Re: Unreliable 'git rebase --onto'

["SZEDER Gábor" <szeder.dev@gmail.com>]

On Wed, Jan 08, 2020 at 04:55:46PM -0800, Elijah Newren wrote:
> > Alas, there is unexpected bad news: with that commit the runtime of
> > your 'git rebase --onto' command goes from <1sec to over 50secs.
> > Cc-ing Elijah, author of that patch...
> 
> I see slowdown, but not nearly as big as you report:

The linux repo is big, my notebook is small, the poor thing :)

> $ time git rebase -m --onto v4.18 463fa44eec2fef50~ 463fa44eec2fef50
> warning: inexact rename detection was skipped due to too many files.
> warning: you may want to set your merge.renamelimit variable to at
> least 7216 and retry the command.
> Successfully rebased and updated detached HEAD.
> 
> real 0m13.305s
> user 0m9.644s
> sys 0m3.620s

> Interestingly, turning off rename detection only speeds it up a little bit:
> $ time git rebase -m -Xno-renames --onto v4.18 463fa44eec2fef50~
> 463fa44eec2fef50
> Successfully rebased and updated detached HEAD.
> 
> real 0m11.955s
> user 0m8.732s
> sys 0m3.424s
> 
> 
> This is an interesting testcase; I'm going to try to find some time to
> dig in further.

The culprits are two seemingly unnecessary back-and-forth checkouts.

I didn't realize I could use 'git rebase -m', so ran some tests with
it, and turns out that the slowdown started with 68aa495b59 (rebase:
implement --merge via the interactive machinery, 2018-12-11), where
the runtime suddenly went from <1.5s to 45+s.

Running 'git rebase -i --onto <those-same-commits>' is just as slow,
and it appears that it has always been (the oldest I tried was
v1.8.0), and it spends a long time both before and after popping up
the editor for the rebase instructions.  That's highly suspicious, so:

  $ git log --oneline -1
  94710cac0ef4 (HEAD, tag: v4.18) Linux 4.18
  $ git rebase -i --onto v4.18 463fa44eec2fef50~ 463fa44eec2fef50
  hint: Waiting for your editor to close the file... 
  # Hit ctrl-z in the editor
  $ git log --oneline -1
  463fa44eec2f (HEAD) Input: atmel_mxt_ts - disable IRQ across suspend

Oh.

So 'git rebase -i' apparently checks out the tip commit of the
to-be-rebased revision range before invoking the editor for the rebase
instructions, only to check out the --onto commit (i.e. the commit
we've started from!) to apply the selected commit on top.

And indeed those two checkouts account for all the wasted runtime:

  $ time { git checkout 463fa44eec2fef50 && git checkout v4.18 ; }
  Updating files: 100% (49483/49483), done.
  Previous HEAD position was 94710cac0ef4 Linux 4.18
  HEAD is now at 463fa44eec2f Input: atmel_mxt_ts - disable IRQ across suspend
  Updating files: 100% (49483/49483), done.
  Previous HEAD position was 463fa44eec2f Input: atmel_mxt_ts - disable IRQ across suspend
  HEAD is now at 94710cac0ef4 Linux 4.18
  
  real    0m48.801s
  user    0m13.963s
  sys     0m5.114s