The Google Cloud Professional Machine Learning Engineer (PMLE) exam is one of the most challenging certifications on the market. The following is my account of how I prepared for and passed the PMLE exam in late 2021. Let me start by saying that it took me about three months to prepare for the exam. But I also did some unnecessary things, so I hope this article will help anyone interested to prepare for the exam more efficiently.


In the beginning I did the example questions at

Unfortunately these are only 11 questions (the real exam has 60 questions) and they don’t change even if you do them again. You don’t get a score at the end, so you have to calculate it yourself (to pass the exam one needs 80% correct answers). But you get a first impression about the type of questions.

I then followed the learning path for the “Professional Machine Learning Engineer” on . But I realized that it will take a very long time, if you really want to do all the suggested courses. Especially doing the hands-on qwiklabs is a bit of a hassle. For every lab one needs to setup the VM, clone the git repo start the Jupyter notebook, execute many BiqQuery SQL queries for data preprocessing and repeating this over and over will take much of your time. In addition I encountered quite a few bugs (e.g. access rights) and struggled with incomplete documentation and outdated tensorflow versions, which many labs still required. Also the code TODOs in the labs are often impossible to do, because the information for the task given is simply not enough. I often just copy pasted the solution from the solutions notebook to go on. So all in all I would recommend to everyone: Skip the labs. They will not help you to pass the exam at all.

Fortunately I then found the following post from a Google Sales Lead:

He recommends to do only the following courses:

  • Launching into Machine Learning
  • Introduction to TensorFlow
  • Feature Engineering
  • Art and Science of Machine Learning
  • End-to-End Machine Learning with TensorFlow on GCP
  • Production Machine Learning Systems
  • ML Ops Fundamentals

And he suggest to skip the following courses (which I did, unfortunately I already had done the first two before I found the article):

  • Google Cloud Big Data and Machine Learning Fundamentals
  • How Google does Machine Learning
  • Image Understanding with TensorFlow on GCP
  • Sequence Models for Time Series and Natural Language Processing
  • Recommendation Systems with TensorFlow on GCP

In addition he mentioned the company Whizlabs, which offers example questions and test exams. Since after doing all the courses I still felt quite unsure about my knowledge I decided to purchase the practice test and the video course from

There are some issues, though:

  • The company claims to sell you 125 questions, 2*55 and 15 free example questions. But it turned out the 15 free example questions were included in the first test exam with 55 questions. So effectively you will only buy 95 questions.
  • The questions turned out to be too recent. They were often about Googles new Vertex AI service. However in my exam only the old AI Platform was mentioned. This problem is resolved nowadays, because Google changed its syllabus for the PMLE exam in Feb. 2022, but it was still an issue in November 2021.
  • The sample exams have many questions where one needs to select multiple answers. This does not reflect the real exam, where you will always only select a single correct answer.
  • The video course is not of the same quality as the Coursera/Google lectures. In the end I didn’t watch them.

But I worked a lot with the two example exams. These are done online and you will see a timer running, so it gives you a good hint on your answering speed. In the end you will get a nice report, with your score even grouped by topic. Especially nice is that you get also explanations of the answers and links to further material to look up for the topic. What you can’t do though is copy-pasting the questions, answers and explanations.

For making most out of the example questions I did the following: First I took the first example exam and scored about 60%. I then went through all the questions (especially the incorrect ones, but also the ones I just had guessed correctly) and made handwritten notes based on the given explanations from Whizlabs (sometimes simply copying the sentences, sometimes following the references and writing a short summary). After doing that I took the first exam a second time and scored > 90%. Then I did the same with the second exam. So in the end I had about 20 pages of handwritten notes and had scored >90% in both example exams. This gave me some confidence that I had at least learned something (although as a good data scientist you know that the 90% score the second time is basically overfitting on the example questions ;-) ).

After that I registered for my real exam for 200$ plus tax. Until the exam I tried to read and memorise my handwritten notes. I also looked again at the blog post of the Google Sales lead and some blog posts listed at

to identify further topics I had not learned very well (e.g. Tensorflow distributed training strategies). I made handwritten notes about all these topics, which at least gave me a good feeling.

All in all I would recommend to do at first the Google example exam to get a first impression. But then you should probably go and buy the Whizlab example questions and work through the explanations making handwritten notes. After that you can still watch the Google/Coursera classes (without doing the labs), but then you already know which topics are of interest for you. If I would do it again I would definitely try to make some handwritten notes from the videos in addition.


I took the exam offline at a test center in Berlin. Some other colleagues had struggled with online exams for other cloud providers, because workplace PCs didn’t allow to install the necessary software, so I decided to do it offline. I came half an hour early and was allowed to immediately start the test. I had to sign documents, that I agree to video surveillance during the test.

  • The exam took 2 hours and consisted of 60 questions.
  • Except for one question all other question were different from the test exams I had seen before.
  • Many questions had roughly the following schema:

You are a Lead Data Scientist in a company A. Your team has done B to solve task C, but has (encountered difficulties D | wants to improve performance | wants to move to Google Cloud | etc). What solution do you propose, if you (want to use as little code as possible | get the highest performance | save costs) ?

  • Most questions were about how to combine several services from Google Platform in the right way. The technologies mentioned most are BigQuery ML (Often the right simple answer), AI Platform, Tensorflow, Kubeflow (Often the wrong overcomplex answer), Pub/Sub. Firebase was mentioned once. No questions about Vertex AI. Some questions were purely data science question, often about precision/recall/accuracy, ROC-AUC, hyperparameter tuning/over-optimization/regularization or the correct way to do cross validation and feature preprocessing.
  • There were always 4 answers given (A, B, C, D) and you had to select exactly one.
  • I didn’t have a very secure feeling then answering the questions. Many times you can “guess” the right answer by searching for certain keywords in the question and checking if the answer does mention them, too. Or you first exclude the ridiculous answers and then make an educated guess on the reasonable answers left. Sometimes it also helps to think about what Google would probably like to hear. At least for the pure data science questions there was often a definite correct answer.
  • I needed roughly 1:30 h to finish the test. Then I used the ReviewAll Button, jumped again to the first question and by selecting next again and again went through all the questions a second time. This took all my time I had left, but in the end I managed to review all the questions (but skipped very quickly through some of them) with a few seconds left on the timer.

In the end the exam software showed in a very small window only the words: PASSED. No grade, percentage or other evaluation was given. The result displayed is also only “provisional” as the person in charge of the test told me. The final result was emailed to me by Google a few days later. So don’t expect a champagne shower when you pass the exam. But you will be rewarded with a link to a website with free Google Merchandise products. So you can order a nice mug to prove that you really have passed this challenging exam.



Andreas Maier

PhD in Astrophysics, currently working as Senior Data Scientist & Machine Learning Engineer