Re: [RFH] git cherry-pick takes forever
Hello,
Junio C Hamano wrote:
Show 10 quoted lines
>Michal Vitecek <fuf@mageo.cz> writes:
>> I have two git repositories: one is the origin of the other. However no
>> merging is being done as the projects in the repositories quite differ
>> but still use the same core. So to propagate changes I cherry-pick
>> those which are useful from one repository to another.
>>
>> however 'git cherry-pick' has lately started to last almost forever:
>
>Can you define "lately"? Is it a function of your git version, or is it a
>function of the age of your repositories?
It's a function a cherry-picking some commits - afterwards
cherry-picking crawls. One of the commits removes a number of files
(614) and also renames some (303).
Show 13 quoted lines
>> $ time git cherry-pick b42b77e66a83f1298d9900a9bb1078b9b42e8618
>> Finished one cherry-pick.
>> Created commit 7caef83: - removed some superfluous newlines
>> 2 files changed, 0 insertions(+), 2 deletions(-)
>> git cherry-pick b42b77e66a83f1298d9900a9bb1078b9b42e8618 282.97s user 34.69s system 100% cpu 5:17.63 total
>>
>> Both repositories have approximately 16k commits and their forking
>> point (merge base) is 250 to 490 commits far away.
>
>When talking about cherry-pick, the size of the history (unless the
>repository has too many objects and badly packed) does not matter; the
>operation is purely about your current state, the cherry-picked commit
>itself, and the parent commit of the cherry-picked one.
I too thought so but after cherry-picking starting taking so long I
began to doubt my thoughts :)
Show 14 quoted lines
>Taking 5 minutes to cherry-pick a change to only two paths, one line
>deletion each, is plain ridiculous, but if the tree state of cherry-picked
>commit and the tree state of the target is vastly different (e.g. almost
>no common pathnames), the behaviour is certainly understandable. Ancient
>git used straight three-way merge for cherry-pick, but recent ones use
>more expensive "recursive-merge", which tries to detect renames. If the
>states of trees are very dissimilar, you can end up wasting a lot of time.
>
> $ H=$(git rev-parse 7caef83^) ;# the commit before cherry-pick
> $ C=b42b77e6 ;# the cherry-picked one
>
>cherry-pick operation roughly runs these two diffs:
>
> $ time git diff --shortstat -M $H $C
$ time git diff --shortstat -M $H $C
2 files changed, 0 insertions(+), 2 deletions(-)
git diff --shortstat -M $H $C 0.00s user 0.00s system 72% cpu 0.006 total
> $ time git diff --shortstat -M $H $C^1
$ time git diff --shortstat -M $H $C\^1
git diff --shortstat -M $H $C\^1 0.00s user 0.00s system 0% cpu 0.003 total
>and uses the result to perform its work. Can you clock these?
>
>If you rarely have renames, it may be much more efficient to run "git
>format-patch -1 --stdout $C | git am -3" instead of cherry-pick.
Turning off renames detection in diff (via 'git config --add diff.renames
false') helped and 'git cherry-pick' is instant now.
Maybe my repositories are "strange" in some way. I would be more than
happy to provide more information if needed.
Thank you,
--
Michal Vitecek (fuf@mageo.cz)