SDS 599: MLOps: Machine Learning Operations

This is episode number 599 with Mikiko Bazeley, senior software engineer at MailChimp.

Welcome to the SuperDataScience Podcast, the most listened to podcast in the data science industry. Each week, we bring you inspiring people and ideas to help you build a successful career in data science. I’m your host, Jon Krohn. Thanks for joining me today, and now, let’s make the complex simple.

Welcome back to the SuperDataScience Podcast. We’ve got a special episode today, all about MLOps, machine learning operations, featuring the MLOps expert, Mikiko Bazeley. Mikiko is a senior software engineer at MailChimp, the leading email newsletter platform, which was acquired last year by the financial tech giant, Intuit. In her engineering role at MailChimp, she’s responsible specifically for MLOps. Previously, she held technical roles at a range of Bay Area startups with responsibilities including software engineering, MLOps, data engineering, data science, and data analytics. She is a prominent content creator on MLOps across live workshops, her YouTube channel, her personal blog, and the blog of Nvidia, the AI hardware company.

Today’s episode will appeal primarily to hands on practitioners such as data scientists and software engineers. In this episode, Mikiko details what MLOps is, why MLOps is critical for the efficiency of any data science team, the three most important MLOps tools, the four myths holding people back from MLOps expertise, the six most essential MLOps skills for data scientists, and her productivity tricks for balancing software engineering, content creation, and her athletic pursuits. All right. Are you ready for this enriching technical episode? Let’s go.
Mikiko, welcome to the SuperDataScience Podcast. I’ve been excited to have you on air. I’ve known about you for so long, and I can’t believe that now I get to talk to you and ask you all of the questions I’ve been waiting to ask. Where in the world are you calling in from, Mikiko? 

Mikiko Bazeley: 00:02:20

I am coming from the beautifully muggy and foggy city of San Francisco. 

Jon Krohn: 00:02:25

Yes. What’s that? There’s a quote by a famous American author. It’s something like, “I’ve never had a older winter than my summer in San Francisco,” or something. Ernest Hemingway or something? 

Mikiko Bazeley: 00:02:39

Oh, that sounds absolutely accurate, although I guess people who are in Boston would be like, “Ah, it’s from a Boston winter. Why don’t you?” 

Jon Krohn: 00:02:50

At least they have a summer, I think was the point. Was that- 

Mikiko Bazeley: 00:02:53

Yeah. 

Jon Krohn: 00:02:53

In San Francisco, it’s winter year round. Well, it is a hop-in city for machine learning, machine learning operations. So I imagine lots of interesting things for you to do out there, even despite the maybe not ideal weather. Plus, you also have lots of wonderful weather climates within a few hours drive, all kinds of treats. 

Mikiko Bazeley: 00:03:22

Absolutely. I have to say I was born and raised in the city, so I’m a little bit biased, but there’s no other place I’d really want to be, especially for the work that I do, and also, too, being close to my family. It’s been great. 

Jon Krohn: 00:03:36

That is nice. That is an important thing. All right. So we had never met before recording this episode. We had never had a conversation before, but I’ve known who you are through various people for it feels like years. So it seems like perhaps the linchpin in all of this is Harpreet Sahota, who was in episode number 457 on the SuperDataScience Podcast, because he runs his Artist of Data Science happy hours, where anybody can drop in. Those are on Fridays, right? 

Mikiko Bazeley: 00:04:15

Yeah. Fridays at, what was it? It’s 2:00, 3:00 PM PST. So don’t remember what time. It’s Winnipeg time. Yeah. It’s 5:00 PM. 

Jon Krohn: 00:04:23

It’s Winnipeg time. I guess it would be 3:00 PM in Winnipeg, I think, for Harpreet locally. You’ve met tons. We have tons of mutual connections from you doing that. I’ve actually never been to the happy hour. I would love to do that, but tons of guests on the show have been in that happy hour. You’re saying Mark Freeman. He was a recent guest. Serg Masís is the researcher for the SuperDataScience Podcast. So before a guest comes on air, Serg does a detail deep dive into their history. He studies YouTube videos, lectures that they’ve given, blog posts that they’ve written, journal articles that they’ve written, and he comes up with the most amazing questions that bridge different parts of people’s careers, that tie threads together across their careers, and also, he asks questions that nobody else on the planet could be asked.

He’s got questions for you that I’ll ask you today based on your specific experience, Mikiko. Nobody else could possibly be asked the same question. Serg is absolutely brilliant, and he also has a very popular SuperDataScience episode number 539. He wrote to me, before he provided me with the questions for you for today, he also said that, “Mikiko is so smart and always has something interesting to say.” So he’s looking forward to hearing this episode, no doubt. Yeah. So that’s someone else from those Artists of Data Science happy hours that you know. 

Mikiko Bazeley: 00:06:01

Yeah, absolutely. Serg is such a great guy. It’s so funny because I guess you never know who’s watching you in a way. A lot of these people I consider my friends, but I also consider them people that I look up to, especially when thinking about how I develop and grow my own career in MLOps. That’s so nice to hear. 

Jon Krohn: 00:06:19

Yeah. Serg is absolutely wonderful. Yeah. So listeners, check out. If you’ve got questions about your data science career that you’d like to ask stable, brilliant data science leaders and content creators, that happy hour is a great way to do that. So speaking of Serg’s questions, here is one from him. So he points out, Mikiko, that you’re a content creator on Medium, Substack, YouTube, and Nvidia’s blog, lots of other places. We’re going to talk about some more of those later in this episode, and you use all this content creation to share your tips on machine learning operations, MLOps, as well as production machine learning, and entering the field of data science machine learning.
Previously, you mentored folks at Springboard. You also mentor interns at your job. You wrote a software engineering course called Software Engineering Skills 101, and you run the data book club at MailChimp where you work. So clearly, mentorship and teaching is a big part of your life. It seems like it’s something that’s very important to you. Where does that come from? When did that get started? 

Mikiko Bazeley: 00:07:38

Yeah. I think it comes from two or three main places. One was the fact that whenever I was at a critical juncture in terms of my career, I always had, even if I didn’t have resources, I had one or two people, key linchpins that were able to help teach me the skills I needed to get through to the crucible and onto the next stage. That happened the first time I got a job at a startup and I had no programming skills. My friend, who was a researcher who used R a lot, helped me understand, pick up enough of it to write up a script to do the analysis for the take home interview. That was one key point because it was my very first job in tech, and I didn’t have anyone in my family that I knew of that worked in tech before.
There were so many linchpin critical moments like that, where someone was willing to reach out a hand and help pull me through in the classic mentor-teacher apprentice style. So that’s something that I want to help give to other people, understanding that not everyone has the right resources, has the right educational background to be able to succeed in a very, very demanding industry. So that’s part of it.

A second part was when I was younger, I had a lot of issues with reading, actually, which is funny because I love reading now, but in elementary school, I was considered very slow and, actually, that’s been the feedback I’ve gotten a lot of times when I was growing up both in elementary school, middle school, and high school was that I was hard working, but I was not quite bright, just not bright enough to pass the line.
I think part of that was because I was surrounded by people who didn’t have that growth mindset. They really thought in terms of you either got it or you don’t. I think that’s a very damaging attitude. So eventually when I graduated college and I had to reset what I knew and understood of how someone succeeds in the world, part of it was learning about the growth mindset and understanding that people’s skills and experiences are very, very adaptable and flexible, and that was just life-changing for me because before, I was almost helpless to this idea that, “Oh, I’m hard working, but I’m not quite bright,” like a golden retriever.

No one wants to be told that, but when I had that critical piece of information, I was able to then start to teach myself how to fish.
I think those experiences of constantly being underestimated, constantly being towed, “Oh, you’re so sweet, you’re so nice. Maybe you should try going into something with people, away from technical stuff. Oh, maybe you’re not good with numbers,” or what have you, made me understand that, look, at the end of the day, you should allow people and you should give people the tools for them to self-determine their own destiny and all those people were wrong. I mean, maybe they’re not wrong. Maybe I’m not quite bright. Maybe I am still more hardworking than I am bright, but I have to say I enjoy my life a lot more, and certainly my career. 

Jon Krohn: 00:11:05

I don’t know. Just from the conversation that we were having prior to recording, I can confidently say that you are in fact quite bright.

Mikiko Bazeley: 00:11:12

Thank you. 

Jon Krohn: 00:11:14

So yeah, and that growth mindset or whatever has certainly has transformed you or allowed parts of you to come out and flourish. Yeah. You’re an exceptional leader in this space. To wit, you have a machine learning operations course that you’re currently working on. So first of all, maybe we should break down for the listener what machine learning operations is. 

Mikiko Bazeley: 00:11:46

Absolutely, and this is something that I was developing a blog post series about it. I had initially scoped it to five pages because I’m like, “I don’t know if people will even read five pages,” and I just kept writing and writing. It ended up being a 24-page blog that I’m currently trimming down to now turn to videos because defining what is MLOps, it’s so controversial right now in that space. Everyone has their different lines. Everyone has their diagrams that they’re putting out there. I would say for me the simplest explanation of MLOps is you are developing the tooling and the infrastructure to make developing, productionizing, and deploying models a lot easier. 

Jon Krohn: 00:12:30

Nice. 

Mikiko Bazeley: 00:12:32

We could add in more things, too, like easier, more accurate, more robust, all that, but that’s ultimately what it is. 

Jon Krohn: 00:12:39

So in a big enough team that can afford to have more than just data scientists and software engineers, you have this additional role machine learning operations that helps data scientists to create models more efficiently, to do their work more replicably, to ensure that there’s redundancy if something goes wrong, that experiments are recoverable, all those kinds of things. Yeah. 

Mikiko Bazeley: 00:13:09

Yeah, absolutely. It’s funny because the data scientist role, it’s evolved. I think there’s debate about what is the difference between a data scientist versus an MLOps engineer versus a machine learning engineer versus a devops engineer versus a data engineer. In my head, those distinctions become much more relevant in a mature enterprise setting, but at the end of the day, I think MLOps, it’s a role or it’s a discipline or domain that still works with other roles.
So for example as an MLOps engineer, I collaborate really closely with my data science counterparts, with my data engineering counterparts. We do have those that exist. I have to say, to me, it’s a very valuable collaboration, and more importantly, we all have skillsets and experiences in different areas.

Our concerns are going to be much more different. We don’t necessarily want data scientists to constantly be retooling the same pipelines over and over again when maybe the only difference is actually the model and the data input.
Now, that’s significant, but in terms of all the other infrastructure that develops around that, there’s no reason why that shouldn’t be templated and Cookiecuttered and why they can’t just go, “I know I need this pattern. I know I want to do a streaming recommendation service,” for example. So they should be able to say, “Okay. Give me the pattern that is best for a streaming recommendation service, and we should essentially have that rate to serve.”
That’s my personal philosophy of the, yes, we’re an infrastructure engineering team, but we are also, if we think about it, internal developer advocates. We are internal tool developers for our main customers, which is data scientists.

Jon Krohn: 00:15:06

Cool. That’s a really nice explanation.
Eliminating unnecessary distractions is one of the central principles of my lifestyle. As such, I only subscribe to a handful of email newsletters, those that provide a massive signal to noise ratio. One of the very few that meet my strict criterion is the Data Science Insider. If you weren’t aware of it already, the Data Science Insider is a 100% free newsletter that the Super Data Science team creates and sends out every Friday. We pour over all of the news and identify the most important breakthroughs in the fields of data science, machine learning, and artificial intelligence. The top five, simply five news items, the top five items are handpicked.

The items that were confident will be most relevant to your personal and professional growth. Each of the five articles is summarized into a standardized, easy-to-read format, and then packed gently into a single email. This means that you don’t have to go and read the whole article. You can read our summary and be up to speed on the latest and greatest data innovations in no time at all.
That said, if any items do particularly tickle your fancy, then you can click through and read the full article. This is what I do. I skim the Data Science Insider newsletter every week. Those items that are relevant to me, I read the summary in full. If that signals to me that I should be digging into the full original piece, for example, to pour over figures, equations, code or experimental methodology, I click through and dig deep.

So if you’d like to get the best signal to noise ratio out there in data science, machine learning, and AI news, subscribe to the Data Science Insider, which is completely free, no strings attached at www.superdatascience.com/dsi. That’s www.superdatascience.com/dsi. Now, let’s return to our amazing episode.
All right. So now that we know what MLOps is, and later on in the program, I’ll have questions for you about things like what the key skills, what the key MLOps skills are for data scientists. So maybe you’re in a smaller organization where you don’t have the luxury of MLOps specialists in your company, so what are the key MLOps things that data scientists need to know on those smaller teams? But before we get to that, let’s talk about your MLOps course that I mentioned. So what are you going to cover in that? When’s it going to come out? When can we get it? 

Mikiko Bazeley: 00:17:40

Yeah. So the first course, I’m thinking of a series of courses, the first one is going to be an intro to MLOps course. When it’s coming out? I’m targeting three or four months from now. We’ll see how that goes, right? You know the whole saying, “Man plans, God laughs.” There’s that. So we’ll see how that works out. In terms of the coverage, so I see a lot of courses out there where they have a certain implementation that they’re proposing and there’s certain assumptions.
Actually, my favorite project out there is the MLOps at reasonable scale slash you don’t need a bigger boat project by Yako. That is fantastic because he directly contradicts this idea that MLOps is an enterprise-only concern. With that being said, there are tons of courses out there, tons of implementations that are going on.

My goal with the intro to MLOps course is to take it a step up, a higher level of abstraction and go, “Well, why do you need all these things?” because what I feel like is people get stuck in the, “Oh, this is the best tool for X, Y, Z,” and what I feel a successful MLOps engineer needs is to be able to think in terms of decision trade offs, and also what is the overall architecture, and is the roadmap and the north star of where they want to get to with their project because the tools or the pattern or architecture that let’s say, for example, a solopreneur sets up, it’s going to be different from the way, for example, we maybe do things at MailChimp or Google scale, Amazon scale or even let’s say a company that’s not traditionally tech, GE maybe.

The current concerns they have and also their starting points are going to be way different. So to me, what is most important is that someone is able to look at the starting point, look at where they want to go, and then start saying, “Okay. What are the concerns we really care about? How do we stack rank it, and then how do we attack it with architecture that is not just for the here now, but is maybe for 10x growth?”
That to me is what I ideally think I want to get at with the course is helping people first understand that higher level of what are the concerns we’re getting at, what are the different designs and trade offs that people need to make, and then following courses after that, we’ll go into a little bit more of the technical implementation. We’ll go into, for example, if you’re a career changer or if you’re someone who’s interested in pivoting into area, what are the things that you would want to develop? How does your portfolio look like? How do you want to prepare your LinkedIn resume? So I’ll have a series of those, but that’s what the first one is about. 

Jon Krohn: 00:20:34

Nice. That sounds great, the high level philosophy of what machine learning ops is for, and how you can be creating infrastructure that allows data scientists to be supported even after 10x growth. That makes a lot of sense to me, and it will be great to have those kinds of career tips as well. I’m sure our listeners would love to receive those, and we will have some career tips, actually, in this very episode. So in addition to the video course that you’re creating, you also have live workshops coming up. 

Mikiko Bazeley: 00:21:09

Yup. Absolutely. I’m very excited for those. Those will be a little bit more on the technical implementation side, but I’ll also have some that maybe talk more to also about best practices around MLOps. 

Jon Krohn: 00:21:24

Nice. Then so with those things forthcoming like your course, your workshops, at the end of the episode, I always ask for how people should follow you, but it feels like I might as well ask right now so that listeners have a sense of what they can be doing to make sure that they don’t miss out when these resources are released. 

Mikiko Bazeley: 00:21:47

Yeah, absolutely. So my main platforms I’m going to be living on are LinkedIn, Medium, and YouTube. I’m grateful to the thousand subscribers on my YouTube channel. Thank you very much for the support. Those are the three where content will either be announced or released with my … I mentioned Substack. Medium and Substack, they’re married for now. So I think either is good. Substack will ultimately be where I land on. People can also follow me on GitHub itself because that’s where I intend to release the code demos and the sample projects. I also occasionally yell at people on Twitter. 

Jon Krohn: 00:22:28

Very good. So with things like your personal Substack, you intend on having a twice monthly cadence with technical blog posts and then on your YouTube channel, which I’m blown away because you already have a thousand subscribers to your YouTube channel, but you’ve only released a few short videos. They’re really short. They’re less than a minute. So people are obviously itching to get the MLOps content. So what you’re planning on doing is having the personal Substack, have these technical blog posts, and then you follow those up with videos on YouTube that take advantage of your capability as someone who can create visuals and videos. You seem to have a lot of experience with that kind of stuff. So you can create fun and entertaining ways of bringing to life the technical stuff that you publish in Substack. 

Mikiko Bazeley: 00:23:25

Yeah, absolutely. I think it is great if someone, so my channels and all that, if you are someone who is looking for that big sister mentor, I’m going to be a gal. Specifically, I try to emulate the mentors and the educators and influencers I had in my life, especially in assuming that most people are probably going to start from a different place than maybe, say, a more traditional candidate. So my goal is to target that audience for people who are like, “Oh, God! This is just so confusing. Why do we have another job title name, data scientist, data engineer? Why do we have more?” So people are a little bit confused and maybe want a little bit of a friendlier voice. I recommend them to definitely come to the channels. 

Jon Krohn: 00:24:22

Nice. Then on top of all that, on top of all of your personal content creation, your MLOps course, your live workshops, your personal Substack, your YouTube channel, you also do writing for other blogs like Nvidia. You recently had a post for Nvidia that was an introduction to speech AI. I don’t know if you want to talk about that at all. 

Mikiko Bazeley: 00:24:46

Yeah. I mean, it was definitely a really fun process. I appreciate my partners in the tech space, especially the ones who are recognizing that, “Hey, MLOps is not just about tools. It’s not just about fancy use cases. It’s not just about helping companies not get sued for stuff, but also recognizing that it is also a valuable area for individual practitioners.” So yeah, so that article is fun. Would definitely check it out, especially since Nvidia is doing some really cool stuff with Texas Speech and also Speech Generation.

Jon Krohn: 00:25:22

Very cool. Yeah, Nvidia is doing a lot of the cool stuff in AI period. So we have alluded to you having a job on top of doing all of this content creation, which I will have questions near the end of the episode on how you manage to do all of this at the same time. So your day job is working at MailChimp, which was acquired by Intuit. There, you’re a senior software engineer responsible for MLOps and infrastructure.
So I was introduced to MailChimp many years ago through their ads on the Serial podcast. So they had these funny, really memorable ads where they had all these people saying, “Mailkimp,” like mispronunciations of MailChimp. I don’t know. They were really catchy, really weird. All these years later, I still use MailChimp for my email newsletters. So I think MailChimp is a cool company. They definitely make it easy for me to make email newsletters. In recent years, they have branched out into lots of different ways of marketing to your audience, not just email newsletters. Then when was the acquisition by Intuit? That was it just in the last year or two, wasn’t it? 

Mikiko Bazeley: 00:26:55

Yep. That was in November of last year. 

Jon Krohn: 00:26:57

November 2021. Yeah. So you’re working on the MailChimp product. Technically, you’re a part of the bigger Intuit company. So what is it like to be a senior software engineer responsible for MLOps and infrastructure at Intuit? 

Mikiko Bazeley: 00:27:19

Yeah, I mean, first off, I love my work. I do love my work. Second part is, so in terms of what my day-to-day looks like, first off, we are a vital partner to our data scientists on pipeline and model architectures. So going back to that definition of MLOps earlier, I really view it, some people view MLOps as a practice or domain that could be shared by multiple teams, data science, data engineering, and so on and so forth. I think in smaller companies, that is definitely the case. A lot of times people are implementing MLOps best practices without necessarily having someone assigned to the role because all of MLOps best practices I would argue is a extension of devops.
So we are a core engineering team. So we consult with the data scientist. We assist with infrastructure and ops specific tasks, troubleshoot bugs in our tooling, develop new features, make it even easier to productionize data science models. There’s also mentoring opportunities.

So currently, I have an intern that I’m mentoring. We also do on-call a lot of other infrastructure and platform teams. So we make ourselves available on certain days from 9:00 to 5:00 to to help unblock any issues that come up.
I have significant opportunities also to impact the culture in the hiring. So I really, really enjoy it, and especially in the MLOps space. It’s like we’re building the plane as we’re flying it. A lot of the practices we are borrowing from other disciplines that came before us. So there are things we’re going to have to as an industry, as a domain, as a practice, not just my team, but also the wider group of MLOps practitioners, there are some things we’re going to have to figure out because traditional software development, there are certain assumptions that are baked in there that a machine learning product just throws out the window. Yeah. So I really enjoy it, and I’m really excited for the kind of work that we’re doing, too, for sure, at MailChimp. 

Jon Krohn: 00:29:34

You mentioned their devops. So that just got me thinking about this potential parallel. I’d love to hear your feedback on this. I’m just riffing on this idea here, but would you say that it’s fair to make the parallel that devops supports software engineering analogously to the way that MLOps supports data science and machine learning model development? Would you say that that’s a reasonable comparison? 

Mikiko Bazeley: 00:30:05

Yeah, absolutely. I think other things too. So MLOps, so it’s quite fascinating. So one blog post I’m working on, I also did Twitter thread on this, was looking at, I don’t want to say the genesis of MLOps, but, yeah, looking at the genesis of MLOps because a lot of the tools that even in devops that we use, for example, Jenkins, Docker, Kubernetes, they were developed in, well, they were released in their first stable versions only between 2012 and 2015, which is a relatively young timeline. When I was going back into because I was very interested in … Our stuff is so new in a way. The practices are so new. Why are some of these things so difficult? When I looked at the tooling because tooling is a big part of the conversation with MLOps, Jenkins, Docker, Kubernetes, TensorFlow, PyTorch, those were only released within maybe the last five to eight years as a first stable version, which is really fascinating. GitHub, I think, was released around 2008.

So it’s fascinating because people enter the field and they don’t always have the history and context to understand what came before. I think having that understanding when you realize so much of the cool technology that we’re doing. Yes, Yann LeCun might have been working on deep learning years ago, and some of the initial foundational deep learning papers were pre-2000s, but in terms of the tooling and the practices and the technology around supporting MLOps like machine learning products at scale, that has only been within a so short period of time. Those tools are still developing. Our ability to use them and how we think we should use them is also just very nascent. So I think there’s still more to come, for sure. 

Jon Krohn: 00:32:14

Let’s talk about some of those key tools in more detail. So you mentioned what are probably the big three in MLOps, I’m going to assume, Docker, Jenkins, and Kubernetes. So let’s start with Docker and explain to the listeners what Docker containers are and how they allow data science work to be easier to spin up and easier to replicate. 

Mikiko Bazeley: 00:32:40

Yeah, absolutely. It was really funny because I had written a blog post. It’s called the Why of Virtual Environments Containers and VMs because I remember when I was first moving more into the ML engineering and then MLOps role, so I had worked as a data scientist for a few years and I had developed models in a Jupyter notebook. I spit out a pickle file and I have no idea what to do with it. More importantly, too, I was just using tools and practices that I was either taught in my bootcamp or that I would pick up from people around me.
For example, a lot of times people would say, “Oh, you’re struggling. Go use Anaconda. Go use XYZ,” but they would never help me understand the why of it. So I like to think of Docker containers. So if we think about, and I have a bias towards Python, so that’s just a warning for, I guess, listeners, right? 

Jon Krohn: 00:33:44

No. You’re doing production deployment, so that makes sense. 

Mikiko Bazeley: 00:33:48

Ah. All the R people are probably up in arms going, “How dare you?” If we look at, for example, we have a computer, right? It could be a laptop. It could be a more server grade computer, but you have the hardware, you have all the applications, everything is on the OS, lots of fun ways to screw it up, but you have so much control. Also, it’s physically in front of you.
We can think of a Docker container and virtual environments as representing different layers of abstraction. So a virtual environment on let’s say for Python, it’s analogous to a project folder. You have a language binary that sits in that folder. You import whatever libraries you need so I could learn, all that. It all sits in there, and then you can switch in and out of it, and the idea is that you’re not messing up your core like Python 2.7 that’s running your laptop.

Then we’ll talk about a virtual machine, right? A virtual machine is just a step up from an actual physical computer. You have lots of options. For example, you can have multiple OSs going on it, which means you can have multiple stacks of languages, of applications, of programs. We go a step up above a VM or a step down, actually, and you have a Docker container. A Docker container, it allows you to have that separation and encapsulation of the work that you’re doing. You still share an OS kernel, but essentially, it’s just one more layer of abstraction, where with a physical computer, you have the most amount of granularity, security, all that, as long as it’s not connected to the internet.
With a VM, you have a little bit less control, a little bit less granularity, but it’s a lot more flexible. You have containers, which are even more flexible. They’re a little bit more lighter weight, and then you have virtual environments.

The fun part is you could have all them on a single thing. So my physical computer can have a VM on it, which can then have multiple containers on it, which can then have multiple virtual environments on it. So it’s not in either or, but it’s that aspect of thinking through the hierarchy of what are your specific needs that you have or your security requirements and how you adjust that. If you’re a data scientist, that is just trying out different projects.
Usually, a virtual environment will be great, but if you’re a data scientist that is seriously considering pointing something to production, you’d want to use a container. More than likely, for people who are like, “What about container versus a VM?” every time you launch a container to GCP or AWS, it’s actually running on a VM. So in this case, it’s not an either or. It’s just some of that decision is being made for you as a data scientist. 

Jon Krohn: 00:36:48

Yeah, and I think Docker containers provide just the right level of abstraction for the vast majority of data scientist use cases because it allows you often. So you provide this Docker file that specifies exactly what operating system and software dependencies will be available once that Docker container is running. There are lots of pre-compiled Docker images that are very easy to download, that have exactly what you need, have Scikit-learn, have TensorFlow, and you know that everything is going to work together nicely.
So yeah, I highly recommend that if you’re not using Docker containers already as a data scientist, that you explore that possibility because it means that you can very easily replicate things because you’re not changing. You don’t necessarily need to change the library versions in that container.

So you can have a particular data science experiment that you run in a particular set of data with specific versions of all the software libraries specified in that Docker file. So you can come back to that a year later or two years later, and maybe even though in your local environment you’ve changed and now you’re using a much newer version of PyTorch, and it would clash with the version that you were using a year ago or two years ago, it doesn’t matter because you had specified in the Docker container you were using for that experiment a year or two ago this specific PyTorch version, and it’s going to run just like it did in the past.
So it makes it really easy to keep being able to run your code years later, and then something that you alluded to there, Mikiko, was security, and this is another great thing about having these Docker containers is that they’re separate from the rest of what you’re doing on your system. So you can mess around in that Docker container and it’s not going to ruin your whole computer. 

Mikiko Bazeley: 00:39:06

Absolutely. Yeah. It was really meant to solve the “It worked on my computer” statement. 

Jon Krohn: 00:39:13

Right. Exactly. 

Mikiko Bazeley: 00:39:14

Yeah, absolutely. I think if you’re a data scientist who is serious about developing applications and software, a cloud is a huge thing. Cloud is a huge thing, and having Docker containers. A lot of people, I think, concern themselves probably unnecessarily about vendor lock in, especially if you’re an individual practitioner, and containers give you some of that flexibility in being able to play around with different cloud providers, for sure. 

Jon Krohn: 00:39:43

Yeah. So you can, I mean, that’s the other key piece here. I gave the analogy of wanting to be able to continue to use some data science code, some data science experiment that you’ve written a year or two ago, but the much more common application, the analogy that I should have made is that it allows you to pass your Docker container off to anybody in your organization or outside your organization, and then they can run your code exactly the way that you did. So thank you for making that point.
All right. So that gives listeners an introduction to Docker containers if they weren’t already familiar with them. Then what can we do with Kubernetes and Jenkins once we have those Docker containers? 

Mikiko Bazeley: 00:40:30

Yeah, absolutely. So with Jenkins, some of the other tools that are used in place of Jenkins include GitHub actions. I think CircleCI, Buildkite, essentially, it’s a CICD tool. So it’s super, super important. 

Jon Krohn: 00:40:47

What is CICD?

Mikiko Bazeley: 00:40:48

Yeah. So CICDCD is continuous integration, continuous delivery, continuous deployment. It describes the practice of, well, it describes a lot of things, but mainly, it’s the practice of automating code for continuous integration. It’s building and testing applications and software for continuous delivery. It’s actually releasing the code to a production server, and then for continuous deployment, it’s actually deploying it either to a server or deploying it to an audience, rolling it out. Rollout is also included in this continuous deployment.
This is a crucial component of MLOps practices because, essentially, there’s a couple reasons why. One reason, obviously, is we don’t want the data scientists to manually be doing something that can be automated.

For one thing, it just makes their lives miserable, but more importantly, it exposes that process to fat-fingered issues, so making mistakes just, for example, setting up the configuration like typing in the wrong thing in the YAML file or whatever.
Also, too, we do need tools like linting. We need to run robust test suites just because we want to make sure we’re getting all the bugs before they get to production. Also, you don’t always have a QA person, and more importantly, the QA person, their job isn’t to find random bugs, but it’s to be an essential partner in terms of thinking about, “Is the software doing what we intended to do? Is it working not in the strictest sense of does the code run, but also, is it doesn’t meet the brand or the audience that we want to meet?” Yeah. So CICDCD, it’s an absolutely crucial part of MLOps and of any software, really.

Jon Krohn: 00:42:54

Nice. So Jenkins is an automation tool that allows data scientists’ processes to be automated in some way, to allow for this CICDCD to take place. 

Mikiko Bazeley: 00:43:08

Yup. Absolutely. 

Jon Krohn: 00:43:10

Cool. Then that leaves us with Kubernetes, which I think is more directly related to Docker containers themselves, right? 

Mikiko Bazeley: 00:43:17

Yeah. It’s a container orchestration tool. So when Docker came out, and it’s funny because I think Docker and Kubernetes, they were actually released, the first stable versions were released within the same year. I think people recognized pretty soon that if you have hundreds and millions of Docker containers all with different ports, all accessing different resources, you need some way to orchestrate it, and to also create pods, which they are the individual subunits, really, within a Kubernetes cluster. We need to figure out a way to actually scale up, scale down, allocate resources, all that good stuff. That is also something that you typically don’t want to do manually.
These are amazing technologies. They had to really be developed by companies that are solving big challenges around big data, which ultimately set the stage for solving big challenges using models that require big data. So these are awesome tools.

Jon Krohn: 00:44:19

Okay. So I don’t know the technical aspects nearly as well as you would, but I think the general idea here is that when a data scientist has defined some machine learning program that can run inside a Docker container or you could even have your entire web application running inside a Docker container, and maybe that web application happens to have some machine learning algorithms running in it, with Kubernetes, with this orchestration tool that you’re describing, you can then have a system that is automatically responsive to the load, “Oh, we have more and more users logging in than we were expecting in our baseline assumption,” you don’t need to wake up and start up some servers and have manually set some Docker containers running on those servers. Kubernetes automatically gets those servers up and running, gets the Docker containers running on those servers, and all of a sudden, bam, your web application can handle a thousand times or a million times more people than you’re anticipating when you went to bed. 

Mikiko Bazeley: 00:45:27

Yeah, absolutely. It is a crucial component of that. Now, it’s fun because the question of should data scientists know infrastructure like Kubernetes, should they have to know how to set it up and how to run it, that is a very polarizing question within the MLOps community because for a time, there was this idea that a data scientist should have to know so much of this infrastructure piece. They should have to know the end to end, especially if they call themselves an ML engineer.
I remember when some key blog posts around that came out and I thought about it and I went, “This is such an unreasonable expectation for data scientists.” So we’re asking them to be responsible for dealing with business partners, working with legal to ensure compliance. They’re also now responsible for future generation and engineering mall training for making sure that their data sets are clean, making sure their mols are performant, making sure they get deployed and selling. Now, they’re also responsible for the infrastructure to develop that when enterprise teams have full-on teams of five to 20 people staff to that.
So it’s nice to see that the pendulum has swung back to no data scientist should not have to ruin their live, focusing on a million things. Instead, it should be a collaboration of partnership between different teams with different skills and experiences to deliver on the same thing. I’m very happy that the industry has swung back to not trying to hurt people.

Jon Krohn: 00:47:00

Nice. So what would you say are the most essential MLOps skills that data scientists should know anyway even if they don’t necessarily have to? 

Mikiko Bazeley: 00:47:10

Yeah, absolutely. So number one, version control. That is so, so important. All of us have had that experience where something goes wrong in Git and we don’t understand what’s going on. You start stack overflowing, searching, and then you start randomly copy-pasting commands into your terminal. That’s just a nightmare thing. So I would highly recommend … So MIT, they had this missing CS semester course or set of lectures. I have to say that had one of the best lectures I’ve ever seen on understanding and rocking the spirit of Git and how it works, specifically version control in general.
I would say the missing CS course or semester set of lectures that has a lot of the key skills that data scientists would need in order to be, I would just say more effective, and I think that’s the most important part is just becoming more effective. So version control is a big part of that.

The second part is understanding essentially Python packaging or if you’re not using Python, you’re using different language, understand the packaging mechanisms for your chosen language is very, very important.
Third, containerization. I think that is something that, because at the end of the day, models do end up being code. So I think that is a crucial part of data scientists being more effective because now, for example, they’re not just relying on virtual environments, they’re not just relying on conda environments. They understand ways that they can share, reproduce code. That’s super important about that.
I would say then I’ll just give another two more. So we talked about version control. We talked about packaging. We talked about containerization. I would say the fourth part is understanding the tools that are useful for templating their projects. So Cookiecutter is an important one. I would say if people understand how to leverage it, then for example, if a data scientist want to create multiple projects and it’s a very similar structured organization, they could do that. They could set up Cookiecutter to do that to set up a consistent project structure so that they’re not reinventing the wheel.

Jon Krohn: 00:49:39

That’s cool. 

Mikiko Bazeley: 00:49:42

Yeah, I love using that tool. Then I would say the fifth one that’s really important is understanding at a high level how to leverage some of the cloud resources that we have. So whether it’s GCP, whether it’s AWS, and I’m not saying become a cloud practitioner extraordinaire, but understand that nowadays most models, they’re going to end up being shared. You want a model to be shared, right? You want to develop software that other people like using. So understanding in your chosen cloud vendor of choice what’s a minimum viable path to getting your model hosted and getting it shared so that you can either put a stream app on it, you can create a mobile or web app. Those are super important. Any other skills after that I’d say they are important, but those first five definitely I’ve seen block junior data scientists and I’ve seen them really struggle with those concepts. 

Jon Krohn: 00:50:42

Nice. I love it. That was such a great list, Mikiko. So number one is version control, tools like Git, and we’ve got the MIT missing computer science course, which I’ll be sure to look up and include in the show notes. Number two is packaging of software libraries. Number three is containerization. Four is templatization with a tool like Cookiecutter. That is one that I need to brush up on of these five, and number five is familiarity with cloud platforms like AWS and GCP. 

Mikiko Bazeley: 00:51:13

Oh, and the last one, and I think you had this author on your show at one point. He wrote Data Science at the Command Line. 

Jon Krohn: 00:51:20

Jeroen Janssens. 

Mikiko Bazeley: 00:51:22

Yeah. That is an under, underutilized tool. What I found was that for various projects where I need to set up data cleaning and processing pipelines, it was just so much faster to use a shell script, so much faster than trying to import Pandas and other kinds of libraries. Those were very, very slow for stuff that a lot of times could be solved through Regex and through some smart shell usage. So I would say that is the powerful skill right there. 

Jon Krohn: 00:51:58

The command line is hugely valuable. It can be a glue between whatever you’re doing on your Unix system, which is almost certainly your production system. Yeah. So Jeroen Janssens, he was on episode number 531. So you can get an introduction to this idea of the command line as a glue between all the programming languages and operations that you’d like to carry out on a given system. Yeah. He’s got his book, Data Science at the Command Line that you can check out as well.
Awesome. Mikiko, such a great list. That was a really valuable part of this episode. Thank you. So let’s move on here from this very specific technical advice related to what you’ve been doing in your current role at Intuit, MailChimp. Let’s talk about what happened before. So you transitioned very rapidly in your career from a data analyst to a data scientist, now to a machine learning ops engineer. So do you want to give us some color as to how that came about? How and why did you make these transitions so rapidly? 

Mikiko Bazeley: 00:53:09

Yeah, absolutely. So I would say when I graduated college, I studied in the humanities and I had initially gone in thinking I want to do med school. This is the Trinity. It’s the doctor, lawyer, engineer, right? 

Jon Krohn: 00:53:25

Right. 

Mikiko Bazeley: 00:53:27

I did not do that. Graduated college and what do I do with my life and how do I pay for roof over my head, all these questions- 

Jon Krohn: 00:53:39

“How do I put rent on the table?” 

Mikiko Bazeley: 00:53:40

Right. Exactly. I was like, “I need to get a job.” A lot of my friends, they were doing their internships or they had gone their first finance or tech job, and I was just like, “Oh, I just need someone to employ me.” So I feel like in especially the early stages, a lot of it was just me trying to pick up skills to be valuable. Eventually after certain point in time, I started to think more about, “Well, what do I actually want to do for my career? What is the body of work I want to be producing and contributing to?” Also, too, “What kind of work fit with the way I view the world and how I engage with it?”
So before my role at MailChimp, before that, I had worked as a data scientist for Teladoc, I was developing models. I was doing the analyses. I was trying to understand the business problems. I’d realized that, for me, the big questions were the bottleneck of getting my models out because I would create these, I would train these models in my Jupyter notebook or at that point, I was just starting to dip the toes into using VS code, and full-fledged ID that wasn’t Jupyter or Eclipse back when I was doing a little bit of Java. I was struggling so much with this.
I think this is a common experience a lot of MLOps practitioners shells.

You’re developing a model or you’re an engineer that’s trying to help data scientists developing models. You’re like, “Why is it so hard to share it and to get in front of people?” because at the end of the day, that is the power of it. It’s not fun when you’re just sharing a notebook and you’re refreshing. Your business partners, they get tired of that eventually.
So in about, what was it? So I joined MailChimp in May 2021. So I would say around, if I remember correctly, six months before that working as a data scientist at Teladoc, it was in the middle of the COVID quarantine. Companies were going through layoffs. Airbnb around that time had a huge layoff. My company, Teladoc, after Livongo had been acquired by Teladoc, they had been giving everyone a one year employment guarantee. My family was like, “You hold onto that job. Hold onto it. Hit that two-year mark,” but I was really thinking about it and I’m like, “I’m really not happy doing this kind of work. I want to be doing more engineering work.”
So I decided to quit. Well, it was funny, actually, it was before that.

So I had applied to open AI’s fellowship because they had this program where basically, “Come join us. Create cool applications and software. We’ll pay a stipend,” that was actually more than what I was making at Teladoc as a data scientist. I had applied for it. I’d put my heart and soul in the application and had gone just rejected. It was so crushing to me. I kept thinking, I’m like, “If I had that fellowship, I could change my career around,” and I didn’t get it.
At that point I’m like, “Okay. So what am I going to do about this?” I said, “Okay. Well, I will go do my own personal six months sabbatical.” So I quit my job, figured out a learning roadmap plan for myself to formally transition to MLOps, figure out what courses or workshops I want to take, and then got started on that path while also contributing side projects for different startups and all that around building ML pipelines. Eventually, I was fortunate to be able to make that transition. I got three offers out of that, out of that job search. What I love was that I didn’t do technical interviews for majority of those. 

Jon Krohn: 00:57:40

When you say job search, so you quit at Teladoc where you were working as a data scientist, and you then developed this six-month curriculum for transitioning into MLOps. 

Mikiko Bazeley: 00:57:55

Yeah, for myself. 

Jon Krohn: 00:57:57

Yeah. Was it near the end of it that then you started doing the search or were you doing this along? Okay. Yeah. 

Mikiko Bazeley: 00:58:03

Yep, yep. Yeah. I’m a big fan of trying to keep multiple options on the table. It’s really how I approach most of my career. For me, eventually, I knew that I wanted to transition into getting a full-time job, but I also knew that, and it was so hard because for me, I grew up in a family that wasn’t rich by any means. For me, money has always been this fear point of, “Do I have enough money to keep going? Do I have enough money to have autonomy in my career?”
So it was such a scary experience, but I knew it had to be done because I had tried for some time to do some projects and workshops and do my own learning around my work schedule. I had started interviewing around a lot. People said, “Look, you seem better as a data scientist.” So I knew I was getting some bites there. So I figured worst case scenario, if my personal sabbatical and my job search don’t churn out well, I could always interview and get an offer as a data scientist, but I really wanted to make that transition into an engineer. So I stuck with that, and I’m really glad I did. It was scary, but I think sometimes you do need to make that big jump for the long term and take some short term pain. 

Jon Krohn: 00:59:34

Nice. Well, it’s obviously paid off. 

Mikiko Bazeley: 00:59:37

Yeah. 

Jon Krohn: 00:59:39

So yeah, congratulations. A big risky step there to leave guaranteed employment and to have the confidence to create your own curriculum, which has obviously gone well, and probably that thing of creating your own MLOps curriculum, that has probably been helpful in developing now your curricula for other MLOps students to come. 

Mikiko Bazeley: 01:00:08

Yeah, absolutely, because when I had first developed that roadmap, I had included a lot of nonsense stuff on there, and I realized I had to hose about 60%, 70% of that roadmap plan and focus on the 30% that was most impactful, some of which includes the skills that we talked about earlier. Some also, which included more content around developing and designing distributed systems, working a lot with data, patching up models, understanding if you step away because a lot of what I was focused on as a data scientist was how do I get my one model out the door, and that’s just a totally different perspective from MLOps, where it’s literally like not how do we get one model out the door, but it’s how do we get 20 or 100 models out the door and how do we do it in a way that we’re also not locked in. So if we need to make changes to those pipelines or the data scientists want to just POC stuff, how do we support that innovation because a lot of times people say, “Well, innovation and stability and providing value, they’re totally mutually exclusive activities,” but I think there’s ways that we can try to meet in the middle. 

Jon Krohn: 01:01:25

Nice. So something related to your development as an MLOps content creator and MLOps practitioner and your career trajectory in general is that you strive to demystify machine learning careers for what you call non-traditional candidates. So what do you mean by that? What’s a non-traditional candidate and what are the biggest myths that aspiring machine learning practitioners or MLOps practitioners face when entering the field? 

Mikiko Bazeley: 01:02:03

Yeah, absolutely. So a non-traditional candidate. To me, a non-traditional candidate is someone who, first off, does not have formal study in the discipline or domain that they’re going into. They are candidates who also potentially come from families who don’t have a history of formal study in that area, and also who I would say are middle class or who essentially don’t come from upper class families. I consider my family to be lower middle class. So candidates who maybe don’t have all the resources and time in the world, candidates who are also job changers, they have other commitments, who essentially don’t fit this mold of the person who went through an Ivy league education, who came from a family that had access to resources and then followed this path of then going into engineering.
I am really interested in candidates who are very much so like myself, where you didn’t have family or you didn’t have that roadmap paved for you, and maybe you had to be really scrappy and maybe you still have to be really scrappy in both time and resources.

So those are the people that I’m very interested in helping out because I see a lot of myself in them.
In terms of the myths, and this came out, too, both in terms of my career working data analytics and data science, but also, too, as I was developing that personal roadmap for myself, the big one to get out of the way is formal study because MLOps and its parent, devops, they’re really about building tools, and a lot of that building is being done in industry. It’s not being taught in classrooms.
So on the one hand, having a formal CS background, it does help. I’m not going to say it doesn’t, but I don’t want to say either that it is a prerequisite for having a successful career as an MLOps practitioner.

The second myth that I want to take out is that you need to be a deep learning machine learning expert to be successful in MLOps. You need an appreciation and understanding of the workflows and the stages that data scientists go through, but I feel like a lot of that can be gotten from, first off, practicing empathy, talking to and working with data, understanding their world perspective and how they operate, and being able to do research. That is a huge part of it is I do so much research, especially that’s the hardest part in developing the courses is I’m going through, I’m reading blog posts, I’m talking to people, I am asking really stupid questions on forums that people are not to dunk me for. It’s a lot of that. It’s figuring out what is going on right now, and being able to consolidate that into actionable stuff. Empathy and research are much more valuable skills than necessarily being an experienced practitioner or working practitioner in machine learning. 

Jon Krohn: 01:05:29

Nice. 

Mikiko Bazeley: 01:05:31

I would say the third myth is that it’s all about the tooling. It’s really not. The tooling will come and go. I’d say we’re really at the beginning of the explosion of potential tools in MLOps. I really think it’s about the foundational knowledge and practice. I think that that is more important than being an expert in tooling.
The fourth one I want to just get out there is accreditation. I talked about you don’t need a prerequisite. You don’t need to have studied the field, but more importantly, I think I like certificates because they give me a structure for how to approach learning a specific topic or tool, but at the end of the day, it really is about can you build stuff. So I think credentials and certificates are really useful in the learning process, but I don’t think they will help as much as building actual software, building programs, building pipelines. 

Jon Krohn: 01:06:41

Awesome. Those myths, Mikiko, were crystal clear. So number one is that you don’t need a formal CS degree to get into MLOps. Number two is that empathy and research are more important than machine learning experience. Foundational knowledge is more valuable than specific tooling experience, and actually building software applications is more important than a formal accreditation. It makes perfect sense. I love those.
So alongside these general tips that you have for people getting into MLOps, which have been hugely valuable, you also, as part of your role at MailChimp, you are on the global engineering hiring committee. So through your vantage point on that hiring committee, what specific advice do you have for people who are looking to get hired into an MLOps role? What are the most important skills? 

Mikiko Bazeley: 01:07:42

Yeah, absolutely. So part of what I’ve done at MailChimp is I’ve helped to design and develop our interview process for some of our data engineering roles and also our MLOps roles, and also contribute, as you mentioned, to the global engineering hiring committee for MailChimp, especially the focus on values and culture, and that’s huge. We really care about culture ad as opposed to culture fit. We want people who it’s not just they’re cool to work with, but also they’ll continue to uphold the principles of diversity, equity, and inclusion because, honestly, that’s been a really special part of working at MailChimp.
First off, I’ve never seen so many female engineers under one roof as I … So MailChimp is an Atlanta-based company. When I was working in Bay Area startups and companies, I did not see that many female engineers under the same roof. On my team alone, we have at least four female MLOps engineers, which is great. My manager is female too, but it’s been so great.
Also, too, for example, in the Bay Area, you just don’t see that cultural and ethnic diversity that we also see at MailChimp, which is really wonderful. I like working with people who look like everyday people that I would interact with. It’s fantastic, people with different abilities, different perspectives and values, but all around this idea of you’re adding something.

So that’s a huge, huge part of what we care about.
So now to get into the actual specific parts of the interviews, so a couple tips of what I’ve seen, especially for engineers or data scientists, first tip is know yourself. So for example, I ramble when I get very nervous, which means I do really well with bullet points. Bullet points are my friend, and more importantly, having those stories framed out ahead of time. We see some engineers who they get really, really stuck in the details of the technical projects they worked on, and then they miss out on the ability to talk about, first off, what was the innovation that you brought to that project, what was the leadership that you engaged in, what were some issues that you proactively identified and then went after it.
There’s a lot of key values that we look for, and I think sometimes engineers they get so focused on the, “Okay. I’m going to talk about this really, really cool project I did where I implemented this microservice and all this other stuff.” It’s like, “No, no, you’re missing out on the human element on the ability to display the fact that, first off, you’re going to be in the trenches with us. You’re going to be delivering software. We can rely on you.” So I’d say that’s the first one is people should know themselves.
The second part should be understand what values, what are some elements that you are displaying in your storytelling. I would say the third part is humility.

So we’ve had some candidates where we asked them, “So when was the time that you really screwed up and you got negative feedback? How did you handle those?”
They’ll say, “Oh, we never got that. That didn’t happen. We’ve never screwed up. We’ve never gotten negative feedback.” You know what that says to me is actually not that, first off, and I could believe that for some candidates, but it tells me two things. One, you’re either lying or two, in some cases, “Okay. So you’ve never been put in a position where you’ve had to take ownership of something. You’ve never been in a position where you had to be accountable for your results. So if we put you, and right now we’re building a lot of stuff like plane in the air, right? We’re building a plane as we fly it. So how can we know that if we handle a crucial piece or element of a project that’s a little bit ambiguous that you’ll be able to dig away at that ambiguity and you’ll also be able to make a decision in terms of how to proceed with that project?”
So that accountability piece, and there is this book called Extreme Ownership. That’s almost an element of what we want is that ownership that a lot of times you do see in people with military experience. I think my friend Albert talked about this as well. 

Jon Krohn: 01:12:16

Albert Boime? 

Mikiko Bazeley: 01:12:17

Yeah, yeah, yeah, and Luke Brewers, too. They’re both former military, they have that experience, and they both talk about extreme ownership. Ownership is something that I’ve found a lot of military to have that innate understanding of, and I think that’s something that some candidates would probably do well to read up on and to really understand what that means. 

Jon Krohn: 01:12:41

Awesome. Really great guidance, Mikiko. So yeah, the key things that you look for in hires from your vantage point as a member of the global engineering hiring committee at MailChimp is to know yourself, to have a clear story, a clear narrative to convey, and humility. I love that you are looking for culture ad over culture fit. That’s a really nice way to describe what you’re looking for. I hadn’t heard that one before and it makes a lot of sense to me.
All right. So you do a ton. On top of your day job, we’ve talked about your MLOps course, your live workshops, your newsletter, your YouTube channel, the writing that you do for other companies like Nvidia. On top of all that, you also are a Crossfitter and a bodybuilder. So that’s something near and dear to my heart. If people follow me on social media, then every once in a while you get subjected to my latest lifting PR, personal record. So do you have any productivity tips or tricks for listeners on how to juggle all of these different things, day job, content creation, and fitness, and who knows what other hobbies you have out there? 

Mikiko Bazeley: 01:14:07

Yeah, absolutely. So something that I had picked up years ago, and this is similar to the … Have you heard of the power list, where it’s like pick your five things in a day, and if you finish those five things, usually it’s two or three one-offs that can directly contribute to your goals, and then maybe two or three habits that you’re working on. Then once you just call those, it’s almost like an example of Warren Buffett’s make list of your top 20 priorities, pick the top five, and then everything else, kill it with fire.
So I do something very similar. At the beginning of every single week, I draw a quadrant for myself that is based on my four key areas with fitness being not one of those quadrants, but it’s a must have. One quadrant is for my work goals. One quadrant is for my content goals. One quadrant is for my personal learning and development, and another quadrant is actually for … So I actually like making and designing clothing. So that’s my dream dream is to eventually have my own streetwear brand. 

Jon Krohn: 01:15:14

Wow. 

Mikiko Bazeley: 01:15:16

Everyone had their activity they did during quarantine. Some people did breadmaking. For me, I took a workshop on how to customize sneakers and how to make them. So that was my thing. I have much of Air Jordans right now that I’m just working on, just taking part of the shoe and buying different colored leathers and doing embellishments on it, rebuilding the shoe up. It’s really fun. 

Jon Krohn: 01:15:39

Wow. 

Mikiko Bazeley: 01:15:40

So I set those quadrants for myself and I say, “These are the things I must accomplish,” and then what I used to do is I used to be part of the 5:00 AM Club. I used to do my two, three hours ahead of time, ahead of the workday, get that out the way. Unfortunately, I work on East Coast hours and I’m on the West Coast. So for me to get two, three hours ahead of my peers would be really early in the morning. So I don’t do that for myself, but what I have done is I have, instead of doing two, three hours within the day, I try to listen to what my energy needs are. Instead, what I do is Friday, Saturday, and Sunday is when I work on my content projects.
I have my workshop right next to me. So if I need a break from work, I literally just churn around and then I just start working on my clothes, on my sneakers, learning new different techniques for creating heirloom stuff, but a more modern contemporary vision.

So that’s how I do it is there’s certain things I’ve determined are energetically intensive.
So working on work stuff, working on content is very, very energetically intensive, and other stuff where it’s a little bit more fun for me, I instead try to sprinkle it throughout the day or even at the end of the night so I can decompress and making sure that I have those boundaries because if I don’t have those boundaries, I become useless very quickly, including working out, and that’s a must have. So I have my own home gym. I have my power cage also in the living room, and I incorporate that. Usually, it’s around 3:00 to 4:00 PM when I’m just most tired and I’m most drained, and that’s when I have actually the best workouts. 

Jon Krohn: 01:17:30

Amazing. Well, Mikiko, thank you for all of the insights that you’ve provided today across MLOps and productivity, and just generally making the most out of a career wherever you start. So as we wind down episodes, I always ask our guests for a book recommendation. Do you have one for us? 

Mikiko Bazeley: 01:17:54

Yeah, absolutely. So right now, my focus is continue to develop my writing skills, especially for technical writing. So two books I found to be immensely useful are The Art and Business of Online Writing, and Documentation for Developers from Apress. Then in terms of like engineering books, I have a bad habit of just bouncing between stuff. So right now, I’m really enjoying Joe and Matt’s data engineering book that they released, as well as Visualizing Google Cloud, which is great, and also, The System Design Interview books by Alex Xu. 

Jon Krohn: 01:18:38

Awesome. Those are a lot of great recommendations from clearly an avid reader, an outstanding reader far beyond what most people are capable of. So I also then end episodes always asking guests for how we should follow them. I think we’ve got a pretty good sense of this from your episode already. We’ve got LinkedIn, Medium, YouTube, Substack, and even a little bit of Twitter. Is that right? 

Mikiko Bazeley: 01:19:08

Yeah. 

Jon Krohn: 01:19:09

Nice. All right, Mikiko. Well, thank you so much for coming on the Super Data Science Show. It’s been wonderful to get to know you on air and we’ll have to check in sometime maybe in a few years and see how your MLOps world is developing along. I look forward to checking out your sweet, sweet shoes at that time as well. 

Mikiko Bazeley: 01:19:28

Absolutely. It was so great talking with you and thanks so much, Jon. 

Jon Krohn: 01:19:36

Well, I hope you enjoyed that deep dive into MLOps with Mikiko as much as I did. In today’s episode, Mikiko filled us in on how MLOps is to data science like devops is to software development, how Docker, Kubernetes, and Jenkins work together as three of the key tools in the MLOps practitioners belt, how culture ad is more important than culture fit in prospective hires, and the six most essential MLOps skills for data scientists, namely version control, software library packaging, containerization, templatization of project structure such as with Cookiecutter, familiarity with cloud environments like AWS and GCP, and a strong command of the command line.

As always, you can get all the show notes, including the transcript for this episode, the video recording, any materials mentioned on the show, the URLs for Mikiko’s social media profiles, as well as my own social media profiles at www.superdatascience.com/599. That’s www.superdatascience.com/599.
If you enjoyed this episode, I greatly appreciate it if you left a review on your favorite podcasting app or on the Super Data Science YouTube channel. I also encourage you to let me know your thoughts on this episode directly by adding me on LinkedIn or Twitter and then tagging me in a post about your feedback. It’s invaluable for helping us shape future episodes of the show. 

 

All right. Thanks to my colleagues at Nebula for supporting me while I create content like this Super Data Science episode for you, and thanks, of course, to Ivana Zibert, Mario Pombo, Serg Masis, Sylvia Ogweng, and Kirill Eremenko on the Super Data Science team for managing, editing, researching, summarizing, and producing another wicked episode for us today. Keep on rocking it out there, folks, and I’m looking forward to enjoying another round of the Super Data Science Podcast with you very soon.