Performance on Osman Goni Nahidhttps://osmangoninahid.github.io/tags/performance/Recent content in Performance on Osman Goni NahidHugoen-usSat, 19 Sep 2026 00:00:00 +0400We were hashing every byte twice: making 500 GiB volume imports fasthttps://osmangoninahid.github.io/hashing-every-byte-twice/Sat, 19 Sep 2026 00:00:00 +0400https://osmangoninahid.github.io/hashing-every-byte-twice/<p>On the platform I work on, a <em>workspace volume</em> is a Kubernetes PVC that users fill from Hugging Face or an S3 bucket and then mount into training or inference workloads. The import runs as a one-off pod: download, then record metadata (a file tree with per-file hashes, and one digest for the whole volume) back to the control plane.</p> <p>Customers importing 500 GiB+ models started reporting that the download finished in a reasonable time and then the pod sat for <strong>hours</strong> in the metadata step. This is the story of why, and what fixed it.</p>