This changes the upload-pack response copy buffer size in both Git and Gitaly. This is an efficiency improvement for Git fetch traffic. It only makes a difference when the pack-objects cache is enabled; without that it is performance-neutral.
Also see gitlab-git!12 (closed).
I did an experiment where I cloned gitlab-org/gitlab from a local VM, while running perf record at 99Hz. This is with the pack-objects cache enabled, and it's a cache hit. What is nice that we not only save CPU cycles in git and gitaly (the binaries touched by this MR) but also in praefect and gitlab-workhorse.
| Process | CPU samples before | CPU samples after |
|---|---|---|
git |
326 | 179 |
gitaly |
444 | 347 |
praefect |
480 | 333 |
gitlab-workhorse |
442 | 367 |