Cloudflare Tests Cache Transcoding to Reduce Storage Requirements

Cloudflare is experimenting with a new system called Cache Transcoding that aims to squeeze more storage space out of its caches. The technology compresses specific types of data, such as HTML, JSON, CSS, and JavaScript, using a compression algorithm called Zstandard. By doing so, Cloudflare estimates that it could unlock petabytes of extra cache capacity.

The compression process occurs once when the content is stored in the cache and then decompressed when the data is served to users. This process requires minimal additional CPU power, allowing Cloudflare to store more data on its existing servers. Furthermore, the compression reduces the amount of data that needs to be transferred between data centers, leading to cost savings.

Cloudflare's Cache Transcoding only targets content that is uncompressed, successfully returned, and at least 4 KiB in size. It excludes certain types of content, such as images, video, and fonts, as well as responses with unknown sizes. This targeted approach ensures that the compression is applied efficiently without wasting resources on already compressed content.

The compression algorithm used in Cache Transcoding is Zstandard, a lossless compression tool developed by Facebook for real-time applications. The encoding cost is incurred only the first time the content enters the cache, and the storage and bandwidth savings persist with every subsequent reuse of that content. Cloudflare has found that compressible text accounts for a significant portion of its traffic, representing 67.3% of requests and 22.3% of bytes.

While the compression process may introduce a small increase in CPU usage, the benefits of reduced storage and bandwidth requirements far outweigh this minor trade-off. Cloudflare is still refining the Cache Transcoding system and plans to conduct further testing to evaluate its performance across different compression levels, content types, object sizes, and cache scenarios.