AI Signal 307 2 feeds carried it
How we could save petabytes of cache storage with Zstandard and Pingora
Cloudflare prototyped a Cache Transcoding system that compresses eligible cache content with Zstandard inside Pingora, shrinking eligible assets to roughly one-third their original size at a minor CPU cost.
For operators of large-scale caching infrastructure, this demonstrates a concrete trade of a few percent more CPU for petabytes of effective storage and reduced inter-datacenter bandwidth. The approach is selective, only compressible text assets above 4 KiB are encoded, because re-compressing already-compressed media wastes CPU with no storage benefit.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
Eligible compressible text assets shrank to roughly one-third their original on-disk size using zstd level 3 compression.
Encoding cost is paid once per cache fill at 4.31 ns/byte, while decoding at 1.56 ns/byte is paid on every serve.
The prototype transcodes all eligible compressible text at or above 4 KiB, as limiting to only popular content reduced storage savings without proportionally cutting CPU.
THE CLUSTER
↗