How we could save petabytes of cache storage with Zstandard and Pingora
What Happened
Cloudflare prototyped Cache Transcoding, encoding eligible cache assets with Zstandard (zstd) inside Pingora to reduce on‑disk size to about one‑third on average. The encoding is applied once when an asset enters the cache, then the compressed form is stored and transferred between data centers, and decoded before serving to clients. The extra CPU cost in the origin‑facing proxy is small, while the savings accrue in storage and cross‑data‑center bandwidth each time the asset is reused.
Why It Matters
Enterprise architects can view this as a way to extend cache capacity without adding physical memory, lowering storage costs and reducing inter‑data‑center traffic. The modest CPU overhead may be acceptable if the organization prioritizes cost savings and bandwidth efficiency, but it requires monitoring CPU utilization and ensuring that compression/decompression does not impact latency or throughput.
The Limitation
The compression benefits depend on asset eligibility and may not apply to all content types; also, the initial CPU cost could become significant under high request volumes or with slower CPUs.
What You Can Do
Implement a pilot of zstd‑based cache compression on a subset of non‑critical assets to measure storage savings and CPU impact before rolling out broadly.