So, I recently got certified as a Google Cloud professional machine learning engineer. Now, how did I do that, and for that matter, why did I do that? Well, keep watching 'cause I'm going to tell you all about it. I'm Richard, and this is Richard on Data. So, yeah, there's this piece of paper out there that claims I know a thing or two about machine learning engineering on GCP—that is, the Google Cloud Platform. I'm going to start with why I wanted to do this thing in the first place and what this exam even is. Then, I'm going to get into exactly what resources I used, how I prepared for this thing in the days leading up to it, and even some of my impressions of the actual test once I started taking it. So, watch to the very end for that. Before I do that, take just one moment to smash the like button and subscribe to this channel if you haven't already. All right, so as you all know, I'm a data scientist. So, why exactly did I want to become a certified professional machine learning

engineer well. Part of it for me had to do with upskilling. A few years ago, all I knew about working in the cloud was laying in the grass and pointing up at things and saying, Huh, that kind of looks like an elephant. I'm exaggerating slightly, but not by much. Frankly, Cloud platforms aren't the only way that you can deliver production data science solutions, but they greatly expand your capabilities to do so. Now, at my company, I had recently started extensive work in the Google Cloud platform space, where there was an explicit expectation that my team was going to be able to deliver solutions. And not only that, just for the sake of my own professional development, I just wanted to be as competent and confident as possible in the devops and production side of solution delivery. So, I set a goal at the beginning of 2023, which was, even with all this crazy personal stuff I was doing, like planning a wedding and buying a house, I would prepare for this exam, take it, and pass it before the end of the year. And just

to give some more full context on my background here. I was far short of the expertise that Google Cloud recommends for this. Google Cloud recommends about one year of expertise with their platform before you take this thing. Well, I had worked on a migration project from Microsoft Azure to GCP for about four or five months before I took this exam, and I really didn't have that much more expertise hands-on beyond that. Now, that was incredibly valuable experience that I definitely needed. So, it's one thing to be on the internet somewhere and read about the difference between, let's say, Cloud Run and Kubernetes pipelines. It's a completely different experience when you're sitting in a role and you need to make some kind of decision about the best solution for orchestrating a pipeline that has to execute once a day. So, all that is to say Google Cloud's recommendations are probably a little bit exaggerated, but you definitely would want to come into this exam with some experience. So, now that we've established why I wanted to take this thing, let's talk about what the exam

even is so. The Google Cloud professional machine learning exam is a 60-question exam that you have 120 minutes to take, and it's all multiple-choice questions. Now, the best summary of this thing is its official exam guide, but keep in mind I took this exam in November of 2023, and I'm recording this video in December of 2023, and the exam has already changed since then, and it's almost certainly going to change again soon to reflect generative AI since that's such a hot topic right now. So, let's briefly talk about what's on this exam and, arguably, even more importantly, what's not on it. First of all, there's almost zero coding questions on it. I think when I took it, there was one question with some SQL code in it, and that was pretty much it. Also, since I took it, there seemed to be a lot less questions about classification metrics, so things like precision, recall, PAU, F1 score, all that good stuff. When I took it, there were a lot of questions about how to frame a problem, like is ML even a good solution?

for this problem. Is this a good case for a regression, or clustering, or classification, or recommendation engine solution? I'm not totally sure on this, but just from perusing the exam guide now, it looks like those kinds of questions are significantly scaled back. Instead, you do get a ton of questions about AutoML, pre-trained APIs, and which one is better. There's a lot of specific questions about those solutions and different types of data, data pre-processing, some stuff on PII and PHI. Now, I don't remember when I took it anything about Jupyter notebooks, but I guess that's in there now. Then you get into section three, which is all about scaling prototypes into ML models. Now, one point I want to call attention to is this piece about appropriate hardware. I remember more than a few questions on the exam about CPUs versus GPUs versus TPUs, as well as distribution strategies. Then moving on to section four, there's a ton of questions about the distinctions between anything served online versus with batch serving. There's a big topic here around scaling served models. Then there's a ton of questions around

orchestrating pipelines. This is where knowing the difference between these different tools is really helpful. For example, Cloud Build is for continuous integration and continuous deployment. Cloud Run is for running containerized applications. And I remember seeing lots of questions testing your knowledge of distinctions like that. And then under this last section six, this is another section I remember seeing a lot around. You've got questions around bias and fairness. You get questions about training, serving skew, and feature drift, and a few about what even makes sense as a baseline for monitoring model performance. So now that we've established what is and isn't on this exam, how in the world do you actually prepare for it? This is the hard part because there isn't one single comprehensive One-Stop shop that you can go to prepare for this exam. And even if there were, it would have to be updated constantly. But I'll detail all the resources that I personally used. In total, I used one book, some Medium articles, some Coursera courses, some practice exams, and Chat GPT. First things

first, I bought this book, Journey to Become a Google Cloud Machine Learning Engineer. The link to that will be in the description of this video. Honestly, there were a few chapters of the book that I just kind of skimmed over, but the content around developing and deploying ML models and the content around neural networks and T Flow and Keras is excellent. Same with this whole chapter on Vertex AI. From looking at the new exam guide, Vertex AI stuff seems more and more important, so that's another great one to know really well. Then, perhaps most importantly, there's also a practice exam in the back with 30 questions. I found these to be excellent questions and to match the difficulty level of the questions on the actual exam pretty much perfectly. Then, sometimes it's helpful to hear perspectives from real-world people who've taken this exam too, so I read several Medium articles from people who had passed the exam where they laid out their recommendations, their approaches, and their opinions, and it's really helpful just because with an exam with this much content on it, it's useful to

get some kind of feel for what you should prioritize and what topics seem to come up the most. A lot of these articles are going to point you to Coursera as the backbone of any sort of prep that you do. Now, I'll admit I'm not the world's largest fan of Coursera, and the courses for this particular program vary pretty wildly in quality. I did find some of them pretty dry, but more of them than not were at least somewhat worth doing. I think the most important ones for me were Enterprise machine learning, production machine learning systems, and pipelines on the Google Cloud. I found these courses to be both highly instructive and highly relevant to the actual content on the exam. So, what I would usually do is put the video lectures on 1.25x to 1.5x speed. I would go through them and take notes on some key concepts, and I would also go through the labs pretty quickly. Granted, I could have done some more and gotten more out of these labs, but since I already knew there was

pretty much no coding on the exam. I was more focused on just making sure that I had the key concepts down and essentially moving on to the next thing. I'd say the most valuable content from these courses is the quizzes. These really help to drive home key concepts from the courses, and a lot of them were pretty comparable in quality and difficulty to exam questions. So, for me, it was about one week or so before the exam, but at some point, it's absolutely critical to test your knowledge and start taking practice exams. For me, I also started going over a lot of the quiz questions from Sarah again too. There's a few different practice exams floating around out there on the internet, but a couple that stand out to me are the one in the journey to become a Google Cloud machine learning engineer book and Google's own sample questions. The interesting thing about Google's own practice exam is that I actually found the average difficulty of these questions to be significantly higher than the actual exam questions. So, let's go over a couple of these questions.

Here's one example. It's about you need to develop an online model prediction service that accesses near real-time features and returns a customer churn probability value. The features are in BigQuery and updated hourly. And now this is really important. It says the service needs to be low latency and scalable and require minimal maintenance. So you get all these different possibilities here. You've got you could use Memory Store, you could use Vertex AI Feature Store, you could use Google Kubernetes Engine, you could use a custom prediction endpoint in Vertex AI. And once you start looking at the answers, you see the problems and the distinctions between some of these. But basically, the Memory Store is going to be lower latency than the Vertex AI Feature Store. And you don't want to use the Google Kubernetes Engine because it's just going to be higher maintenance. So the question's really testing your knowledge of these sorts of distinctions. Now here's another sample question, and this one is all about PII and privacy issues, which is a pretty hot topic on the exam. It says you've developed a tree.

model based on a feature set of user behavioral data. Now, there's new regulations that require anonymizing PII. Uh, you know that that exists, and you want to update your model pipeline to adhere to the new regulations while minimizing drop in model performance. What should you do? And the keys here are you end up having to train the model from scratch because if you just tune the existing model, it's not recommended because the original training might have memorized sensitive information, and you don't want to use something like masking because it doesn't enforce referential integrity, and it's probably going to drop the performance of the model again. Just the kind of things that the exam tests you on that you need to know. What I ended up doing is taking the exams over and over again to a point where I basically memorized the questions, but the point is you have to be really attentive to the specific verbiage in these questions, and you have to be intimately familiar with the differences in different technologies. But over time, you're going

to start seeing some patterns show up. I remember a number of questions talking about accuracy on the training set was like 99%. So, it turned into pretty much a dead giveaway that there's a problem, and that problem is data leakage. Now, this is where I started using chat GPT a little bit in the days leading up to the test to really quickly summarize and iron out the differences between these Technologies and Concepts. Here's an example where I ask chat GPT what's the difference between data flow and data proc. And here's a hint: you need to know this on the exam because there's a lot of questions about it. And in the response, there's a few key points here. Number one, data proc is used for the deployment of Apache Hadoop or Spark Frameworks. It also doesn't handle native support for stream processing, but data flow excels at that. So, that's one example. Here's another: I didn't know as much about TensorFlow distribution strategies, so I asked for a summary of the difference between mirrored strategy, multi-worker TPU, and parameter server.

and this was a super amazing summary, which was of great help for me when I was getting ready. I've never personally been really big on creating flashcards, but I know a lot of people are, and this is kind of a great application for that. So then, I finally got to the test day, and that was an experience. I took the exam on my own personal computer in an online proctor environment, and the exam ended up getting delayed by an hour and a half due to weird technical problems. So, we were already off to a great start. Then, when I finally got to take the actual test, I got to the first two or three questions, and I thought, Why exactly did I do this to myself again? But by the time I got through every single question and then went back over everything and reviewed, I realized I was pretty much 100% sure on roughly half of the questions, and for a lot of the rest of them, I at least had a pretty good guess. There were probably about 10% of the questions which I had just

absolutely no idea about whatsoever, and I just put B for all of them. Now, the way the exam results are set up, it doesn't break down what you got right or wrong, and it doesn't even give you a final percentage score. My best guess as to why that is is because a lot of the questions ask for the best way to do something, so multiple options are technically correct, but some are better than others. So, Google might be allotting partial credit for some of the answers. It's just a theory; it's really impossible to say. All of that to say, though, it felt really good, and frankly, really surprising when I submitted the test, and then not 2 seconds later, I got back this screen that said result pass. So, to recap, that was my experience with the Google Cloud professional machine learning exam, and to sort of summarize here, the exam is really hard, but all of this is a pretty doable process if you're a person with some data and machine learning exposure and just a little bit of cloud experience.

you can get through it. My final comment would be if you put the time into this thing and be smart about how you're using your time. You take advantage of all the different resources that are out there, and you test your knowledge. You too can get a passing score on this exam. So thanks for watching this video. If you enjoyed it, hit the like button. If you hated it, I guess you're more than welcome to hit the dislike button. Then leave me a comment down below and let me know your thoughts. Do you have any questions about my experience taking the exam? Are you interested in taking it yourself? Let me know. Then I'll see you all in the not-too-distant future. Until then, Richard on data.

How I Passed the Google Cloud Professional ML Engineer Exam

How I Passed the Google Cloud Professional ML Engineer Exam