Quick summary

  • Cloudflare is prototyping Cache Transcoding in Pingora, encoding eligible assets with Zstandard before storage and decoding them at delivery time.
  • The design shows how a small CPU tradeoff can produce large storage and inter-data-center bandwidth savings at CDN scale.
  • Canary the design on high-hit-rate assets and compare compression, CPU, latency, and bandwidth before widening eligibility.

What happened

Trading CPU for cache capacity

Cloudflare is prototyping an architecture called Cache Transcoding as memory and disk costs rise. When an eligible response enters cache, Pingora encodes it with Zstandard before writing it to disk. The compressed representation remains in place while the asset is cached and while Tiered Cache moves it between data centers. It is decoded only before the response is delivered to a client.

Initial tests reduced eligible assets to roughly one third of their original on-disk size on average. Encoding is paid once when the object enters cache, while storage and cross-region bandwidth savings recur each time the object is reused. The tradeoff is a modest increase in CPU at the origin-facing proxy, so the decision is an optimization of total infrastructure cost rather than a free compression switch.

What web infrastructure teams can learn

The internal storage representation can differ from the client representation without changing the reconstructed bytes. CDN teams should measure compression ratio, encode and decode CPU, tail latency, cache hit rate, and inter-data-center traffic for each content class. Objects with high reuse and long residence times are stronger candidates than already compressed or rarely requested files.

The experiment also illustrates a broader systems principle: optimize the complete data lifecycle. A bounded transformation at ingestion can create repeated savings across reads and replication. Production rollout still needs canaries, MIME allowlists, CPU budgets, observability, and an immediate rollback path so storage efficiency does not become compute saturation.

Source

How we could save petabytes of cache storage with Zstandard and Pingora

Why developers should care

The design shows how a small CPU tradeoff can produce large storage and inter-data-center bandwidth savings at CDN scale.

  1. 1Canary the design on high-hit-rate assets and compare compression, CPU, latency, and bandwidth before widening eligibility.