Getting Data Science / ML Certified across Azure, AWS and GCP

Recently between March and June, I had gotten Machine Learning and Data Science certified across the major cloud players - AWS, Azure and GCP. And I decided to blog about that experience for people considering stepping up their ML game to hyper-scale.

Why I wanted to get these certifications - Back in my undergrad days, I had taken special interest in getting myself trained in cloud technologies, just when it was starting to be a buzz word. And I had spoken out about this interest so much that some of my classmates started calling me "Sanjana Cloud". Fast-forward to today and it's not just a hype, but the present and future of how digital businesses are run. My career trajectory up to this point, has been along the lines of an applied ML Research Engineer and a Data Scientist. So getting these certifications was, for me - a great opportunity to combine these two valuable skill streams to be ready for the industry's future.

To folks who feel the same about these certifications, and are considering taking 1 or 2 or all of the following:

• Microsoft Certified: Azure Data Scientist Associate
• AWS Certified Machine Learning - Specialty
• Google Cloud Certified: Professional Machine Learning Engineer

I use this article to compare these certifications under 8 categories:

1. Ease of preparation
2. Affordability
3. Exam experience
4. Challenge the Data Scientist in you
5. Scope of improvement post exam
6. Post certification benefits
7. Which one should you take?
8. Where would I use each?

Now that I've revealed the whole content, I would also like to mention that these cloud providers and exam providers keep changing their exam content and format at intervals they find fit. So, some of the points I am providing now in July 2021 might not apply if you're reading this too far out. Still, I hope some timeless points in this comparison help you make informed decisions.

Ease of preparation:

The first step to taking these certification is learning about what these providers have to offer in end-to-end ML solutions. There's also the ML foundations and concepts to be strong on. But such materials can be found in various courses and online sources and doesn't really change from one provider to another.

So it comes down to how easy / accessible is it, to gain provider specific knowledge (theory and hands-on) on data pre-processing, exploratory analysis, modeling, deployment and operations. Azure was the most accessible as they provide a free learning path which is very detailed, has everything you need to know in one place and also lets you spin up assets to try out their offerings. They also offer free official practice tests.

AWS would be the next, where they provide this exam readiness course, which serves as an introduction and overview, but I definitely wouldn't say is exhaustive and a single stop for the exam. To be ready for the exam, requires in addition to strong ML basics, thorough knowledge on the ocean of cloud ML offerings AWS has. And being the oldest cloud layer, the ocean is wide and deep. Especially, the data engineering offerings took me a while to wrap my head around the number of alternate instances and services being built for the same overlapping use cases, with only minute differences.

For this, one might use external paid courses like the one offered by SunDog Education, but also make sure to go through the developer documentation across all offerings from AWS website. A trick I used to make my life easier with the developer docs was to use the Edit -> Speech -> Start Speaking (or similar functionality) provided by most browsers. This worked out for me as I can remember stuff I hear better than the ones I read.

GCP was the hardest to prepare for since the certification itself is very new and was also in Beta mode a while ago. If you look on their webpage for this exam, they seem to be providing a learning path, but it only points you to several external links and courses, a lot of which are paid. And hence, definitely not a single stop resource. There is this Coursera course which might serve as a single stop resource, but I did not try it as I was not willing to spend for it (didn't want to rush through the 7 day free trial either), but people willing to spend can try it out and let me and others know if it was useful.

So how did I prepare for the GCP ML exam? I relied on a lot of Medium blogs from past test takers. I should say that I did get side tracked from some of the blogs from Beta exam takers, since the beta exam ground and the current exam group happen to be different. I came across Sathish VJ's blog and used that as a guide to land at initial developer docs and navigate to related developer docs from there. And used the same Edit -> Speech -> Start Speaking technique again.

I can say this blog was a very good outline, but some of the links it points to may not be valid since GCP seems to keep rebranding stuff like AI Platform to Vertex AI etc. So whatever you read, you'll have to read with the old branding and new branding in mind as it's not clear which one is going to be asked about in the exam. GCP has an opportunity to improve themselves in this regard and provide a better exam readiness course, which is possibly also updated with each of their rebranding. I did also have access to the partner learning provided by GCP for a short period of 2 weeks through my employer, but that can use some restructuring too as it has a lot of repetitive information especially around AI Platform Pipelines and Kubeflow and is not organized very well under reflective headings and sections, in my opinion.

So for this category my rating is: Azure > AWS > GCP

Affordability:

Phew, now that we've gotten past the biggest task of preparing for the exam, let's sign up to take the exam and obtain the certification.

The affordability is as follows: Azure > GCP > AWS

While Azure costs about $100, GCP cost $120 and AWS cost $300 (this is all before tax). And one should consider validity along with cost too. While both Azure and GCP ML certifications are valid for 2 years, AWS is valid for 3 years. Even with that one extra year, AWS costs 2x of Azure and a little above 1.5x of GCP on a per year cost.

This is when we are considering only the ML exams against each other. But, one more thing about AWS is that, if you can take another cheaper AWS exam first, you get a 50% off for the ML Specialty exam and also get a free practice test. And that can bring the cost and investment on this single exam down a bit.

Exam Experience:

This includes experience during the duration of the exam and up to the point you receive results.

Both Azure and AWS are 3 hour long exams with Azure having a varying number of questions (up to 80) and AWS having a fixed 65. The exam provider I used for both is PearsonVue and while I noticed a calculator being present for Azure, did not find it for AWS and did not end up requiring it either. GCP has 60 questions and a duration of 2 hours and is provided through the test provider Kryterion. Between PearsonVue and Kryterion, I liked Kryterion's UX better.

One might be wondering how someone can manage GCP when it has nearly the same questions as AWS but provides an hour less. I would say GCP is ideally timed since even 2 hours is a lot of time to be looking non-stop into a system screen (with even looking away from screen being looked at as a sign of cheating in remote-proctored exams during the pandemic). Besides, for both AWS and Azure I ended up having half an hour and an hour extra respectively. GCP was the one I ended right on time.

When it comes to variety of questions Azure had a greater variety like fill in the blanks, multiple choice, multiple response etc. and that variety removed the monotony associated with these testing experiences, at least for me. AWS had the longest questions - took me longer to figure out the question than the answer. Maybe they were trying to teach us the life skill of surviving meetings that should've been an email. And since I took GCP right after AWS on the same weekend, seeing direct questions on GCP was a relief. And partly that's what enabled finishing almost the same number of questions on time, even with one less hour.

After finishing the exam and submitting it, you will see a Pass / Fail status on the test software for all 3 exams. For Azure, your scores are computed and displayed right away. For AWS, you get the score only later, when you receive your certificate. For GCP, there is no score provided at any time. Since these testing softwares kill and disable all other apps in your system, they don't really allow you to take a screenshot of that momentary flash of your result. With that being the case, speed and mode of delivering proof of result plays an important role.

For Azure, I got the final result delivered in my email inbox within an hour of taking the test. For AWS, you do not get preliminary results displayed or provided anywhere and it can take up to 5 business days to get the result finalized and get your certificate, scorecard and badge. For GCP, it can take up to 10 days but you will see the preliminary result displayed on your WebAssessor candidate profile (the same one you used to register for the test).

I am guessing both AWS and GCP take time to rewatch your remote proctoring video to ensure there has been no cheating. Kind of invalidates the presence of the proctor, but I much prefer GCP's way of providing proof of preliminary results first and then taking their own sweet time to validate your result, over AWS's method of causing unnecessary anxiety to test takers who have no proof of their result, even if it's for a few days. Imagine the amount of effort and money put into this certification, only to cost you more anxiety and stress after successfully completing it.

So for this category my rating is: Azure > GCP > AWS

Challenge the Data Scientist in you:

This I think is inversely proportional to the ease of preparation. Because, if everything is handed to you, what's the challenge in it. It is also in line with the type of questions and the racing against the exam clock. While I did say that AWS questions were really long, I feel like that's more of a test on someone's reading comprehension skills (something my colleague also mentioned) rather than pure Data Science and analytical skills. GCP I found was testing more on core Data Science and applied Data Science skills in a more direct manner and also testing how well you do under pressure by being the shortest of the exams. If you're someone who likes to challenge yourself,

The rating is: GCP > AWS > Azure

Scope of improvement post exam:

One can improve only when there's feedback. And that is something you definitely get with the AWS certification. You get scores plus section wise rating on whether you performed well or whether you need improvement in that section. Azure also displays score but no detailed rate-card. It still helps assess how far you have to go to reach the goal you've set for yourself. GCP however does not provide any scores or feedback, so it's hard to assess where you stand.

For this category my rating is: AWS > Azure > GCP

Post certification benefits:

Congratulations, you've come this far, you feel like you challenged yourself and proved to yourself that you're a good Data Scientist, and one who knows how to do ML at scale in that. But does the journey have to end just with that feeling? This is where post exam community and benefits come in. I did not find any for Azure and hence an area that Azure has an opportunity to improve in.

GCP maintains a Google Cloud Certified Directory which maintains a focussed list and view of all the GCP certified professionals. AWS provides you multiple benefits like 50% off on the next exam, free practice test and also the ability to join their unlisted LinkedIn group. Between the GCP Directory and the AWS LinkedIn group, I found the AWS one to be more active (since obviously it's a group with lively people rather than a directory) and found the GCP one to provide a clean directory with links to people's profiles, credentials etc.

Merch!!! Let's talk about merchandise. Both AWS and GCP let you get your hands on really cool merchandise post certification. But the AWS one costs and is not customized to your specific certification - it just says "AWS Certified". GCP provides limited but free merchandise and since it's free they are often in back-order I believe. It's also awesome that they are customized to your specific certification - for example I ordered (back-ordered) a tumbler and a notebook saying "Professional Machine Learning Engineer" and my dad has called dibs on the notebook already.

Taking all this into consideration, I rate: AWS > GCP > Azure

Which one should you take:

Obviously you should get certified in the current cloud infrastructure you have access to and are using, or is in the plans to be subscribed by your group or employer.

But if you are just getting started and would like to have one Cloud ML certification, I would suggest Azure. From the well designed and comprehensive learning path and practice tests to the exam experience, it feels like Azure cares about the candidates and wants them to succeed. Hence, Azure is definitely the best one to start with.

Where would I use each:

While the ML basics and the end-to-end core components of an ML system do not change, each of these cloud providers have their own offerings and take on how such a system comes together to function. Having gone through all 3, these are the unique roles I see for each (purely personal preference):

• Azure: I would use when I have a quick deadline and want a well-guided approach to developing my ML solution at scale and it has to be very user-friendly. I do not have the time to get distracted nor experiment with multiple options. I want a quick, clean and efficient prototype - I will choose Azure.
• AWS: I would like access to multiple options and flavors enabling the core components of my ML system. I would also like access to proprietary algorithms with years of research behind it and be able to use them right off the shelf. And I want to do this while taking advantage of multiple savings options available for my infrastructure - I will choose AWS.
• GCP: I would like to bring in a lot of my own ML models and I want to experiment or research with combinations of custom and off the shelf tools and be able to do this at scale. I also want full control of my deployments and want granular visibility into cost - I will choose GCP.

Summary:

There goes the full picture I had to provide and I wanted to close it by saying - irrespective of which one (or all) of these certifications you prepare for and take, and irrespective of the pass / fail / score, the experience will get your ML foundations strengthened and make you ready to apply those foundations at hyper-scale. So I encourage people to definitely add these certifications to their learning goals. Happy learning!