# Mike Smith - Rethinking Netflix’s Edge Load Balancing (Highlights) ![rw-book-cover|256](https://miro.medium.com/v2/da:true/resize:fit:862/0*7V7WBr802_zlfJjV) ## Metadata **Review**:: [readwise.io](https://readwise.io/bookreview/61975054) **Source**:: #from/readwise #from/reader **Zettel**:: #zettel/fleeting **Status**:: #x **Authors**:: [[Mike Smith]] **Full Title**:: Rethinking Netflix’s Edge Load Balancing **Category**:: #articles #readwise/articles **Category Icon**:: 📰 **Document Tags**:: #work **URL**:: [netflixtechblog.com](https://netflixtechblog.com/netflix-edge-load-balancing-695308b5548c) **Host**:: [[netflixtechblog.com]] **Highlighted**:: [[2026-07-10]] **Created**:: [[2026-07-18]] ## Highlights - When possible, instead of configuring static thresholds, use adaptive mechanisms that change based on the current traffic, performance and environment. ([View Highlight](https://read.readwise.io/read/01kx70ssrvn27mf39a4khkezbb)) ^1033287442 - An overarching idea was that while the best source of data for a servers’ *latency* is the client-side view, the best source of data on a servers’ *utilization* is from the server itself. And that combining these 2 sources of data should give us the most effective load-balancing. ([View Highlight](https://read.readwise.io/read/01kx70tybte5e10x3mr9zh4pv7)) ^1033288109 - **Decay** of collected server stats to zero over time. ([View Highlight](https://read.readwise.io/read/01kx70whgvwcxr6wx8rat1vycp)) ^1033289057 - Join-the-shortest-queue works very well for a *single* load-balancer, but has a significant problem if used in isolation across a *cluster* of load balancers. The problem is that the load balancers will tend to *herd* and all choose the *same* low-utilized servers at the same time, thus overloading them, and then moving on to the next least-utilized and overloading that, and onward … ([View Highlight](https://read.readwise.io/read/01kx713jry3wz834ks5at5jhyx)) ^1033290568 #caveat Solution: use with choise-of-2 - Using each server’s viewpoint on their utilization has the advantage that it provides the aggregate of all load balancers that are using that server, and therefore avoids the JSQ problem of an incomplete picture. ([View Highlight](https://read.readwise.io/read/01kx717c3bxqx1bwwj06cd994d)) ^1033290796 - **Passively** track responses from the servers annotated with their current utilization data. ([View Highlight](https://read.readwise.io/read/01kx717vwn6jp1bg08m77apkrb)) ^1033290804