DEV Community: Vishal Chandak

Why AWS Is Tearing Down the "Tree"?

Vishal Chandak — Mon, 27 Jul 2026 18:40:24 +0000

Amazon CEO Andy Jassy has been on a highly publicized crusade lately to crush corporate bureaucracy. His goal? Flatten the org chart, remove the layers of middle management that slow things down, and let individual teams move faster.

As it turns out, AWS’s network engineers took his mandate incredibly literally—they just applied it to their data centers.

For thirty years, the networking industry has been stuck with a rigid hierarchy known as the "fat-tree." In this model, data center architecture looks exactly like an aging, bloated corporate org chart. To move a packet of data from one server rack to its neighbor, the data can’t just walk across the hallway. It has to climb the corporate ladder—from Top-of-Rack (ToR) switches, up to aggregation layers, and finally to the "executive" spine switches—before trickling back down to its destination.

Every layer in this network acts exactly like a middle manager: it adds latency, drains power, and serves as a potential bottleneck. While the fat-tree has been the workhorse of the cloud, AWS is officially tearing it down. They are replacing it with a radically flat alternative called the Resilient Network Graph (RNG). By taking quasi-random graph theory out of academia and into hyperscale production, AWS has figured out how to commoditize randomness.

Here is how AWS is slashing their hardware footprint while actually making the network faster.

Takeaway 1: Why "Random" Beats "Ordered"

In traditional networking, order equals efficiency. But RNG proves the exact opposite: randomly connecting routers is mathematically superior to stacking them in a neat hierarchy.

In a traditional fat-tree, traffic is limited to strict pathways. Think of it like forcing all highway traffic through a few toll booths; it creates "small cuts," or bottlenecks where upper-layer links are jammed while other routes sit completely idle. Because of this, a fat-tree can strand up to 60% of its capacity.

In an RNG fabric, those toll booths don't exist. Every node has high bandwidth to all other nodes. This is powered by the Expansion Property of random graphs, creating "capacity fungibility"—a fancy way of saying no bandwidth ever goes to waste.

So why didn't we do this decades ago? Because it looked terrible on paper. As Giacomo Bernardi, AWS Principal Applied Scientist, noted:

*“It was typical for academia. Everybody's excited, but then the real world hits.”*

The real world meant an unmanageable hairball of physical cables and routing memory requirements that no standard switch could handle. Until now.

Takeaway 2: The "Magic Number" 69%

By firing the "middle managers" (removing the aggregation and spine layers entirely), AWS is pushing more data using drastically less metal. This shift to RNG is one of the most massive infrastructure optimizations in cloud history.

The impact breaks down into four staggering metrics:

69% fewer networking devices (routers and switches).
Up to 33% higher throughput compared to traditional fat-trees.
40% less power consumption for network equipment.
9% to 45% in cost savings , depending on the setup.

Beyond the balance sheet, eliminating thousands of power-hungry routers directly shrinks the carbon footprint of AWS's global grids. After all, the greenest energy is the power you never have to use in the first place.

Takeaway 3: The ShuffleBox (Taming the Spaghetti Monster)

The biggest barrier to flat networks has always been the physical wiring. Connecting routers hundreds of meters apart in a truly random pattern usually results in a logistical nightmare affectionately known as the "spaghetti monster."

AWS solved this with a secret weapon called the ShuffleBox.

Think of the ShuffleBox as a master illusionist. It’s a passive optical device that internally scrambles fiber wiring into a deterministic, random pattern. To the network, the topology looks wonderfully chaotic and random; to the technician on the floor, the cabling looks perfectly structured, neat, and maintainable. Because it’s completely passive, it requires no power and has no active parts that can fail.

Crucially, the ShuffleBox fixes the expansion problem. In older random graphs, adding a new server rack meant unplugging live cables, risking a total outage. With ShuffleBoxes, technicians can plug in new racks just like they always have, growing the network without disrupting live traffic.

🎥 Watch: How AWS Tames the Data Center "Spaghetti Monster" 📦⚡

Takeaway 4: Spraypoint Routing (Taking the Side Streets)

To navigate a structure-less network, AWS had to throw out the traditional networking rulebook. Standard routing logic uses too much memory—scaling it up would require 20 to 80 times more memory than commodity switches actually have.

Instead, AWS built a lightweight protocol called Spraypoint. It works on a simple "Spray and Point" strategy:

Spray: The source router essentially blasts the data out to all its immediate neighbors.
Point: The packets bounce through "rings" of waypoints surrounding the destination, guiding the data in from every possible angle.

It sounds counterintuitive. Why take a longer, zig-zagging path? Because in a flat network, taking a dozen different side streets is actually faster than getting stuck in a traffic jam on the main corporate highway.

🎥 Watch: SprayPoint Explained

Takeaway 5: Surviving the Blast Radius

The best argument for RNG brings us right back to Andy Jassy’s war on corporate structure: removing single points of failure.

In a hierarchical fat-tree, if a critical "VP" spine router fails, it can wipe out half the capacity for massive regions of a data center. The blast radius is catastrophic.

In an RNG topology, because there is no "top" of the network, a failure is barely a blip. If 1% of the routers fail, you only lose 1% of your capacity.

The network degrades proportionally, making outages virtually invisible to the end user. As Matt Rehder, VP of Network Engineering, put it:

“You turn it on, it functions. It's not something we want our customers to think about at all.”

The Future of the Flat Fabric

As of April 2026, this isn't just an ambitious science experiment. RNG is the global default for all new general-compute AWS data center builds. The network "tree" has been officially chopped down for standard workloads.

The final boss is specialized AI training. Because massive AI clusters require highly synchronized, centralized traffic patterns, they still rely on AWS's structured UltraServer architecture. But as RNG research evolves to handle these coordinated workloads, even those AI islands might eventually get flattened.

AWS has proven that chaos—when properly managed—scales better than order. The question now is whether competitors will follow them into the mesh, or remain stubbornly stuck in the trees while AWS enjoys a 69% hardware advantage.

Serverless vs Managed Servers

Vishal Chandak — Mon, 14 Dec 2020 05:56:43 +0000

Serverless seems really nice to begin with but does it make your bills a lot more heavier as you grow?

Uing AWS Lambda when the traffic received is around 100 transactions per seconds works fine and costs nominal but if the Transactions per seconds were to grow ten fold (or more) is Lambda still a viable option?

Along with Lamba, the Cloud watch costs extra to house the logs this is not an issue with managed servers.

Demystifying Machine Learning

Vishal Chandak — Thu, 29 Oct 2020 07:26:51 +0000

Even though a quick google search can give you plenty of “answers” for this question, it appears that they have actually not been satisfactory enough. The fact that you are reading this post should be enough to back my claim but if you are still not convinced, take a look at the google trends report for the term “Machine Learning”.

So what is Machine Learning? Let’s find out!

Before you go any further, I should warn you that this post is intended for people with little or no idea about the topic and thus has a super-simplified view of what Machine learning is.

To put it in Layman’s terms— Machine Learning is the ability to let machine form an equation to correctly solve any problem with little or no human intervention.

With Machine Learning, instead of programming the machine to do something, we let the machine learn to find a solution to the problem by itself.

“It sounds cool, but how is that it even possible.” — Your Brain

Hold your horses, we will get to that question in a minute. Before that, let’s understand how is it different than the “usual” way of solving a problem.

To understand this analogy better, let’s take a simple real-world example of — “Classifying shapes”.

Problem:

Given an image of a particular shape, let’s find out if the image is for a Square or Triangle.

Non Machine Learning Approach

The pseudo-code (fancy term programmers use to define “steps”) for that would look like:

Take the input Image
Find the number of corners in the image. This operation would basically include: (Don’t worry if you don’t understand all the terms below, neither do I)
- Convert the image from RGB (i.e. Colored Image) to Grayscale (i.e. B/w)
- Smooth-en out the image
- Use thresholding to convert the image to Bi-Level Binary Image
- Use an edge-detector algorithm to get the list of edges using pixel gradients
- Use contours to get the list of corners in the image
Write an if-else block to return the shape of the object based on the number of corners returned

Here’s a small visualization to help you better visualize what the above steps might look like:

Huh! You must be thinking — That seems a lot of work for such a trivial task. You are not entirely wrong but this has been the general approach to solve such problems until a few years back. Even though this approach “worked”, it had a lot of issues. For starters, it only worked well with perfectly captured images i.e. pictures taken under proper light and clear background which means that it will have a high percentage of failure in real-world scenarios.

Machine Learning Approach to the rescue

Let’s see what the same would look like with Machine Learning:

Gather some images for different shapes — squares and triangles.
Feed it to a ML Algorithm (more technical term would be a Convolutional Neural Network)
Run the image to be tested on the above trained model
Sit back and relax!

Now you must be thinking “Ok. That seems to be awfully easy. But how does that even work?” Let me answer that question with a question of my own — How would you teach a child a difference between a square and a triangle?

You won’t start by telling him about the number of corners or that the angles between them should be 90; Rather you would show him some images and help him relate that one of them is called a square and other is called a triangle.

Exactly! This is the way Machine Learning works too — based on feedback from the label provided with the data.

The below image might help understand it better:

Let’s take off the training wheels and see how far our newly trained child can go:

“All that’s cool but I don’t need a machine to categorize shapes.” — Almost everyone

To give a little more perspective on its potential - Machine Learning is the magic behind how Google lets you find your favorite cat photos (Who doesn’t love cats!) on the internet. If you look closely, almost everything you interact with on a day-to-day basis has machine learning weaved into it — Amazon product recommendations, Netflix movie recommendations, predicting stock prices, Alexa Voice, Sales forecasting, Automatic subtitle generation, Language translation and a lot more. The details pertaining to those is outside the scope of this post.
The major advantages with the Machine Learning approach is that:

You don’t need any domain knowledge to get started
The same classifier which can recognize shapes can also recognize numerical digits, cats vs dogs etc. since the Machine trains itself based on the input data
Can be continuously improved based on its own performance

The few disadvantages with Machine Learning are:

Needs a lot of training data for better predictions
Needs a lot of processing power for training the model and deriving inferences

However, the advent of the information age, providing us with humongous amount of data, and the drastic reduction in CPU and GPU costs over the years have helped solve both the above problems. Thus, it makes sense now more than ever, to have a generalized solution for such cases rather than a problem-specific solution.

That’s where Machine Learning comes into play. Sounds really cool, doesn’t it? If it’s so fascinating, why isn’t everyone trying to get their hands dirty with Machine Learning?

I will be addressing this question in my next post. Stay Tuned!

Until then, That’s all folks!