How I passed the Google Cloud Professional Machine Learning Engineer exam (Vertex AI)

For those unfamiliar, the Google Cloud Professional Machine Learning Engineer (PMLE) certification is an exam designed by Google to certify one in the following skillsets:

• Frame ML problems
• Architect ML solutions
• Design data preparation and processing systems
• Develop ML models
• Automate & orchestrate ML pipelines
• Monitor, optimize, and maintain ML solutions

This certification is relatively new (announced in November 2020) as compared to the related but more established Professional Data Engineer (PDE) certification.

Beginning 22 February 2022, the syllabus of the PMLE exam has been updated to focus on Google Cloud’s Vertex AI, a unified platform for the entire ML workflow instead of the older AI Platform.

I sat for the PMLE exam on 26 February 2022. This post details my experience preparing for the exam with the updated Vertex AI syllabus and offers some advice on how you can strategically prepare for it too in less than two months.

For a more holistic and detailed view of the exam and topics, please refer to this Github repo that has a compilation of useful resources for the PMLE exam.

My preparation approach:

A little background: I’ve been practicing data science/ML in a financial institution in Kuala Lumpur, Malaysia for about three years . However, this was my first cloud certification which is why I fumbled around a little as I started my learning journey.

I commenced my preparation effort with the official learning path recommended by Google, taking courses and doing quests according to the program. Soon, I found myself misremembering the various Google Cloud services and confusing them with one another (e.g. Dataproc, Dataflow, Dataprep, Data Fusion, etc.) as the course went deeper into the weeds of technical details. I started getting frustrated midway as I wasn’t sure if I was on the right track or if I’d gone down the rabbit hole. I hopped around from Google Cloud documentation to how-to-guides (like this one!) contributed by the community and discovered Whizlabs as recommended by Rolf Siegel. The Whizlabs practice exams, while imperfect, gave me a better yardstick to understand the depth of knowledge required for each topic. Coupled with Google’s official sample exam questions, they allowed me to calibrate my efforts when studying the relevant materials which drastically reduced my overall preparation time.

From a logistics standpoint, I’d initially planned to complete the exam on 21 February 2022, right before the syllabus was slated to be amended to reflect the Vertex AI update. However, due to some scheduling error, I wound up having to push my exam date back a week. I used the extra week to learn more about Vertex AI and unlearn what AI platform had to offer. In totality, it took approximately 2 months for me to prepare for the exam with intermittent studying after work hours and on weekends. However, that duration can be reduced if you combine some machine learning experience with a more concerted study effort.

In terms of the Vertex AI syllabus update, the Vertex AI technology stack was quite similar to that of AI Platform. I don’t think the Vertex AI syllabus update changed the PMLE exam drastically. This is because many of the underlying components were largely the same, just shuffled around with tighter integration alongside other Google Cloud services. Aside from understanding the differences between AI Platform and Vertex AI, I took some time to read up on the new services offered in Vertex AI such as Feature Store. In short, if you’re familiar with AI Platform but unfamiliar with Vertex AI, you have little to worry about. The following diagram nicely summarizes the new components moving from AI Platform to Vertex AI.

https://github.com/GoogleCloudPlatform/training-data-analyst/tree/master/self-paced-labs/vertex-ai/vertex-ai-qwikstart

The preparation approach I would recommend (in sequence):

The following are the steps I would’ve taken if I’d known what I now know. They’re broadly the steps that I’d taken but stripped off of unnecessary ones and with some steps reorganized to maximize preparation efficiency.

1. Sign up for Google Cloud Skillsboost (30 days free), complete as many of these quests as possible. I recommend jumping straight into the quests/labs at the beginning and not the courses because I feel that the courses go into too much depth about technical details that may not be relevant for the exam. You can complete the courses later if you wish to deepen your understanding. However, for a start, the quests provide a sufficient introduction to the Google Cloud ecosystem such that you understand how the various technologies fit together in the ML workflow. You may also wish to skip the challenge lab at the end of each quest if you’re pressed for time.
2. Review the sample questions from Google Cloud. This should allow you to at the very least, partially calibrate the amount of effort required for each topic that you study. If you find them insufficient, skip to step 6 and attempt the first practice exam.
3. Read GCP documentation. Now that you’ve gotten an introduction to the various GCP services/tools from Skillsboost quests and have, focus next on reading documentation. I find that they provide the necessary depth (but not too much like the courses do) to ace the exam. Don’t focus too much on the code but pay attention to configurations and settings available in the tools. Follow the links sprinkled throughout the key documents below and branch out to learn more about the various services/tools. I think that the following documents tie in the various Google Cloud tools and concepts nicely. Pay special attention to Dataflow, BigQuery, and pipelines (both TFX and Kubeflow).
   – Best practices for implementing machine learning on Google Cloud.
   – Best practices for performance and cost optimization for ML.
   – Architecture for MLOps using TFX, Kubeflow Pipelines, and Cloud Build
4. Spend some time focusing on Vertex AI. Since Vertex AI is the primary platform to run ML workflows in Google Cloud, you may want to reinforce your understanding of the platform with the following resources:
   – Watch the YouTube series on Vertex AI for a high-level overview/recap.
   – Study the various ML workflow implementations in the Google Cloud Platform Github repo on Vertex AI and in the codelabs. Note that the tutorials in both sources may overlap.
   – Read through the official Vertex AI documentation end-to-end.
   – If you went through step 1, you would’ve completed the “Build and Deploy ML Solutions on Vertex AI” quest. Else, it’d be a good idea to work on it now.
5. Skim through Google’s Machine Learning Crash Course. Even if you already have a working knowledge of ML, I would still recommend running through the material to understand the way Google does ML. Having said that, only a few of the exam questions touch on ML techniques. Your attention should be spent on the usage of Google Cloud software and MLOps.
6. Purchase and attempt the practice exams from Whizlabs. I recommend taking the practice exams earlier (perhaps a week before) to identify areas where you need to refine your knowledge and dive deeper. However, the questions are a mix of topics from the new Vertex AI and the outdated AI Platform. Hence, I wouldn’t give too much attention to questions about AI Platform but focus on how Vertex AI can replace AI Platform under those specific scenarios. As of this writing, their practice exams consist of two practice tests with a total of 110 questions (55 questions/exam). Whizlabs includes 15 free practice exam questions but they’re just questions extracted from the first paid practice exam. Also, note that the actual PMLE exam has 60 questions, but that shouldn’t be a big issue. While the practice exams were beneficial overall, I wouldn’t consider the scores obtained from the practice exams to be indicative of the score you would obtain in the actual exam. Some of the questions, answers, and solution explanations from the practice exams were poorly phrased which I felt didn’t accurately represent the types of questions from the actual exam. Finally, a quick google search would get you a Whizlabs 50% discount code, reducing the cost of the practice exams to $15. While I did purchase the Whizlabs course video, I didn’t gain much value out of them and wouldn’t recommend you to do so. Much of the video content is a rehash of Google Cloud documentation while glossing over important implementation details.
7. Read more GCP documentation/tutorials. If you have extra time, go through as many of these as possible and/or these courses.

Conclusion

I sincerely hope that the sharing of my exam preparation experience would help some of you in saving time and avoiding the mistakes I made while navigating the process. Just remember not to dive too deep into any one topic (save a few key ones mentioned above) but understand how the pieces fit together. The bulk of your effort should focus on steps 1, 3, 4 detailed in the preparation section above. Finally, as with all ML workflows, I’d suggest performing all the steps iteratively in a cycle. Rinse and repeat. You’ll be fine. Good luck!