Cloudflare Saves Petabytes of Cache Storage with Zstandard and Pingora

As memory and hard disk costs surge, Cloudflare developed a 'Cache Transcoding' prototype leveraging Zstandard (zstd) compression within its Pingora architecture. This approach reduces the disk size of assets stored in the cache to an average of 1/3, securing petabytes of effective cache capacity and inter-data center bandwidth while accepting a slight increase in CPU usage.

Efficiency of Zstandard Compression

Zstandard is a lossless compression algorithm developed by Facebook, offering balanced performance that is 42% faster than Brotli while generating files 11.3% smaller than gzip. To prevent speed degradation, Cloudflare applied zstd level 3, which provides an optimal balance between compression ratio and processing speed.

Scope and Limitations

Not all assets are candidates for compression. Already compressed media such as images, videos, and fonts are excluded to avoid wasting CPU resources. Instead, text-based assets like HTML, JSON, CSS, and JavaScript are the primary targets; test results showed these assets were compressed by approximately 2.8x. Specifically, for compressible text larger than 4 KiB, nearly all potential storage savings across total traffic are achieved while operating within the CPU budget.

How It Works and Benefits

Upon a cache miss, the Pingora proxy encodes the response body with zstd before storing it on disk, then decodes it before delivering it to the client. In this process, the compression cost is incurred only once when the asset enters the cache, but the benefits of increased storage density and reduced data transfer during Tiered Cache movements persist every time the asset is reused.