Big Disk Slow Pipe

An Open Connect site holds two generations of appliance in one cluster:

Type Storage Throughput
Disk server about 200 TB about 40 Gbps
Flash server up to 18 TB up to 100 Gbps

Content is placed on a consistent hash ring, each server taking about 1000 positions, weighted by storage. Requests follow the content: a server is asked for exactly the titles it holds.

What breaks, and what does Netflix do about it?

A

Store the full catalog on both types and let the load balancer split requests in proportion to throughput.

B

Weighting by storage hands the flash box a sliver of the catalog and therefore a sliver of the requests, while it can serve more than twice the disk box's throughput. Netflix splits the catalog by popularity across two weighted rings.

C

Weight the single ring by throughput instead of storage. Give the flash server two and a half times the positions and traffic matches capacity.

D

Increase the number of hash positions per server so the distribution smooths out.