Editing History

How to modify repository history.

1. Alternatives to editing history

If you would like to undo a changeset, but don't mind having it preserved in history, you can use the backout command to reverse it. This is generally preferred to changing history as it's non-destructive and lets future developers use that history to avoid making the same mistakes.

2. Why changing history is hard

First, consider changesets. Each changeset's id is a cryptographically strong hash of the changeset data, which recursively includes all of the changeset content (data and metadata) as well as the ids of its parents. Change any bit in a changeset itself or the history it's based on and you will change its id. This makes Mercurial changesets tamperproof: it is computationally infeasible to make a tampered changeset that has the same changeset id as a given changeset.

Second, Mercurial's network protocol assumes history is append-only. Pushing and pulling only ever add history to repositories, never remove it. If the history you want to modify has already been published to public repositories, there is no way to recall it except with the cooperation of everyone who has pulled a copy, which is generally not practical.

3. Consequences of editing history

If you edit your repository history, the changeset IDs (i.e., the identity of the changesets) will be changed from the point of the edit forward. Suppose for example that the bad revision is number 3. Then, before your change, the repository will look like this:

After your edit, changes BAD, R4 and R5 will have new changeset IDs:

As long as nobody else has seen the repository before the change, this is okay. If people have already pulled from your repository, then things become more complex. Suppose someone cloned your repository before you edited BAD, and then pulled afterwards. They would see completeley new changesets, BADa, R4a and R5a, and their tree of changes would look like this:

This is exactly what one would expect - Mercurial always works this way when you pull in changes from others: It takes two directed acyclic graphs (one representing your repository, the other representing the repository you pull from) and merges common nodes in the graphs to produce a new acyclic graph.

Notice that the changesets R4 and R4a and R5 and R5a are identical changes (neither of you edited them), but they have different hash values since they have different histories (because of BADa).

So you can simply ask your friend to strip BAD, which will strip R4 and R5 as well - their changes are preserved in R4a and R5a. This makes your friend's repository identical to yours:

On the other hand, if your friend had already committed new work on top of R5, then his repository looks like this before the pull:

After the pull he gets

Now it is not simply enough to strip BAD since that destroys the work in R6 and R7. But you can ask him to import R6 and R7 into MQ, pop the queue, strip BAD and apply the queue again (this time to R5a). This is rebasing his changes on R5':

 hg qimport -r R6:R7
 hg qpop -a
 hg strip BAD
 hg update -C R5a # might not be necessary
 hg qpush -a
 hg qdelete -r qbase:qtip

4. Motivation

Having said all of this, there are good reasons why people might want to change history. Here are some examples:

As long as you understand the implications, it is possible to do this.

The key implication is that the changeset IDs will change from the point at which the revision occurs. This means that developers with clones will need to rebase their changes, and care must be taken to manage the effects of the revision. This document does not attempt to cover this process. It assumes that if you need to do this, you will ensure that you know what to do, and you make sure it happens.

5. Mercurial safety features

Since version 2.2, Mercurial includes a feature called phases that will prevent you from editing history that has been pushed or pulled to another repository. If the above motivations apply, you can still forcibly change the phase of published changesets. It's also possible to set up a repository as "non-publishing", such that when you push changes to it, they do not get marked as public. See phases for more information.

6. Scenarios

6.1. Basic changeset removal with clone

One of the simplest tasks is removing the most recent commits in a repository. This can be done non-destructively with clone:

hg clone -r LASTGOODREVISION oldrepo newrepo

- and then perhaps move oldrepo away as a backup and rename newrepo to take its place. If the old repository have multiple heads you might want to pull them too.

6.2. Amending the latest changeset with commit --amend

Another common history-editing task is changing something in the most recent commit, either to modify the code or to edit the commit message. Since version 2.2, the commit command has a --amend option that will fold any changes into your working directory into the latest commit, and allow you to edit the commit message.

hg commit --amend can in fact be used on any changeset that is a (topological) branch head, that is, one that has no child changesets. It need not be the tip revision.

6.3. Editing recent history with MQ

Recent history can be modified fairly easily with the MQ extension:

Some caveats exist. First, MQ can't operate on merge changesets. Second, by default MQ works with textual changes. If the history you edit contains binary files, permission changes or other non-textual changes, enable extended diffs for your repo. Add the following section to your hgrc:

git = True

(alternatively remember to add --git to every qimport and qrefresh invocation below).

Let's pretend you comitted the file which should not be commited in revision BAD. Then doing

hg qimport -r BAD:tip

will import the changesets into MQ. You can find newly created patches in .hg/patches.

Those patches are nevertheless still applied, to strip them from the history, you need the qpop command. Issue

hg qpop -a

Now all changes since revision BAD are no longer available in your repository history. They are saved as patches in .hg/patches - and only there.

If you want to undo the entire changeset BAD (obliterate it!), then do this:

 hg qdelete BAD.diff

If you only want to edit the changeset (remove some edits, avoid commiting one file while leaving the remaining changes etc), then do this:

 hg qpush BAD.diff
 # edit files, remove passwords, revert newly added files etc.
 hg qrefresh

Note that qdelete removes whole patch, while qrefresh modifies it.

Now if you want to combine the next two patches (say GOOD1.diff and GOOD2.diff) into one patch, do this:

 hg qpush GOOD1.diff
 hg qfold GOOD2.diff

The changes from GOOD2.diff have been integrated ("folded") into GOOD1.diff and the GOOD2.diff patch itself has been deleted from the queue (it can be kept there but unmanaged, by using qfold -k).

To go back to standard Mercurial changesets you do

 hg qpush -a
 hg qfinish -a

The qfinish command turns an applied patch into a real Mercurial changeset. Here we use it to turn all applied patches into normal changesets.

One thing this process over-simplifies is that the hg qpush -a step may fail, if later changes depend on the obliterated data. In that case, you have to fix the problem manually - there's no easy answer here, after you edit history, you need to manage the consequences, in your own repository as well as elsewhere. So you need to push conflicting patches one by one, edit them as appropriate, and qrefresh the changes. Read chapter about Mercurial Queues from Mercurial Book for details.

Here's a real example (captured on Windows). I add a file I shouldn't in revision 1, edit it in revision 2, then try to obliterate revision 1. After doing so, I need to resolve the issue with revision 2 (editing what is now a nonexistent file). In this case it's easy, we just drop revision 2 as well. In other cases, this could be much harder to deal with, possibly even so much harder that you decide it's not worth it. It depends on your data (and possibly your lawyers!).

>hg init
>echo Line 1 >a
>hg commit --addremove -m "Added a"
adding a

>rem Add the launch codes for the nuclear arsenal here...
>echo Super secret >b
>hg commit --addremove -m "Added b"
adding b

>hg log
changeset:   1:65bcb0d3f953
tag:         tip
user:        "Paul Moore <user@example.com>"
date:        Sat Mar 22 16:43:00 2008 +0000
summary:     Added b

changeset:   0:5dd6949828e1
user:        "Paul Moore <user@example.com>"
date:        Sat Mar 22 16:42:40 2008 +0000
summary:     Added a

>rem Here we compound the error...
>echo More secret stuff >>b
>hg commit -m "Edited b"

>echo More safe stuff >>a
>hg commit -m "Edited a"

>hg log
changeset:   3:ea4f8ad48048
tag:         tip
user:        "Paul Moore <user@example.com>"
date:        Sat Mar 22 16:43:46 2008 +0000
summary:     Edited a

changeset:   2:6bb0d654a0a6
user:        "Paul Moore <user@example.com>"
date:        Sat Mar 22 16:43:32 2008 +0000
summary:     Edited b

changeset:   1:65bcb0d3f953
user:        "Paul Moore <user@example.com>"
date:        Sat Mar 22 16:43:00 2008 +0000
summary:     Added b

changeset:   0:5dd6949828e1
user:        "Paul Moore <user@example.com>"
date:        Sat Mar 22 16:42:40 2008 +0000
summary:     Added a

>rem We realise our mistake. We need to get rid of changeset 1,
>rem so that file b is no longer in out repository!
>hg qinit
>hg qimport -r 1:tip
>hg qpop -a
Patch queue now empty

>rem Delete changeset 1
>hg qdelete 1.diff

>rem Now start to put everything back
>hg qpush -a
applying 2.diff
unable to find 'b' for patching
1 out of 1 hunk FAILED -- saving rejects to file b.rej
patch failed, unable to continue (try -v)
b: No such file or directory
b not tracked!
patch failed, rejects left in working dir
Errors during apply, please fix and refresh 2.diff

>rem Hmm, change 2 depends on file b. Fix things up. Luckily, this is easy, just delete change 2 as well.
>rem In reality, change 2 may contain other edits, and we'd need to do some further fixing.
>hg qdelete 2.diff
abort: cannot delete applied patch 2.diff

>rem Even this isn't as simple as all that. Back out change 2 so we can delete it.
>hg qpop -a
Patch queue now empty

>hg qdelete 2.diff

>rem And now we're good to go.
>hg qpush -a
applying 3.diff
Now at: 3.diff

>hg qdelete -r qbase:qtip

>rem No sign of file b, and the world is safe again.
>rem Except, of course, that evil Doctor Death pulled from us 5 minutes ago.
>rem But at least as we all get blown up, we can be glad that it's not a technical problem :-)
>hg log
changeset:   1:a50e33884959
tag:         tip
user:        "Paul Moore <user@example.com>"
date:        Sat Mar 22 16:43:46 2008 +0000
summary:     Edited a

changeset:   0:5dd6949828e1
user:        "Paul Moore <user@example.com>"
date:        Sat Mar 22 16:42:40 2008 +0000
summary:     Added a

The process of going up through the patch stack, tidying up the debris (as in our example, where change 2 wouldn't apply as it depended on the obliterated file "b"), is what is generally referred to as "rebasing" the changes. It can be simple, in the case of a localised change, but it can be arbitrarily complex. Before you start editing history, you need to be sure that you know what to do to rebase.

7. Other options

There are other options that may be more appropriate in particular circumstances.

8. See also


EditingHistory (last edited 2013-08-29 16:11:40 by EgorKuropatkin)