Stress Test the Cloud: Alibaba Cloud, AWS, Azure, GCP

For example, a large component of "reliability" comes down to your implementation (multi-regional, graceful fail-over, etc.). It is entirely possible that the statistical reliability of individual components of one of these clouds is worse, but they have engineered their architecture on top of it to handle these failures gracefully. It is very difficult to estimate the probability of an unforeseen multi-regional black swan event that would actually threaten these businesses.

It is also worth noting that these businesses have a choice between their original on-prem datacenter, which was likely built under a similar engineering and management culture (and therefore similar quality), and their cloud offering, which is probably equal or better than on-prem (generally public products will be more polished, better documentation, more consistent, etc.). Factor in the waste of maintaining extra data centers and capacity when their cloud is likely not 100% utilized by customers and the decision seems obvious. Cloud and on-prem could both be terrible and they would still likely migrate to their own cloud for sheer efficiency reasons. They really don't have a choice to use a different cloud.

The author also tries to reason about the complexity of the requests being handled by each business (e-commerce vs. email) and I think it is very difficult to do this meaningfully. All of these companies have to integrate systems, developed by separate teams, which contain lots of unique functionality and data. I'm not sure I agree that e-commerce is necessarily more complex than Google maps or docs. I would have speculated that Google's architecture is generally more complex than Amazon's based on my subjective interactions with the platforms and experience with the two engineering cultures. Even if there are large differences in complexity, this strikes me as using layer 7 information to reason about layer 1-4. As we move to serverless and managed offerings this might make more sense, but I think for most businesses, clouds are very much just compute, storage, and network right now, which doesn't have much bearing on how complex of an architecture you can engineer on top of it.

Another thing I don't see discussed a lot is documentation and stability of API. I have never had an unpredictable API response from AWS or GCP. Their docs may not be perfect but they are generally accurate. Alicloud definitely gets respect for handling Singles Day, however I have run into very unreliable documentation and API behavior. It is always a bit of an adventure re-applying a Terraform template that I made only a few months ago. Perhaps their internal team gets more heads up when things change, but I consider DevOps reliability to be a real issue there. I have also noticed occasional issues with single-zone network latency, managed Hadoop version compatibility, and other general fit-and-finish things.

In general I consider all of these clouds to be equivalent enough in reliability that it just isn't a deciding factor. The location of clients and APIs that I'm working with is a much bigger factor (if they are already in a cloud zone somewhere). I subjectively prefer AWS because I have worked with them longer and I am more familiar, but I have had projects where it made sense to use each of the 4 clouds and I think being able to accommodate those requirements with equal degrees of reliability is exactly what good engineering is about.