• Sources: primary, discussion
  • Summary: Cloudflare published a post describing an intern prototype that transcodes cached objects to Zstandard inside its Pingora proxy, and it publishes per-byte encode and decode costs alongside the compression ratio. The post reports 2.834x on a deliberately compressible two-asset corpus and states that a broader corpus is needed before treating that figure as fleet-wide. It also reports that restricting compression to hot assets performed worse than compressing all eligible text, and nothing described in the post is deployed.
  • Why it matters: The post publishes the measured trade rather than a product claim, including per-byte encode and decode costs and the finding that restricting compression to hot assets performed worse than compressing all eligible text.
  • Follow-up: Track whether Cloudflare measures the prototype against a representative corpus or moves any part of it into production.

send feedback on this story