Jeremy: Right. Right. And you also said that the markets a little bit flooded and, I mean, I know of SageMaker and then, of course, there's all these tools like, what's it called, Recognition, a bunch of things at AWS, and then Google has a whole bunch of the Vision API and some of these things and Watson's Natural Language Processing over at IBM and some of these things. So there's all these different tools that are just available via an API, which is super simple and great for people like me that don't want to get into building TensorFlow models and things like that. So is there an advantage to building your own models beyond those things, or are we getting to a point where with things like ... I mean, again, I know SageMaker has a whole library of models that are already built for you and things like that. So are we getting to a point where some of these models are just good enough off the shelf or do we really still need ... And I know there are probably some custom things. But do we still really need to be building our own models around that stuff?

Alexa: So to be honest, I think most of the data scientists, they are using off-the-shelf models, maybe not the serverless API type of models that Google has, but just off-the-shelf TensorFlow models or SageMaker, they have these built-in containers for some really popular model architectures like XGBoost, and I think most of the people they don't tweak these, I mean, as far as I know. I think they just use them out of the box, and they really try to tweak the data instead, the data that they have, and try to have these off-the-shelf models with higher and higher quality data.

Jeremy: So shape the data to fit the model as opposed to the model to fit the data.

Alexa: Yeah, exactly. Yeah. So you don't actually have to know ... You don't have to know how those models work exactly. As long as you know what the input should be and what output you expect, then I think you're good to go.

Jeremy: Yeah, yeah. Well, I still think that there's probably a lot of value in tuning the models though against your particular data sets.

Alexa: Yeah, right. But also there are services for hyperparameter tuning. There are services even for neural architecture search, where they try a lot of different architectures for your data specifically and then they will tell you what is the best model architecture that you should use and same for the hyperparameter search. So these can be automated as well.

Jeremy: Yeah. Very cool. So if you are hosting your own version of this ... I mean, maybe you'll go back to the MLOps piece of this. So I would assume that a data scientist doesn't want to be responsible for maintaining the servers or the virtual machines or whatever it is that it's running on. So you want to have this workflow where you can get your models trained, you can get them into production, and then you can run them through this loop you talked about and be able to tweak them and continue to retrain them as things go through. So on the other side of that wall, if we want to put it that way, you have your ops people that are running this stuff. Is there something specific that ops people need to know? How much do they need to know about ML, as opposed to ... I mean, the data scientists, hopefully, they know more. But in terms of running it, what do they need to know about it, or is it just a matter of keeping a server up and running?

Alexa: Well, I think ... So I think the machine learning pipelines are not yet as standardized as a traditional software engineering pipeline. So I would say that you have to have some knowledge of machine learning or at least some understanding of how this lifecycle works. You don't actually need to know about research and things like that, but you need to know how this whole lifecycle works in order to work as an ops person who can automate this. But I think the software engineering skills and DevOps skills are the base, and then you can just build this knowledge on top of that. So I think it's actually quite easy to pick this up.

Jeremy: Yeah. Okay. And what about, I mean, you mentioned this idea of a lot of data scientists aren't actually writing the models, they're just using the preconfigured model. So I guess that begs the question: How much does just a regular person ... So let's say I'm just a regular developer, and I say, "I want to start building machine learning tools." Is it as easy as just pulling a model off the shelf and then just learning a little bit more about it? How much can the average person do with some of these tools out of the box?

Alexa: So I think most of the time, it's that easy, because usually the use cases that someone tries to tackle, those are not super edge cases. So for those use cases, there are already models which perform really well. Especially if you are talking about, I don't know, supervised learning on tabular data, I think you can definitely find models that will perform really well off the shelf on those type of datasets.

Jeremy: Right. And if you were advising somebody who wanted to get started... I mean, because I think that I think where it might come down to is going to be things like pricing. If you're using Vision API and you're maybe limited on your quota, and then you can ... if you're paying however many cents per, I guess, lookup or inference, then that can get really expensive as opposed to potentially running your own model on something else. But how would you suggest that somebody get started? Would you point them at the APIs or would you want to get them up and running on TensorFlow or something like that?

Alexa: So I think, actually, for a developer, just using an API would be super easy. Those APIs are, I think ... So getting started with those APIs just to understand the concepts are very useful, but I think getting started with Tensorflow itself or just Keras, I definitely I would recommend that, or just use scikit-learn, which is a more basic package for more basic machine learning. So those are really good starting points. And there are so many tutorials to get started with, and if you have an idea of what you would like to build, then I think you will definitely find tutorials which are similar to your own use case and you can just use those to build your custom pipeline or model. So I would say, for developers, I would definitely recommend jumping into TensorFlow or scikit-learn or XGBoost or things like that.

Jeremy: Right, right. And how many of these models exist? I mean, are we talking there's 20 different models or are we talking there's 20,000 models?

Alexa: Well, I think ... Wow. Good question. I think we are more towards today maybe not 20,000, but definitely many thousands, I think. But there are popular models that most of the people use, and I think there are maybe 50 or 100 models that are the most popular and most companies use them and you are probably fine just using those for any use case or most of the use cases.

Jeremy: Right. Now, and speaking of use cases, so, again, I try to think of use cases or machine learning and whether it's classifying movies into genres or sentiment analysis, like I said, or maybe trying to classify news stories, things like that. Fraud detection, you mentioned. Those are all great use cases, but what are ... I know you've worked on a bunch of projects. So what are some of the projects that you've done and what were the use cases that were being solved there, because I find these to be really interesting?

Alexa: Yeah. So I think a nice project that I worked on was a project with Lush, which is a cosmetics company. They manufacture like soaps and bath bombs. And they have this nice mission that they would like to eliminate packaging from their shops. So they asked us, when I worked at Datatonic, we worked on a small project with them. They asked us to create an image recognition model, to train one, and then create a retraining pipeline that they can use afterwards. So they provided us with many hundred thousand images of their products, and they made photos from different angles with different lightings and all of that, so really high-quality image data set of all their products.

And then, we used a mobile net model, because they wanted this model to be built-in into their mobile application. So when users actually use this model, they download it with their mobile application. And then, they created a service called Lush [inaudible], which you can use from within their app. And then, people can just scan the products and they can see the ingredients and how-to-use guides and things like that. So this is how they are trying to eliminate all kinds of packaging from their shops, that they don't actually need to put the papers there or put packaging with ingredients and things like that.

And in terms of what we did on the technical side, so as I mentioned, we used a mobile net model, because we needed to quantize the model in order to put it on a mobile device. And we used TF Lite to do this. TF Lite is specifically for models that you want to run on an edge device, like a mobile phone. So that was already a constraint. So this is how we picked a model. I think, back then, like there were only a few model architectures supported by TF Lite, and I think there were only two, maybe. So we picked MobileNet, because it had a smaller size.

And then, in terms of the retraining, so we automated the whole workflow with Cloud Composer on Google Cloud, which is a managed version of Apache Airflow, the open source scheduling package. The training happened on AI Platform, which is Google Cloud's SageMaker.

Jeremy: Yeah.

Alexa: Yeah. And what else? We also had an image pre-processing step just before the training, which happened on Dataflow, which is an auto-scaling processing service on Google Cloud. And after we trained the model, we just saved the model active artifact in a bucket, and then ... I think we also monitored the performance of the model, and if it was good enough, then we just shipped the model to developers who actually they manually updated the model file that went into the application that people can download. So we didn't really see if they use any shadow model thing or anything like that.

Jeremy: Right. Right. And I think that is such a cool use case, because, if I'm hearing you right, there were just like a bar soap or something like that with no packaging, no nothing, and you just hold your mobile phone camera up to it or it looks at it, determines which particular product is, gives you all that ... so no QR codes, no bar codes, none of that stuff. How did they ring them up though? Do you know how that process worked? Did the employees just have to know what they were or did the employees use the app as well to figure out what they were billing people for?

Alexa: Good question. So I think they wanted the employees as well to use the app.

Jeremy: Nice.

Alexa: Yeah. But when the app was wrong, then I don't know what happened.

Jeremy: Just give them a discount on it or something like that. That's awesome. And that's the thing you mentioned there about ... Was it Tensor Lite, was it called?

Alexa: TF Lite. Yeah.

Jeremy: TF Lite. Yes. TensorFlow Lite or TF Lite. But, basically, that idea of being able to really package a model and get it to be super small like you said. You said edge devices, and I'm thinking serverless compute at the edge, I'm thinking Lambda functions. I'm thinking other ways that if you could get your models small enough in package, that you could run it. But that'd be a pretty cool way to do inference, right? Because, again, even if you're using edge devices, if you're on an edge network or something like that, if you could do that at the edge, that'd be a pretty fast response time.

Alexa: Yeah, definitely. Yeah.

Jeremy: Awesome. All right. So what about some other stuff that you've done? You've mentioned some things about fraud detection and things like that.

Alexa: Yeah. So fraud detection is a use case for Wise. As I mentioned, Wise services international money transfer, one of its services. So, obviously, if you are doing anything with money, then a full use case is for sure that you will have. So, I mean, in terms of ... I don't actually develop models at Wise, so I don't know actually what models they use. I know that they use H2O, which is a Spark-based library that you can use for model training. I think it's quite an advanced library, but I haven't used it myself too much, so I cannot talk about that too much.

But in terms of the workflow, it's quite similar. We also have Airflow to schedule the retraining of the models. And they use EMR for data preparation, so quite similar to Dataflow, in a sense. A Spark-based auto-scaling cluster that processes the data and then, they train the models on EMR as well but using this H2O library. And then in the end, when they are happy with the model, we have this tool that they can use for releasing shadow models in production. And then, if they are satisfied with the performance of the model that they can actually release into production. And at Wise, we have a custom micro service, a custom API, for serving models.

Jeremy: Right. Right. And that sounds like you need a really good MLOps flow to make all that stuff work, because you just have a lot of moving parts there, right?

Alexa: Yeah, definitely. Also, I think we have many bits that could be improved. I think there are many bits that still a bit manual and not streamlined enough. But I think most of the companies struggle with the same thing. It's just we don't yet have those best practices that we can implement, so many people try many different things, and then ... Yeah, so I think it's still a work in progress.

Jeremy: Right. Right. And I'm curious if your economics background helps at all with the fraud and the money laundering stuff at all?

Alexa: No.

Jeremy: No. All right. So what about you worked in another data engineering project for Vodafone, right?

Alexa: Yeah. Yeah, so that was a data engineering project purely, so we didn't do any machine learning. Well, Vodafone has their own Google Analytics library that they use in all their websites and mobile apps and things like that and that sense Clickstream data to a server in a Google Cloud Platform Project, and we consume that data in a streaming manner from data flows. So, basically, the project was really about processing this data by writing an Apache Beam pipeline, which was always on and always expected messages to come in. And then, we dumped all the data into BigQuery tables, which is data warehouse in Google Cloud. And then, these BigQuery tables powered some of the dashboards that they use to monitor the uptime and, I don't know, different metrics for their websites and mobile apps.

Jeremy: Right. But collecting all of that data is a good source for doing machine learning on top of that, right?

Alexa: Yeah, exactly. Yeah. I think they already had some use cases in mind. I'm not sure if they actually done those or not, but it's a really good base for machine learning, what we collected the data there in BigQuery, because that is an analytical data warehouse, so some analysts can already start and explore the data as a first step of the machine learning process.

Jeremy: Right. I would think anomaly detection and things like that, right?

Alexa: Yeah, exactly.

Jeremy: Right. All right. Well, so let's go on and talk about serverless a little bit more, because I know I saw you do a talk where you were you ran some experiments with serverless. And so, I'm just kind of curious, where are the limitations that you see? And I know that there continues ... I mean, we now have EFS integration, and we've got 10 gigs of memory for lambda functions, you've even got Cloud Run, which I don't know how much you could do with that, but where's still some of the limitations for running machine learning in a serverless way, I guess?

Alexa: So I think, actually, from this data science lifecycle, many bits, there are Cloud providers offer a lot of serverless options. For data preparation, there is Dataflow, which is, I think, kind of like serverless data processing service, so you can use that for data processing. For model training, there is ... Or the SageMaker and AI Platform, which are kind of serverless, because you don't actually need to provision these clusters that you train your models on. And for model serving, in SageMaker, there are the serverless model endpoints that you can deploy. So there are many options, I think, for serverless in the machine learning lifecycle.

In my experience, many times, it's a cost thing. For example, at Wise, we have this custom model serving API, where we serve all our models. And if they would use SageMaker endpoints, I think, a single SageMaker endpoint is about $50 per month, that's the minimum price, and that's for a single model and a single endpoint. And if you have thousands of models, then your price can go up pretty quickly, or maybe not thousands, but hundreds of models, then your price can go up pretty quickly. So I think, in my experience, limitation could be just price.

But in terms of ... So I think, for example, if I compare Dataflow with a spark cluster that you program yourself, then I would definitely go with Dataflow. I think it's just much easier and maybe cost-wise as well, you might be better off, I'm not sure. But in terms of comfort and developer experience, it's a much better experience.

Jeremy: Right. Right. And so, we talked a little bit about TF Lite there. Is that something possible where maybe the training piece of it, running that on Functions as a Service or something like that maybe isn't the most efficient or cost-effective way to do that, but what about running models or running inference on something like a Lambda function or a Google Cloud function or an Azure function or something like that? Is it possible to package those models in a way that's small enough that you could do that type of workload?

Alexa: I think so. Yeah. I think you can definitely make inference using a Lambda function. But in terms of model training, I think that's not a ... Maybe there were already experiments for, I'm sure there were. But I think it's not the kind of workload that would fit for Lambda functions. That's a typical parallelizable, really large-scale workloads for ... You know the MapReduce type of data processing workloads? I think those are not necessarily fit for Lambda functions. So I think for model training and data preparation, maybe those are not the best options, but for model inference, definitely. And I think there are many examples using Lambda functions for inference.

Jeremy: Right. Now, do you think that ... because this is always something where I find with serverless, and I know you're more of a data scientist, ML expert, but I look at serverless and I question whether or not it needs to handle some of these things. Especially with some of the endpoints that are out there now, we talked about the Vision API and some of the other NLP things, are we putting in too much effort maybe to try to make serverless be able to handle these things, or is it just something where there's a really good way to handle these by hosting your ... I mean, even if you're doing SageMaker, maybe not SageMaker endpoints, but just running SageMaker machines to do it or whatever, are we trying too hard to squeeze some of these things into a serverless environment?

Alexa: Well, I don't know. I think, as a developer, I definitely prefer the more managed versions of these products. So the less I need to bother with, "Oh, my cluster died and now we need to rebuild a cluster of things," and I think serverless can definitely solve that. I would definitely prefer the more managed version. Maybe not serverless, because, for some of the use cases or some of the bits from the lifecycle, serverless is not the best fit, but a managed product is definitely something that I prefer over a non-managed product.

Jeremy: Right. And so, I guess one last question for you here, because this is something that always interests me. Just there are relevant things that we need machine learning for. I mean, I think the fraud detection is a hugely important one. Sentiment analysis, again. Some of those other things are maybe, I don't know, I shouldn't call them toy things, but personalization and some of the things, they're all really great things to have, and it seems like you can't build an application now without somebody wanting some piece of that machine learning in there. So do you see that as where we are going where in the future, we're just going to have more of these APIs?

I mean, out of AWS, because I'm more familiar with the AWS ecosystem, but they have Personalize and they have Connect and they have all these other services, they have the recommendation engine thing, all these different services ... Lex, or whatever, that will read text, natural language processing and all that kind of stuff. Is that where we're moving to just all these pre-trained, canned products that I can just access via an API or do you think that if you're somebody getting started and you really want to get into the ML world that you should start diving into the TensorFlows and some of those other things?

Alexa: So I think if you are building an app and your goal is not to become an ML engineer or a data scientist, then these canned models are really useful because you can have a really good recommendation engine in your product, you could have really good personalization engine in your product, things like that. And so, those are, I think, really useful and you don't need to know any machine learning in order to use them. So I think we definitely go into that direction, because most of the companies won't hire data scientists just to train a recommender model. I think it's just easier to use an API endpoint that is already really good.

So I think, yeah, we are definitely heading into that direction. But if you are someone who wants to become a data scientist or wants to be more involved with MLOps or machine learning engineering, then I think jumping into TensorFlow and understanding, maybe not, as we discussed, not getting into the model architectures and things like that, but just understanding the workflow and being able to program a machine learning pipeline from end to end, I think that's definitely recommended.

Jeremy: All right. So one last question: If you've ever used the Watson NLP API or the Google Vision API, can you put on your resume that you're a machine learning expert?

Alexa: Well, if you really want to do that, I would give it a go. Why not?

Jeremy: All right. Good. Good to know. Well, Alexa, thank you so much for sharing all this information. Again, I find the use cases here to be much more complex than maybe some of the surface ones that you sometimes hear about. So, obviously, machine learning is here to stay. It sounds like there's a lot of really good opportunities for people to start kind of dabbling in it and using that without having to become a machine learning expert. But, again, I appreciate your expertise. So if people want to find out more about you or more about the things you're working on and datastack.tv, things like that, how do they do that?

Alexa: So we have a Twitter page for datastack.tv, so feel free to follow that. I also have a Twitter page, feel free to follow me, account, not page. There is a datastack.tv website, so it's just datastack.tv. You can go there, and you can check out the courses. And also, we have created a roadmap for data engineers specifically, because there was no good roadmap for data engineers. I definitely recommend checking that out, because we listed most of the tools that a data engineer and also machine learning engineer should know about. So if you're interested in this career path, then I would definitely recommend checking that out. So under datastack.tv's GitHub, there is a roadmap that you can find.

Jeremy: Awesome. All right. And that's just, like you said, datastack.tv.

Alexa: Yes.

Jeremy: I will make sure that we get your Twitter and LinkedIn and GitHub and all that stuff in there. Alexa, thank you so much.

Alexa: Thanks. Thank you.