Amazon

Context

During the summer of my freshman year, I joined Amazon as a Software Engineering Intern on the Astrolabe team. Our team focused on improving traffic routing efficiency for data centers across the world.

Data centers are the physical facilities that host the servers powering AWS cloud services, which support products like Prime Video.

The Problem

Amazon's data centers are overcrowded and customers in rural regions experience higher latency due to inefficient rouitng.

As Amazon scaled, the default traffic routing logic struggled to consistently deliver low-letency connections in edge-case regions. Some regions–namely PDX, DUB, and ZAZ–host data centers that support a significant portion of AWS cloud services. These are known as mega-regions.

Mega-regions are gradually approaching full capacity, meaning there is limited physical hardware available to support the increasing number of requests and services. Furthermore, customers (e.g., Prime Video users) may experience higher latency when streaming, due to traffic being routed to distant, crowded data centers in mega-regions.

When a user in a rural area streams a video via Prime Video, their traffic might be routed to an overcrowded region like DUB–despite closer and less-crowded alternatives–leading to buffering or timeouts.

Without region-aware routing, customer requests are often directed based on static priorities rather than real-time availability or proximity. This results in degraded performance, especially during peak traffic hours, ultimately impacting reliability and user satisfaction across AWS-hosted services.

The Solution

To address this routing inefficiency, I built the GeoTokenHeaderExtractor (yeah, the name is super long), a plugin designed to take the first step towards region-aware routing for Prime Video. Prime Video embeds a GeoToken in the header of each HTTP request, which contains a two-character region code identifying the user's country of origin.

How did this plugin work? My plugin intercepted incoming HTTP requests, extracted and parsed the GeoToken, and rerouted traffic to a closer AWS center based on the regional information from the token. By using geographic context directly into the routing logic, the solution reduced the likelihood of sending traffic to distant, overcrowded mega-regions, ultimately improving streaming performance for end users in underrepresented areas.

Impact

Editing.

What I Learned

Editing.