Cloudflare tests zstd compression to expand cache capacity

Cloudflare is testing a cache architecture that could substantially increase effective storage capacity by compressing eligible content with Zstandard (zstd) before it reaches disk. The prototype, called Cache Transcoding, is built into Cloudflare’s Pingora proxy and is aimed at reducing both cache footprint and traffic between data centers.

The approach targets text such as HTML, JSON, CSS, and JavaScript—not images, video, or fonts, which are generally compressed already. In Cloudflare’s traffic sample, text accounted for 67.3% of requests and 22.3% of bytes; about 71% of that text arrived without content encoding. Eligible assets at least 4 KiB in size are compressed at zstd level 3 when they enter the cache, kept compressed across cache tiers, and decoded only on the client-facing path.

In a controlled corpus, the assets became roughly 2.8 times smaller, with a measured compression ratio of 2.834x. Encoding cost 4.31 nanoseconds per byte—about 232 MB/s—and decoding cost 1.56 nanoseconds per byte, or roughly 641 MB/s. Encoding happens once per cache fill, while decoding occurs on each serve, making the CPU tradeoff central to the design.

For teams building AI products, the experiment highlights a practical infrastructure lever: compressing reusable payloads can delay hardware expansion, reduce eviction, and lower internal bandwidth costs without changing the bytes delivered to users. Cloudflare estimates the strategy could provide petabytes of effective cache capacity, though the article describes it as a prototype rather than a general availability feature.

Source: CloudFlare