squash commits deep down in history

Hi,

i need to squash several commits into a single one in a automated way. I know that there is interactive rebase, which can also be automated using GIT_SEQUENCE_EDITOR. Unfortunately my history is very large and i need to squash deep down in the history several times. So using interactive rebase seems not to be the right tool.

I wonder if i can solve this task using filter-branch? I have a file that list the SHA1s for each squash operation per line. All SHA1s of a line are in chronological order (youngest to oldest), or in other words: the first SHA1 is the child of the second, and so on.

| ab4423e 3432343 3234252
| 2324342 5232343
| ...

Lets say there are N lines in that file. Each line means: squash all SHA1s of this line into the first (or last) SHA1 of this line. Performing this task with rebase would require N rewritings of the history. So e.g. HEAD (but many others too) would be rewritten N times even if it is not directly part of a line. My thinking is, that a filter-branch can do this in a single rewrite and therefore would be much more performant.

How can i solve this? Any ideas?

Best regards Henning

Am 23.10.2014 um 14:34 schrieb Henning Moll:

Use git-replace do construct a squashed commit that replaces the last child in each run such that its parent points to the parent of the first in the run. Then use a git-filterbranch without filters to burn the parenthood into the history.

-- Hannes

Am 23.10.2014 um 22:08 schrieb Johannes Sixt:

Wow, cool. Thanks a lot for the hint.

I managed to do the following: Suppose i have the following history

A - B - C - D   <-master

and i want to squash B and C together. First i create a temporary branch on A and switch to it. Then i cherry-pick B and C without commit. So all the changes are merged into the index. Then i commit a single commit "M". The history now looks like

   M             <-temp
  /
A - B - C - D   <-master

Using

$ git replace C M

leads to

    temp
         |
         v

A -     M - D   <-master

As far as i understand, such a repalce is something like a "symbolic link". Is that right?

This is interesting: Before filter-branch, the deletion of the replace would lead to the initial situation (ABCD). After filter-branch, the history is (AM'D'). The replace still exists but is now something like a "dead link". I may come to life again after "git resetting" to somewhere before filter-branch. Even "before" creating the replace (the replace ref itself is not part of the history). Even if you gor before creation of M: The link still works as long as M is not garbage collected. I was very irritated after resetting to where i commited D, and still seeing A-M-D.

Somehow logical, somehow crazy! I am still a git learner ;-)

Two questions remain for me:

1. Is there a more elegant way to build M?
2. Having a replace "in place" before filter-branch: Is there a way to 
find out what the replace is hiding? A simple "git log C" produces 
"M-A". How to get "C-B-A"?

Best regards Henning

W dniu 2014-10-23 22:08, Johannes Sixt pisze:

Using grafts (which are only about parentage, and are purely local) with filter-less git-filter-branch might be easier than git-replace.

If you have the following history:

     ... <--- O <--- A <--- B <--- C <--- ...

and you want to squash A, B, C into

     ... <--- O <----------------- C <--- ...

simply add the following to .git/info/grafts

    <sha-1 of C> <sha-1 of O>

for example with

    echo "$(git rev-parse C) $(git rev-parse C^^^)" >>.git/info/grafts

Check if the history is correct e.g. with "git log --graph --oneline", then use filter-branch.

BTW. do you know if it is possible to run filter-branch without replacing replaces?

Just for completeness, yet another way to do scripted history rewriting is to use git-fast-export + some editing script (e.g. reposurgeon[1]) + git-fast-import.

[1]: http://www.catb.org/esr/reposurgeon/

On Thu, Oct 23, 2014 at 8:34 AM, Henning Moll <newsScott@gmx.de> wrote:

I've often felt there should be some simple commands to do these kinds of edits. For example, it is easy to amend HEAD, but an order-of-magnitude more difficult to amend HEAD^. I imagine commands like this:

   git rebase --reword HEAD^
   git rebase --edit some-old-commit
   git rebase --squash ab4423e 3432343 3234252

But each time I think of implementing one of these I am daunted by the many exceptional cases.

I know you solved this already with filter-branch, but that seems like a complicated solution to me. I think the --interactive rebase would have been useful for you. You should keep in mind that you do not need to repeat the command multiple times for your multiple squashes. For example, if your to-do list looks like this simple example:

pick 0000000 pick ab4423e pick 3432343 pick 3234252 pick 0000001 pick 0000002 pick 2324342 pick 5232343 pick 0000003

I think you could get the desired effect by changing it to this:

pick 0000000 pick ab4423e squash 3432343 squash 3234252 pick 0000001 pick 0000002 pick 2324342 squash 5232343 pick 0000003

And then running that all in one interactive rebase.  Does that make sense?

Phil