Menu

How Uber saved millions and cut P99 latency by deleting an entire fleet of edge proxies

Reading 0:00
15s threshold

When designing global systems, the conventional wisdom is simple: put your front doors (proxies/edge servers) as close to the end user as possible. More regions = lower latency.

Uber previously ran a setup following this exact blueprint:

1. Traffic entered Google’s Global Load Balancer (with Cloud Armor and Cloud CDN).

2. Traffic was handed off to a fleet of Envoy proxies running on VMs across 16 Google Cloud regions.

3. Traffic crossed Cloud Interconnect into Uber’s on-prem data centers.

On paper, having Envoy proxies in 16 regions meant bringing services closer to users.…

Continue reading — create a free account

Join HashtagPLUS to read full articles, follow hashtags, vote, and join the conversation.