Patrick Steinhardt reported that git log --stat --follow -- include/linux/rsc_table.h in linux.git takes roughly 33 seconds with Git 2.55 but almost 8 minutes with Git 2.56. He bisected it to 304812ed33 (log: improve --follow following renames for non-linear history). The report came from a user on IRC.

Junio C Hamano replied to a suggestion from Justin Tobler. He called it sensible and far better than reverting and declaring that the original "checkbox hack" must stay forever because doing it properly would be too expensive. He said following a single file is an important operation in some workflows.

Steinhardt said he does not know how close the follow-up series is to being merged. He proposed first landing a single-patch series that adds a knob, independent of the rest. He gave two ways to opt in: extend --follow to take a value such as --follow=merges, with a matching value for log.follow, or add a separate setting. He said GitLab deploys from Git's master branch, so the cost is a real problem for them. He offered to help but did not want to leave the whole burden on the author.

Miklos Vajna said he has no problem making the behavior opt-in rather than default. He would prefer not to revert, since a lot of effort went into the change and some see the old behavior as a bug that hides real history. He said he is unhappy with the complexity of the last patch in the series. He asked whether a new config option similar to log.follow would meet Steinhardt's needs.

Steinhardt said he is leaning that way too. He has no strong view on a full revert and finds the patches hard to review, which he attributes to his own unfamiliarity with the code. If the change is kept, he wants a toggle to disable it as a first step.

What remains open is the spelling of the opt-in, whether a new --follow value or a separate config, and who writes the patch.