On Thu, Oct 19, 2017 at 11:04 PM, Jeff King <peff@xxxxxxxx> wrote: > On Thu, Oct 19, 2017 at 10:27:28PM -0700, Stefan Beller wrote: > >> > If my analysis above is correct, then it's already fixed. You just had >> > leftover corruption. >> >> Well fetching yesterday worked and the commit in question is from >> 8/23, the merge 8a044c7f1d56cef657be342e40de0795d688e882 >> occurred 9/18, so I suspect there is something else at play. >> (I do not remember having a gc between yesterday and today. >> Though maybe one in the background?) > > Even a gc between yesterday and today should have used the new code, > which would have been safe. So yeah, maybe it is something else > entirely. Oh, yeah. > >> I am curious how you can have a worktree owned by multiple >> repositories [1] (?). > > Sorry, I forgot my footnote. I saw this with my "ci" script: > > https://github.com/peff/git/blob/7905ff395adecdd2bb7ab045a24223dfb103e0e9/ci > > I check out the contents of my "meta" branch as "Meta", and it contains > that script. It basically just waits for ref updates, then walks over > all the commits and runs "make test" on them in the background (caching > the results, thanks to the git-test[1] script). So I kick off "Meta/ci" > in a terminal and forget about it, and magically it builds my commits in > the background as I work. > > It operates in a worktree inside the Meta directory (Meta/tmp-ci), so as > not to disturb what I'm doing. So far so good. > > But I actually have _two_ clones of Git on my system. One on which I do > most of my work, and then the other which has the fork we use in > production at GitHub. I symlink the Meta directory from the first into > the latter, which means they both see the same worktree directory. And > somehow running "Meta/ci" in the second corrupted things. > > I can get some funniness now, but I think it's mostly caused by the > script being confused about the worktree existing but not having access > to our branches. That's not a corruption, just a confusion. I _think_ I > had a bogus HEAD in the worktree at one point, but I may be > mis-remembering. I can't seem to trigger it now. Thanks for these insights. I played around with Meta a bit, but I did not feel it would enhance my workflow enough as I am not involved with any maintainance of git using git. The git-test from Michael sounds intriguing. Initially I put off using it as I had my main working dir (or rather test dir) on a spinning disk, still. Now I test in memory only, which is a lot faster, so I could try if git-test can keep up with my local commit pace. Thanks, Stefan
