Notes
Write-ups
Irregular. Mostly measurements and decisions we want to be able to point at later.
-
Deduplication numbers from real pipelines
What 64 KiB chunking actually saves on four different build trees, and why the frontend bundle behaves nothing like the Go binary.
-
Why BLAKE3 and not SHA-256
Hashing was the bottleneck at 1.2 GB/s. Switching the digest moved the ceiling to the disk instead, which is where it belongs.
-
Making output byte-identical
Directory iteration order, sparse files, and the three timestamp fields that had to be thrown away to get reproducible archives.
-
Format version 2
Moving the manifest to the front of the file so archives could be streamed, and what that broke.