How can I "split the branch further" in mercurial? - mercurial

Admittedly a misleading title, but I didn't know how to put it better.
So I made the mistake of pushing all my project specific changes BEFORE I opened a branch. That way, there are some changeset sitting in the default branch that does not belong there. They should be in my newly opened branch.
Do I have to backout all the changesets first and then push them again to the correct branch? Does that even work?
Thank you for your help.

Without resorting to EditingHistory (which is generally advised agains, and is certainly in opposition to the mercurial "immutable history" design goals) those changes will be in your history forever. You can, however, undo them in your 'tip' revision and then trickle them back in multiple subsequent changesets spread over multiple branches.
Let's imagine this is your current history:
[R0] -> [R1] -> [R2] -> [R3]
Where revisions R2 and R3 have some stuff you wanted in the default branch and some stuff you wish was in a different branch. You can create a new revision, R4, that undoes R2 and R3. That can be done using backout as you suggest, or since you're doing multiple changesets, perhaps more easily using 'revert'.
hg update tip ; hg revert --all -r R2 ; hg commit -m 'undid changes R2 and R3'
then your history will look like this:
[R0] -> [R1] -> [R2] -> [R3] -> [R4]
then you can split your work into the changesets you wish you'd initially done, yielding a history like this:
[R0] -> [R1] -> [R2] -> [R3] -> [R4] -> [R5] -> [R7]
\
--> [R6] -> [R8]
R6 and R8 can be on the branch named 'default' or on a different named branch. Either way, you can push just R0 through R5 and R7 by doing:
hg push -r R7
That works because a push pushes a revision and all of its ancestors -- but R6 and R8 aren't ancestors of R7, so they won't get pushed no matter what branch they're on.

Related

How do I copy commits from one branch to another?

I have a branch with a few revisions in it. I want to try making some code changes that require reordering and patching those commits with histedit, but I want to keep the original branch around in case it doesn't go well. How can I do that?
Example
Before:
master -> change 1 -> change 2 (branch A)
After:
master -> change 1 -> change 2 (branch A)
-> change 1 -> change 2 (branch B)
The integrated and recommended way to copy or cherry-pick commits from one branch to another is using
hg graft -r XX YY ZZ.
where XX YY ZZ etc. are the revisions to copy to your currently checked-out branch. By default they are committed, but you can also use the --no-commit flag so that you can edit changes. Or do it one-by-one and make additions using hg commit --amend.
Compared to exporting and importing it has the added benefit of using the configured merge programme, should a merge be required - but it will NOT be a merge, so no line of ancestors is established from the current commit to the one you copy from.

Finding Merges in the Wrong Direction

We have three major releases a year, with co-responding branches: e.g. 2013A, 2013B, and 2013C. When we create each branch, it is started from default. Changes in each branch should only be merged forward, e.g. 2013A -> 2013B -> 2013C -> default. We have a push hook on the server that checks if a pushed merge is in the wrong direction, i.e. default -> 2013C, 2013C -> 2013B, etc.
We also have team-specific branches, some of which are working on features for a release, and others which are working on the next release, e.g. default. While a team is working on a release, they merge to/from the release branch. When the team is ready to work on the next release, they start to merge to/from the default branch.
The other day we had a situation where a new developer merged default into his team branch before the team was ready to move on to the next release, then merged the team branch into a previous release, i.e. default -> TeamBranch -> 2013B. Unfortunately, our hook didn't take this situation into account.
Essentially, this is what happened:
2013B A---o---o---o---o---B---o
/ \ / \
Team / o---o---o---C---o---o
/ / \
Default D---o---o---o---o---o---o---o---o
A = Creation of 2013B branch
B = Merge into release branch
C = The bad merge. We want to detect and prevent these whenever we merge into a release branch.
D = The first common ancestor of the release branch and default.
So, I've re-written our hook to check that when a change is merged into a release branch, it doesn't merge backward. For each merge into a release branch, I check that there aren't any ancestor merges from a forward branch. Here's the revset query I'm using:
> hg log -r "limit(descendants(p1(first(branch('2013B')))) and reverse(p2(ancestors(branch('2013B'))) and branch('default')),1)"
This works. However, we have a large repository (111,000+ changesets,) and the check takes 30 to 40 seconds. I wanted to know if there is a quicker/faster/more efficient way of writing my revset query, or another approach I'm not seeing.
I sent this same question to the Mercurial mailing list and received an answer. The branch() query is the performance bottleneck. It causes Mercurial to unroll all the changesets on a branch. Mercurial doesn't cache the results of this, so each call will unroll changesets.
Instead of using branch() I switched to using descendants() and ancestors():
limit(children(p2(2013BBaseline:: and ::2013B and merge()) and branch(default)) and reverse(::2013B))
p2(2013BBaseline:: and ::2013B and merge()) and branch(default) grabs the second parent (the incoming branch) for all merges between the start of the 2013B branch and its head, and returns just those on the default branch. [1]
The clause above is then wrapped with children() to go back down to the children of that parent.
and reverse(::2013B) then gets the children that are ancestors of the 2013B branch, i.e. the bad merge(s).
limit() then returns just the first of those bad merges.
The query above takes about 1.5 seconds.
Thanks to Matt Mackall for suggesting the solution.
2013BBaseline is a tag which identifies the changeset in the default branch from which 2013B branch was created, otherwise I would have had to replace 2013BBaseline:: with:
p1(first(branch(2013B)))::
to discover the release branch's baseline, and that is not very performant.
To detect the merge C, you should be able to use
$ hg log -r "parents(branch(Team) and merge()) and branch(default)"
This gives you changesets on default that were merged into Team. If there are any, then someone merged into their team branch by mistake. I think you need to attack this from the point of view of the team branch: you cannot forbid merges to/from the release branch since some of them will be legitimate.
If your team branches follow a consistent naming scheme (they should), then you can use a regular expression with the branch() predicate to select them. Something like
$ hg log -r "branch('re:team-.*')"
would match branches starting with team-.

'Forgetting' a dead-end branch

I've got a mercurial repository. It was on rev A. I made some changes, committed (to rev B), and pushed. However, later, I realised I didn't want to make those changes. I updated back to rev A, and made some alternative changes, to rev C.
C
| -
| B
|/
A
However, now I can't push rev C, because it complains that it would create a new remote head (which it would). How do I make the remote mercurial simply forget about rev B and all the changes therein, so I can push rev C and carry on from there?
Editing History is hard. Once you push a changeset to a public repository, it can no longer be easily pruned from the history.
The most direct solution to your problem is:
hg update <tip of branch you want to forget>
hg commit --close-branch -m "close unwanted branch"
hg update <tip of branch you want to keep>
Push all your changes. As you noted, you will need to use --force since multiple heads on the branch now exist.
If you really need to prune the branch, then read EditingHistory again. If it still seems feasible, you can use one of the methods described in PruningDeadBranches.
Personally I'd close the branch and force the push (as Tim Henigan describes), as it leaves the DAG in a state which is truthful. There is another option though. Doing a dummy merge. This is a merge, but one where you ignore the incoming changes.
hg update C
hg -y merge --tool=internal:fail B
hg revert --all --rev .
hg resolve -a -m
hg ci
The end result is
M
|\
C |
| |
| B
|/
A
... but M doesn't contain any of B's changes.
Use the backout command to reverse the B change.
hg up B
hg backout B
Now, you're working directory is in the same state it was before B, which I'll call A'.
After you backout, make your real change and commit. This is what your history will look like:
C
|
A'
|
B
|
A
Mercurial branches are permanent and registered in commit objects themselves. There are some (not too easy) methods for closing and/or removing branches, mainly listed here. I have even already used some of them before. Note that those are somewhat desperate solutions for people who did use branches thinking they can be temporary.
However, if you really want temporary branches, use bookmarks.
Other people have some good answers for getting rid of 'B', but just to put it out there you can always do:
hg push --rev C
which will push C but not B and not complain about new heads.

What's the best way to back out multiple changesets in mercurial?

Is the most reliable method to go one-by-one, using the backout command for each of many changesets, or is there a way to create one big reversal changeset to cover a whole bunch of [edit: non-contiguous] changesets.
If one-by-one, does order matter? (Should one go last-to-first?)
Does the best method differ if there are merges among different sub-projects along the way?
Does this tend to go smoothly in your experience? :-)
If you have no merges along the way, you can either back out every individual change (in reverse order), or, if there are many of them, do it with one big inverse patch.
If you have good changesets atop the ones you need to back out, better commit the inverse patch on top of the most recent bad changeset, then rebasing them onto the tip of the branch.
1 -- 2 -- A -- B -- C -- 3 -- 4
\
C'B'A'
$ hg up C
$ hg diff -r C:2 > backout.diff
$ hg import --no-commit backout.diff
$ hg ci -m "Backout A, B, C"
$ hg up 4
$ hg rebase -s C'B'A -d .
There will be problems if you want to back out merge changesets, see this wiki page for more information.
In such a case, if possible, consider re-doing the branch and stripping the old lineage. Otherwise, you might have to abandon the branch altogether, salvaging the good changesets via graft or transplant.
There is --collapse option for rebase.
Helgi's answer can be upgraded into:
1 -- A -- 2 -- B -- 3 -- C -- 4 -- 5
\
C' -- B' -- A'
$ hg update --clean C
$ hg backout --rev C --message "Backed out changeset: C"
$ hg backout --rev B
$ hg commit --message "Backed out changeset: B"
$ hg backout --rev A
$ hg commit --message "Backed out changeset: A"
$ hg rebase --collapse --source C' --dest 5
$ hg commit --message "Backed out C, B, A"
which will result in the following
1 -- A -- 2 -- B -- 3 -- C -- 4 -- 5 -- C'B'A'
However, backing out in separate branch may result in [logical] conflict in the subsequent merge.
1 -- A -- 2 -- B -- 3 -- X -- 4
\ \
B' -- A' -- M
if X depends on A or B, then M will have conflict (at least logical conflict).
What I came up with is inelegant, but got the job done, despite that the changes I needed to back out were interspersed with other work and had some internal branching. Here's what I did. (Comments and improvements are welcome.)
Got a list of all of the changesets (which I then used to generate the commands below):
hg log -r 'keyword(xyz)' --template '{rev}\n'
Generated a patch for each changeset:
hg diff -p -U 8 --reverse -c 15094 > 15094.rev.patch
hg diff -p -U 8 --reverse -c 15095 > 15095.rev.patch
...
Then, applied each reverse patch. Here the order matters, last-to-first:
hg import -m "reversing changeset 15302" 15302.rev.patch
hg import -m "reversing changeset 15292" 15292.rev.patch
...
This process was interrupted several times for merges that didn't go through automatically, where I had to manually apply changes to a file from its .rej file and then manually commit, before picking up the imports where it had left off.
Finally (in another clone... did I mention I did this all in a clone?) I compressed the whole set of reverse changesets into one changeset using hg histedit -o and its fold command.
Now I've got a single changeset that I should be able to reverse and apply if I decide to put the work back in at a later date (Although if I cross that bridge, I might apply the "forward" patches piecemeal again in order to get better blame/annotate information)
This is how you can do it with TortoiseHg.
Of course you can do the same with the command line.
Given this history, where you wan't to get rid of changeset A, B and C:
1 -- 2 -- A -- B -- C -- 3 -- 4
First update to revision 2.
Then rebase the first of any later revisions you wan't to keep - in this case revision 3.
Your history now looks like this:
1 -- 2 -- A -- B -- C
\
3 -- 4
Now update to revison 4.
And finally use "Merge with local" to merge revision C onto revision 4.
At this point it is crucial that you select the option "Discard all changes from merge target (other) revision".
The description may not be the most logical, but it means that you merge the old tip C back to the default branch - but without the changesets A, B and C.
The result is:
1 -- 2 -- A -- B -- C --
\ /
3 -- 4
Commit and you're done.
If you don't want the "backout" changesets in your history, you could also do something else:
Make a clone of your repository, but only up to the last changeset that you don't want to get rid of.
See Mercurial: Fix a borked history for an example how to do this.
If your repository was a local one, that's all you have to do.
But if the bad changesets were already pushed to a central repository, you'd need server access to delete the repository there and replace it by your clone.
Plus, if someone else already pulled from the repo with the bad changesets, they need to delete and re-clone (otherwise the bad changesets are in the central repo again as soon as one of the other people pushes again).
So it depends on the circumstances whether this solution is a good one for you...

Mercurial: Picking commits to be included in a release

I have a local mercurial repository (for now) within which I have already made several commits, each commit is a self contained bug fix. Is it possible to pick which of the bug fixes (commits) I want to be included when it is time to build a release version of my application.
To elaborate, assuming A, B, C, D, and E are commits I have already done to my repository and each of them relates to a bug fix like so:
A <- B <- C <- D <- E <- working dir
I need to be able to for example pick which of the bug fixes will go into the release version (this depends on the time allocated for deployment as well as testing outcomes). So for example I might get a report saying the release should only contain bug fixes A, C and D.
Is it possible to construct a release version containing only the A, C and D commits (Keeping in mind that each commit is self contained and does not depend on the other commits to actually be there)?
Probably having a branch for each bug fix and then merging into a release branch is the easiest way to accomplish this (or is it not?), but the current situation at hand is as described above with no branches.
This isn't the normal work mode of Mercurial (or git). A repository can only contain a changeset if it also contains all of that changeset's ancestors. So you can't get D into a repo without also having A, B, and C in there.
So here's:
What you Should have Done
Control the parentage of your changesets. Don't make C the parent of D just because you happen to have fixed D after C. Before you fix a bug hg update to the previous release.
Imagine A was a release and B, C, and D, were all bug fixes. If you do a loop like this:
foreach bug you have:
hg update A
... fix bug ...
hg commit
hg merge # merges with the "other" head
then you'll end up with a graph like this:
---[A]----[B2]--[C2]--[D2]----
| / / /
+-[B] / /
| / /
+-----[C] /
| /
+---------[D]
and now if you want to create a release with only, say, B and D in it you can do:
hg update B
hg merge D
and that creates a new head that has A + B + D but no C.
Tl;Dr: make a change's parent be as early in history as you can, not whatever happens to be tip at the time.
What you can do Now
That's the ideal, but fortunately it's no big thing. You can never bring exactly D across without bringing C (because C's hash is part of the calculation of D's hash), but you can bring the work that's in D into a new head easily enough. Here are some ways, any of which will work:
hg export / hg import
hg transplant
hg graft (new in 2.0)
hg rebase (only possible if you haven't yet pushed)
Any of those will let you bring that patch/delta that's in D over -- it will have a different hash ID and when some day you merge D in for real (using merge) you'll have duplicate work in two different changesets, but merge will figure it all out.
If this was my tree and it hasn't been pushed anywhere, I'd (assuming an empty patch queue and MQ enabled):
hg qimport -g -r B: # import revisions B and later into mq as "git" style patches
hg qpop -a # unapply them all
hg qpush --move C # Apply changes in C (--move rearranges the order)
hg qpush --move D # Apply changes in D
hg qfin -a # Convert C & D back to changesets
hg push <release server> # Push them out to the release branch
Then you can hg qpush -a; hg qfin -a to get B & E back into changesets.
Final Result:
---A---C---D---B---E
Advantages:
Nobody needs know you didn't do things in this order to start with (evil grin)
You could modify any of the change-sets whilst doing this
Alternatively, with graft in 2.0:
hg update -r A # Goto rev A (no need to do anything special for A)
hg graft C # Graft C on to a new anonymous branch
hg graft D # Graft D
This will give you
---A---B---C---D---E
\
--C'--D' <-You are here
An hg push -r D' should just push the new, cherry-picked, head.
You can then hg merge to get one head again with B and E included.
Advantages:
Non destructive, so true history is kept, and no chance of loss if you muck up
hg tags the new changesets with the hash of the original version, so totally trackable
Probably a little simpler.
While it's somehow strange way and release-policy, you can do it in different form. You have to manipulate with two main objects: changesets and branches
Version 1
You use two branches (default + f.e "release 1.0"). Default branch is mainline of your work - all changesets commited to this branch. At release-time, you branch first needed-for-release changeset into (new) branch, transplant or graft rest of needed in release changesets from default to this branch, head of release 1.0 will be prepared for release this way.
Next release will differ only in new branch name
Version 2
One branch used, MQ extension added. Variations:
all changesets are MQ-pathes and only needed for release are applied to repo
changesets are changesets, only unwanted for release converted to mq-pathes, later qfinish'ed and returned to repo