ChatGPT解决这个技术问题 Extra ChatGPT

When do you use Git rebase instead of Git merge?

When is it recommended to use Git rebase vs. Git merge?

Do I still need to merge after a successful rebase?

One problem with people who like to use rebase is that it deters them from pushing their code regularly. So wanting clean history prevents them from sharing their code, which I think is more important.
@static_rtti: That's just not true. You're using a rebase-based flow wrong if it prevents you from pushing your changes regularly.
My heuristic: "Try rebase, if it dissolves into conflict resolution hell give up and merge master into your branch and move on." - more detail in my post timwise.co.uk/2019/10/14/merge-vs-rebase (in the context of commercial development)

C
Community

Short Version

Merge takes all the changes in one branch and merges them into another branch in one commit.

Rebase says I want the point at which I branched to move to a new starting point

So when do you use either one?

Merge

Let's say you have created a branch for the purpose of developing a single feature. When you want to bring those changes back to master, you probably want merge (you don't care about maintaining all of the interim commits).

Rebase

A second scenario would be if you started doing some development and then another developer made an unrelated change. You probably want to pull and then rebase to base your changes from the current version from the repository.


@Rob mentioned maintaining interim commits when merging. I believe by default merging branch B (a feature branch you've been working on) into branch M (the master branch) will create one commit in M for each commit that was made in B since the two diverged. But if you merge using the --squash option, all of the commits made on branch B will be "lumped together" and merged as a single commit on branch M, keeping the log on your master branch nice and clean. Squashing is probably what you want if you have numerous developers working independently and merging back into master.
I believe @spaaarky21's assumption about merging is not correct. If you merge a branch B into the master M, there will be only a single commit on M (even if B has multiple commits), regardless of whether you use a plain or --squash merge. What --squash will do is eliminate the reference to B as a parent. A good visualization is here: syntevo.com/smartgithg/howtos.html?page=workflows.merge
@jpeskin That's not what I'm seeing. I just did a quick test to verify. Create a directory with a text file, init a new repo, add the file and commit. Checkout a new feature branch (checkout -b feature.) Change the text file, commit and repeat so that there are two new commits on the feature branch. Then checkout master and merge feature. In log, I see my initial commit on master, followed by the two that were merged from feature. If you merge --squash feature, feature is merged into master but not committed, so the only new commit on master will be the one you make yourself.
@spaaarky21 It looks like we're both half right. When a fast-forward merge is possible (as in your example), git will default to including all commits in the feature branch B (or as you suggest, you can use --squash to combine into a single commit). But in the case where there are two divergent branches M and B that you are merging, git will not include all of the individual commits from branch B if merged into M (whether or not you use --squash).
Why is that the "(you don't care about maintaining all of the interim commits)" aside still in this answer? It made no sense in '09 and it makes no sense now. Also, surely you would only want to rebase if another developer made related changes that you needed - if they made unrelated changes, your feature branch should merge easily without conflicts anyway, and your history would be maintened.
A
Anton Menshov

It's simple. With rebase you say to use another branch as the new base for your work.

If you have, for example, a branch master, you create a branch to implement a new feature, and say you name it cool-feature, of course, the master branch is the base for your new feature.

Now, at a certain point, you want to add the new feature you implemented in the master branch. You could just switch to master and merge the cool-feature branch:

$ git checkout master
$ git merge cool-feature

But this way a new dummy commit is added. If you want to avoid spaghetti-history you can rebase:

$ git checkout cool-feature
$ git rebase master

And then merge it in master:

$ git checkout master
$ git merge cool-feature

This time, since the topic branch has the same commits of master plus the commits with the new feature, the merge will be just a fast-forward.


but this way a new dummy commit is added, if you want to avoid spaghetti-history - how is it bad?
Also, the --no-ff flag of merge is very very useful.
@アレックス as user Sean Schofield puts it in a comment: "Rebase is also nice because once u do eventually merge ur stuff back into master (which is trivial as already described) you have it sitting at the "top" of ur commit history. On bigger projects where features may be written but merged several weeks later, you don't want to just merge them into the master because they get "stuffed" into the master way back in the history. Personally I like being able to do git log and see that recent feature right at the "top." Note the commit dates are preserved - rebase doesn't change that information. "
I think it bears repeating here -- remember that all these terms (merge, rebase, fast-forward, etc.) are referring to specific manipulations of a directed acyclic graph. They become easier to reason about with that mental model in mind.
@Aldo There's nothing "clean" or "tidy" about a rebased history. It's generally filthy and IMHO awful because you have no idea what really went on. The "cleanest" Git history is the one that actually occurred. :)
P
Peter Mortensen

TL;DR

If you have any doubt, use merge.

Short Answer

The only differences between a rebase and a merge are:

The resulting tree structure of the history (generally only noticeable when looking at a commit graph) is different (one will have branches, the other won't).

Merge will generally create an extra commit (e.g. node in the tree).

Merge and rebase will handle conflicts differently. Rebase will present conflicts one commit at a time where merge will present them all at once.

So the short answer is to pick rebase or merge based on what you want your history to look like.

Long Answer

There are a few factors you should consider when choosing which operation to use.

Is the branch you are getting changes from shared with other developers outside your team (e.g. open source, public)?

If so, don't rebase. Rebase destroys the branch and those developers will have broken/inconsistent repositories unless they use git pull --rebase. This is a good way to upset other developers quickly.

How skilled is your development team?

Rebase is a destructive operation. That means, if you do not apply it correctly, you could lose committed work and/or break the consistency of other developer's repositories.

I've worked on teams where the developers all came from a time when companies could afford dedicated staff to deal with branching and merging. Those developers don't know much about Git and don't want to know much. In these teams I wouldn't risk recommending rebasing for any reason.

Does the branch itself represent useful information

Some teams use the branch-per-feature model where each branch represents a feature (or bugfix, or sub-feature, etc.) In this model the branch helps identify sets of related commits. For example, one can quickly revert a feature by reverting the merge of that branch (to be fair, this is a rare operation). Or diff a feature by comparing two branches (more common). Rebase would destroy the branch and this would not be straightforward.

I've also worked on teams that used the branch-per-developer model (we've all been there). In this case the branch itself doesn't convey any additional information (the commit already has the author). There would be no harm in rebasing.

Might you want to revert the merge for any reason?

Reverting (as in undoing) a rebase is considerably difficult and/or impossible (if the rebase had conflicts) compared to reverting a merge. If you think there is a chance you will want to revert then use merge.

Do you work on a team? If so, are you willing to take an all or nothing approach on this branch?

Rebase operations need to be pulled with a corresponding git pull --rebase. If you are working by yourself you may be able to remember which you should use at the appropriate time. If you are working on a team this will be very difficult to coordinate. This is why most rebase workflows recommend using rebase for all merges (and git pull --rebase for all pulls).

Common Myths

Merge destroys history (squashes commits)

Assuming you have the following merge:

    B -- C
   /      \
  A--------D

Some people will state that the merge "destroys" the commit history because if you were to look at the log of only the master branch (A -- D) you would miss the important commit messages contained in B and C.

If this were true we wouldn't have questions like this. Basically, you will see B and C unless you explicitly ask not to see them (using --first-parent). This is very easy to try for yourself.

Rebase allows for safer/simpler merges

The two approaches merge differently, but it is not clear that one is always better than the other and it may depend on the developer workflow. For example, if a developer tends to commit regularly (e.g. maybe they commit twice a day as they transition from work to home) then there could be a lot of commits for a given branch. Many of those commits might not look anything like the final product (I tend to refactor my approach once or twice per feature). If someone else was working on a related area of code and they tried to rebase my changes it could be a fairly tedious operation.

Rebase is cooler / sexier / more professional

If you like to alias rm to rm -rf to "save time" then maybe rebase is for you.

My Two Cents

I always think that someday I will come across a scenario where Git rebase is the awesome tool that solves the problem. Much like I think I will come across a scenario where Git reflog is an awesome tool that solves my problem. I have worked with Git for over five years now. It hasn't happened.

Messy histories have never really been a problem for me. I don't ever just read my commit history like an exciting novel. A majority of the time I need a history I am going to use Git blame or Git bisect anyway. In that case, having the merge commit is actually useful to me, because if the merge introduced the issue, that is meaningful information to me.

Update (4/2017)

I feel obligated to mention that I have personally softened on using rebase although my general advice still stands. I have recently been interacting a lot with the Angular 2 Material project. They have used rebase to keep a very clean commit history. This has allowed me to very easily see what commit fixed a given defect and whether or not that commit was included in a release. It serves as a great example of using rebase correctly.


I mostly love this answer. But: Rebase doesn't make a "clean" history. It makes a more linear history, but that's not the same thing at all, since who knows now much "dirt" each commit is hiding? The cleanest, clearest Git history is the one that keeps branch and commit integrity.
It's worth mentioning that git has recently changed it's git pull behavior to include the --rebase flag by default. That means that doing rebases on branches used by multiple developers is a little less dangerous. A person pulling your changes might be surprised that there are some conflicts to be resolved during such an operation, but there would be no disaster.
Another con for rebase, imo it's harder in high-velocity git repos especially where the build time is > average time between merges. If branches are constantly being merged in, you need to keep rebasing until it's your turn which can be tricky if you're also coordinating with a build to pass. e.g. monorepos rebase merges to a shared branch might be difficult
P
Peter Mortensen

To complement my own answer mentioned by TSamper,

a rebase is quite often a good idea to do before a merge, because the idea is that you integrate in your branch Y the work of the branch B upon which you will merge. But again, before merging, you resolve any conflict in your branch (i.e.: "rebase", as in "replay my work in my branch starting from a recent point from the branch B). If done correctly, the subsequent merge from your branch to branch B can be fast-forward.

a merge directly impacts the destination branch B, which means the merges better be trivial, otherwise that branch B can be long to get back to a stable state (time for you solve all the conflicts)

the point of merging after a rebase?

In the case that I describe, I rebase B onto my branch, just to have the opportunity to replay my work from a more recent point from B, but while staying into my branch.
In this case, a merge is still needed to bring my "replayed" work onto B.

The other scenario (described in Git Ready for instance), is to bring your work directly in B through a rebase (which does conserve all your nice commits, or even give you the opportunity to re-order them through an interactive rebase).
In that case (where you rebase while being in the B branch), you are right: no further merge is needed:

A Git tree at default when we have not merged nor rebased

https://i.stack.imgur.com/7vsGZ.png

we get by rebasing:

https://i.stack.imgur.com/yCxOO.png

That second scenario is all about: how do I get new-feature back into master.

My point, by describing the first rebase scenario, is to remind everyone that a rebase can also be used as a preliminary step to that (that being "get new-feature back into master").
You can use rebase to first bring master "in" the new-feature branch: the rebase will replay new-feature commits from the HEAD master, but still in the new-feature branch, effectively moving your branch starting point from an old master commit to HEAD-master.
That allows you to resolve any conflicts in your branch (meaning, in isolation, while allowing master to continue to evolve in parallel if your conflict resolution stage takes too long).
Then you can switch to master and merge new-feature (or rebase new-feature onto master if you want to preserve commits done in your new-feature branch).

So:

"rebase vs. merge" can be viewed as two ways to import a work on, say, master.

But "rebase then merge" can be a valid workflow to first resolve conflict in isolation, then bring back your work.


merge after rebase is a trivial fast forward without having to resolve conflicts.
@obelcap: Indeed, this is kind of the idea: you take all the problem-conflict in your environment (rebase master within your new-feature branch), and then co master, merge new-feature: 1 pico-second (fast-forward) if master had no evolutions
Rebase is also nice because once you do eventually merge your stuff back into master (which is trivial as already described) you have it sitting at the "top" of your commit history. On bigger projects where features may be written but merged several weeks later, you don't want to just merge them into the master because they get "stuffed" into the master way back in the history. Personally I like being able to do git log and see that recent feature right at the "top." Note the commit dates are preserved - rebase doesn't change that information.
@Joe: mentally, you are saying "replay any of my changes (done in isolation in my private branch) on top of that other branch, but leave me in my private branch once the rebase is done". That is a good opportunity to clean-up the local history, avoiding "checkpoint commits", broken bisect and incorrect blame results. See "Git workflow": sandofsky.com/blog/git-workflow.html
@scoarescoare the key is to see how you local changes are compatible on top of the latest upstream branch. If one of your commit introduces a conflict, you will see it right away. A merge introduce only one (merged) commit, which might trigger many conflict without an easy way to see which one, amongst your own local commits, did add said conflict. So in addition to a cleaner history, you get a more precise view of the changes you introduce, commit by commit (replayed by the rebase), as opposed to all the changes introduced by the upstream branch (dumped into one single merge).
A
Andrew Arnott

A lot of answers here say that merging turns all your commits into one, and therefore suggest to use rebase to preserve your commits. This is incorrect. And a bad idea if you have pushed your commits already.

Merge does not obliterate your commits. Merge preserves history! (just look at gitk) Rebase rewrites history, which is a Bad Thing after you've pushed it.

Use merge -- not rebase whenever you've already pushed.

Here is Linus' (author of Git) take on it (now hosted on my own blog, as recovered by the Wayback Machine). It's a really good read.

Or you can read my own version of the same idea below.

Rebasing a branch on master:

provides an incorrect idea of how commits were created

pollutes master with a bunch of intermediate commits that may not have been well tested

could actually introduce build breaks on these intermediate commits because of changes that were made to master between when the original topic branch was created and when it was rebased.

makes finding good places in master to checkout difficult.

Causes the timestamps on commits to not align with their chronological order in the tree. So you would see that commit A precedes commit B in master, but commit B was authored first. (What?!)

Produces more conflicts, because individual commits in the topic branch can each involve merge conflicts which must be individually resolved (further lying in history about what happened in each commit).

is a rewrite of history. If the branch being rebased has been pushed anywhere (shared with anyone other than yourself) then you've screwed up everyone else who has that branch since you've rewritten history.

In contrast, merging a topic branch into master:

preserves history of where topic branches were created, including any merges from master to the topic branch to help keep it current. You really get an accurate idea of what code the developer was working with when they were building.

master is a branch made up mostly of merges, and each of those merge commits are typically 'good points' in history that are safe to check out, because that's where the topic branch was ready to be integrated.

all the individual commits of the topic branch are preserved, including the fact that they were in a topic branch, so isolating those changes is natural and you can drill in where required.

merge conflicts only have to be resolved once (at the point of the merge), so intermediate commit changes made in the topic branch don't have to be resolved independently.

can be done multiple times smoothly. If you integrate your topic branch to master periodically, folks can keep building on the topic branch, and it can keep being merged independently.


Also, git merge has the "--no-ff" (no fast-forward) option that allows you to revert all the changes introduced by a certain merge really easily.
Just ti make it more clear: You refer to the situation 'whenever you've already pushed' -- this should be bold. The Link to Linus post is great, btw., clarifies it.
but isn't it best practice to "update" from master into your topic branch, before you merge the topic branch into master via PR (to resolve conflicts in your branch, not the master)? We're doing it like that so most topic branches have as a last commit "merge branch master into topic-..." but here this is listed as a "feature" of rebasing and nobody mentions it for merging... ?
@AndrewArnott "Most topic branches should be able to merge without conflicts into their target branches" How should that be possible when 20 devs are working on 30 branches? There will be merges while you're working on yours - so of course you have to update you topic branch from target before creating a PR... no?
Not usually, @Sumit. Git can merge either direction just fine even though changes have been made to either or both branches. Only when the same lines of code (or very close) are modified across two branches will you get conflicts. If that happens frequently on any team, the team should rethink how they distribute work since resolving conflicts is a tax and slows them down.
J
Jeehut

I just created an FAQ for my team in my own words which answers this question. Let me share:

What is a merge?

A commit, that combines all changes of a different branch into the current.

What is a rebase?

Re-comitting all commits of the current branch onto a different base commit.

What are the main differences between merge and rebase?

merge executes only one new commit. rebase typically executes multiple (number of commits in current branch). merge produces a new generated commit (the so called merge-commit). rebase only moves existing commits.

In which situations should we use a merge?

Use merge whenever you want to add changes of a branched out branch back into the base branch.

Typically, you do this by clicking the "Merge" button on Pull/Merge Requests, e.g. on GitHub.

In which situations should we use a rebase?

Use rebase whenever you want to add changes of a base branch back to a branched out branch.

Typically, you do this in feature branches whenever there's a change in the main branch.

Why not use merge to merge changes from the base branch into a feature branch?

The git history will include many unnecessary merge commits. If multiple merges were needed in a feature branch, then the feature branch might even hold more merge commits than actual commits! This creates a loop which destroys the mental model that Git was designed by which causes troubles in any visualization of the Git history. Imagine there's a river (e.g. the "Nile"). Water is flowing in one direction (direction of time in Git history). Now and then, imagine there's a branch to that river and suppose most of those branches merge back into the river. That's what the flow of a river might look like naturally. It makes sense. But then imagine there's a small branch of that river. Then, for some reason, the river merges into the branch and the branch continues from there. The river has now technically disappeared, it's now in the branch. But then, somehow magically, that branch is merged back into the river. Which river you ask? I don't know. The river should actually be in the branch now, but somehow it still continues to exist and I can merge the branch back into the river. So, the river is in the river. Kind of doesn't make sense. This is exactly what happens when you merge the base branch into a feature branch and then when the feature branch is done, you merge that back into the base branch again. The mental model is broken. And because of that, you end up with a branch visualization that's not very helpful.

Example Git History when using merge:

https://i.stack.imgur.com/Oqqmm.png

Note the many commits starting with Merge branch 'main' into .... They don't even exist if you rebase (there, you will only have pull request merge commits). Also many visual branch merge loops (main into feature into main).

Example Git History when using rebase:

https://i.stack.imgur.com/0ZVla.png

Much cleaner Git history with much less merge commits and no cluttered visual branch merge loops whatsoever.

Are there any downsides / pitfalls with rebase?

Yes:

Because a rebase moves commits (technically re-executes them), the commit date of all moved commits will be the time of the rebase and the git history loses the initial commit time. So, if the exact date of a commit is needed for some reason, then merge is the better option. But typically, a clean git history is much more useful than exact commit dates. If the rebased branch has multiple commits that change the same line and that line was also changed in the base branch, you might need to solve merge conflicts for that same line multiple times, which you never need to do when merging. So, on average, there's more merge conflicts to solve.

Tips to reduce merge conflicts when using rebase:

Rebase often. I typically recommend doing it at least once a day. Try to squash changes on the same line into one commit as much as possible.


I would remove the downside (2) completely from your list because, as you said, squashing is a perfect solution to (2) and it always works
Thanks, very useful. This highlights an important point of there being multiple merge commits that can be avoided using rebase.
I'm commenting here because I have no idea where to complain about this. If I rebase main -> feature, GitHub will show the changes up to the common ancestor, which is abominable because reviewing that PR implies reviewing all of the work that has already been merged to main. It's disgusting, and that's why I use merge instead.
@tfrascaroli I don't get your point, if you rebase your feature branch onto the latest commit of your main branch, then a PR targeting the main branch should only show the changes in the feature branch AFTER the common ancestor. What you describe is only possible if your PR is targeting yet another branch, not main. Or am I misunderstanding something?
Well, this just happened to me yesterday. Branch feature from main. Merge other features to main with Squash and Merge. Rebase main -> feature. PR feature -> main = Impossible review because the changes detected go back to the branching from main. If I instead do Merge main -> feature and then PR feature -> main that doesn't happen.
C
Carl

TLDR: It depends on what is most important - a tidy history or a true representation of the sequence of development

If a tidy history is the most important, then you would rebase first and then merge your changes, so it is clear exactly what the new code is. If you have already pushed your branch, don't rebase unless you can deal with the consequences.

If true representation of sequence is the most important, you would merge without rebasing.

Merge means: Create a single new commit that merges my changes into the destination. Note: This new commit will have two parents - the latest commit from your string of commits and the latest commit of the other branch you're merging.

Rebase means: Create a whole new series of commits, using my current set of commits as hints. In other words, calculate what my changes would have looked like if I had started making them from the point I'm rebasing on to. After the rebase, therefore, you might need to re-test your changes and during the rebase, you would possibly have a few conflicts.

Given this, why would you rebase? Just to keep the development history clear. Let's say you're working on feature X and when you're done, you merge your changes in. The destination will now have a single commit that would say something along the lines of "Added feature X". Now, instead of merging, if you rebased and then merged, the destination development history would contain all the individual commits in a single logical progression. This makes reviewing changes later on much easier. Imagine how hard you'd find it to review the development history if 50 developers were merging various features all the time.

That said, if you have already pushed the branch you're working on upstream, you should not rebase, but merge instead. For branches that have not been pushed upstream, rebase, test and merge.

Another time you might want to rebase is when you want to get rid of commits from your branch before pushing upstream. For example: Commits that introduce some debugging code early on and other commits further on that clean that code up. The only way to do this is by performing an interactive rebase: git rebase -i <branch/commit/tag>

UPDATE: You also want to use rebase when you're using Git to interface to a version control system that doesn't support non-linear history (Subversion for example). When using the git-svn bridge, it is very important that the changes you merge back into Subversion are a sequential list of changes on top of the most recent changes in trunk. There are only two ways to do that: (1) Manually re-create the changes and (2) Using the rebase command, which is a lot faster.

UPDATE 2: One additional way to think of a rebase is that it enables a sort of mapping from your development style to the style accepted in the repository you're committing to. Let's say you like to commit in small, tiny chunks. You have one commit to fix a typo, one commit to get rid of unused code and so on. By the time you've finished what you need to do, you have a long series of commits. Now let's say the repository you're committing to encourages large commits, so for the work you're doing, one would expect one or maybe two commits. How do you take your string of commits and compress them to what is expected? You would use an interactive rebase and squash your tiny commits into fewer larger chunks. The same is true if the reverse was needed - if your style was a few large commits, but the repository demanded long strings of small commits. You would use a rebase to do that as well. If you had merged instead, you have now grafted your commit style onto the main repository. If there are a lot of developers, you can imagine how hard it would be to follow a history with several different commit styles after some time.

UPDATE 3: Does one still need to merge after a successful rebase? Yes, you do. The reason is that a rebase essentially involves a "shifting" of commits. As I've said above, these commits are calculated, but if you had 14 commits from the point of branching, then assuming nothing goes wrong with your rebase, you will be 14 commits ahead (of the point you're rebasing onto) after the rebase is done. You had a branch before a rebase. You will have a branch of the same length after. You still need to merge before you publish your changes. In other words, rebase as many times as you want (again, only if you have not pushed your changes upstream). Merge only after you rebase.


A merge with master could result in a fast forward. In a feature branch there may be some commits, which have minor bugs or dont't even compile. If you do only unit testing in a feature branch, some errors in integration my slip through. Before merging with master, integration tests are required and can show some bugs. If these are fixed, the feature could be integrated. As you don't wish to commit buggy code to master, a rebase seems neccessary in order to prevent an all-commits-fast-forward.
@mbx git merge supports the --no-ff option which forces it to make a merge commit.
It doesn't really "depend" when you can gain the same advantages of a "tidy history" without the massive disadvantages of it by simply using search / filter commands to view the history. Making rebasing practically useless.
P
Peter Mortensen

While merging is definitely the easiest and most common way to integrate changes, it's not the only one: Rebase is an alternative means of integration.

Understanding Merge a Little Better

When Git performs a merge, it looks for three commits:

(1) Common ancestor commit. If you follow the history of two branches in a project, they always have at least one commit in common: at this point in time, both branches had the same content and then evolved differently.

(2) + (3) Endpoints of each branch. The goal of an integration is to combine the current states of two branches. Therefore, their respective latest revisions are of special interest. Combining these three commits will result in the integration we're aiming for.

Fast-Forward or Merge Commit

In very simple cases, one of the two branches doesn't have any new commits since the branching happened - its latest commit is still the common ancestor.

https://i.stack.imgur.com/57Wt0.gif

In this case, performing the integration is dead simple: Git can just add all the commits of the other branch on top of the common ancestor commit. In Git, this simplest form of integration is called a "fast-forward" merge. Both branches then share the exact same history.

https://i.stack.imgur.com/f1VTv.gif

In a lot of cases, however, both branches moved forward individually.

https://i.stack.imgur.com/GlAgx.gif

To make an integration, Git will have to create a new commit that contains the differences between them - the merge commit.

https://i.stack.imgur.com/e3xJE.gif

Human Commits & Merge Commits

Normally, a commit is carefully created by a human being. It's a meaningful unit that wraps only related changes and annotates them with a comment.

A merge commit is a bit different: instead of being created by a developer, it gets created automatically by Git. And instead of wrapping a set of related changes, its purpose is to connect two branches, just like a knot. If you want to understand a merge operation later, you need to take a look at the history of both branches and the corresponding commit graph.

Integrating with Rebase

Some people prefer to go without such automatic merge commits. Instead, they want the project's history to look as if it had evolved in a single, straight line. No indication remains that it had been split into multiple branches at some point.

https://i.stack.imgur.com/b1nud.gif

Let's walk through a rebase operation step by step. The scenario is the same as in the previous examples: we want to integrate the changes from branch-B into branch-A, but now by using rebase.

https://i.stack.imgur.com/Az5LO.gif

We will do this in three steps

git rebase branch-A // Synchronises the history with branch-A git checkout branch-A // Change the current branch to branch-A git merge branch-B // Merge/take the changes from branch-B to branch-A

First, Git will "undo" all commits on branch-A that happened after the lines began to branch out (after the common ancestor commit). However, of course, it won't discard them: instead you can think of those commits as being "saved away temporarily".

https://i.stack.imgur.com/To3Iq.gif

Next, it applies the commits from branch-B that we want to integrate. At this point, both branches look exactly the same.

https://i.stack.imgur.com/Vrfmo.gif

In the final step, the new commits on branch-A are now reapplied - but on a new position, on top of the integrated commits from branch-B (they are re-based).

The result looks like development had happened in a straight line. Instead of a merge commit that contains all the combined changes, the original commit structure was preserved.

https://i.stack.imgur.com/4IKL0.gif

Finally, you get a clean branch branch-A with no unwanted and auto generated commits.

Note: Taken from the awesome post by git-tower. The disadvantages of rebase is also a good read in the same post.


P
Peter Mortensen

Before merge/rebase:

A <- B <- C    [master]
^
 \
  D <- E       [branch]

After git merge master:

A <- B <- C
^         ^
 \         \
  D <- E <- F

After git rebase master:

A <- B <- C <- D' <- E'

(A, B, C, D, E and F are commits)

This example and much more well illustrated information about Git can be found in Git The Basics Tutorial.


( stackoverflow.com/a/804178 ) Rebase into our own dev branch,then merge into master? "You can use rebase to first bring master "in" the new-feature branch: the rebase will replay new-feature commits from the HEAD master, but still in the new-feature branch, effectively moving your branch starting point from an old master commit to HEAD-master.That allows you to resolve any conflicts in your branch (meaning, in isolation,).Then you can switch to master and merge new-feature (or rebase new-feature onto master if you want to preserve commits done in your new-feature branch)."
s
sp00m

This answer is widely oriented around Git Flow. The tables have been generated with the nice ASCII Table Generator, and the history trees with this wonderful command (aliased as git lg):

git log --graph --abbrev-commit --decorate --date=format:'%Y-%m-%d %H:%M:%S' --format=format:'%C(bold blue)%h%C(reset) - %C(bold cyan)%ad%C(reset) %C(bold green)(%ar)%C(reset)%C(bold yellow)%d%C(reset)%n''          %C(white)%s%C(reset) %C(dim white)- %an%C(reset)'

Tables are in reverse chronological order to be more consistent with the history trees. See also the difference between git merge and git merge --no-ff first (you usually want to use git merge --no-ff as it makes your history look closer to the reality):

git merge

Commands:

Time          Branch "develop"             Branch "features/foo"
------- ------------------------------ -------------------------------
15:04   git merge features/foo
15:03                                  git commit -m "Third commit"
15:02                                  git commit -m "Second commit"
15:01   git checkout -b features/foo
15:00   git commit -m "First commit"

Result:

* 142a74a - YYYY-MM-DD 15:03:00 (XX minutes ago) (HEAD -> develop, features/foo)
|           Third commit - Christophe
* 00d848c - YYYY-MM-DD 15:02:00 (XX minutes ago)
|           Second commit - Christophe
* 298e9c5 - YYYY-MM-DD 15:00:00 (XX minutes ago)
            First commit - Christophe

git merge --no-ff

Commands:

Time           Branch "develop"              Branch "features/foo"
------- -------------------------------- -------------------------------
15:04   git merge --no-ff features/foo
15:03                                    git commit -m "Third commit"
15:02                                    git commit -m "Second commit"
15:01   git checkout -b features/foo
15:00   git commit -m "First commit"

Result:

*   1140d8c - YYYY-MM-DD 15:04:00 (XX minutes ago) (HEAD -> develop)
|\            Merge branch 'features/foo' - Christophe
| * 69f4a7a - YYYY-MM-DD 15:03:00 (XX minutes ago) (features/foo)
| |           Third commit - Christophe
| * 2973183 - YYYY-MM-DD 15:02:00 (XX minutes ago)
|/            Second commit - Christophe
* c173472 - YYYY-MM-DD 15:00:00 (XX minutes ago)
            First commit - Christophe

git merge vs git rebase

First point: always merge features into develop, never rebase develop from features. This is a consequence of the Golden Rule of Rebasing:

The golden rule of git rebase is to never use it on public branches.

In other words:

Never rebase anything you've pushed somewhere.

I would personally add: unless it's a feature branch AND you and your team are aware of the consequences.

So the question of git merge vs git rebase applies almost only to the feature branches (in the following examples, --no-ff has always been used when merging). Note that since I'm not sure there's one better solution (a debate exists), I'll only provide how both commands behave. In my case, I prefer using git rebase as it produces a nicer history tree :)

Between feature branches

git merge

Commands:

Time           Branch "develop"              Branch "features/foo"           Branch "features/bar"
------- -------------------------------- ------------------------------- --------------------------------
15:10   git merge --no-ff features/bar
15:09   git merge --no-ff features/foo
15:08                                                                    git commit -m "Sixth commit"
15:07                                                                    git merge --no-ff features/foo
15:06                                                                    git commit -m "Fifth commit"
15:05                                                                    git commit -m "Fourth commit"
15:04                                    git commit -m "Third commit"
15:03                                    git commit -m "Second commit"
15:02   git checkout -b features/bar
15:01   git checkout -b features/foo
15:00   git commit -m "First commit"

Result:

*   c0a3b89 - YYYY-MM-DD 15:10:00 (XX minutes ago) (HEAD -> develop)
|\            Merge branch 'features/bar' - Christophe
| * 37e933e - YYYY-MM-DD 15:08:00 (XX minutes ago) (features/bar)
| |           Sixth commit - Christophe
| *   eb5e657 - YYYY-MM-DD 15:07:00 (XX minutes ago)
| |\            Merge branch 'features/foo' into features/bar - Christophe
| * | 2e4086f - YYYY-MM-DD 15:06:00 (XX minutes ago)
| | |           Fifth commit - Christophe
| * | 31e3a60 - YYYY-MM-DD 15:05:00 (XX minutes ago)
| | |           Fourth commit - Christophe
* | |   98b439f - YYYY-MM-DD 15:09:00 (XX minutes ago)
|\ \ \            Merge branch 'features/foo' - Christophe
| |/ /
|/| /
| |/
| * 6579c9c - YYYY-MM-DD 15:04:00 (XX minutes ago) (features/foo)
| |           Third commit - Christophe
| * 3f41d96 - YYYY-MM-DD 15:03:00 (XX minutes ago)
|/            Second commit - Christophe
* 14edc68 - YYYY-MM-DD 15:00:00 (XX minutes ago)
            First commit - Christophe

git rebase

Commands:

Time           Branch "develop"              Branch "features/foo"           Branch "features/bar"
------- -------------------------------- ------------------------------- -------------------------------
15:10   git merge --no-ff features/bar
15:09   git merge --no-ff features/foo
15:08                                                                    git commit -m "Sixth commit"
15:07                                                                    git rebase features/foo
15:06                                                                    git commit -m "Fifth commit"
15:05                                                                    git commit -m "Fourth commit"
15:04                                    git commit -m "Third commit"
15:03                                    git commit -m "Second commit"
15:02   git checkout -b features/bar
15:01   git checkout -b features/foo
15:00   git commit -m "First commit"

Result:

*   7a99663 - YYYY-MM-DD 15:10:00 (XX minutes ago) (HEAD -> develop)
|\            Merge branch 'features/bar' - Christophe
| * 708347a - YYYY-MM-DD 15:08:00 (XX minutes ago) (features/bar)
| |           Sixth commit - Christophe
| * 949ae73 - YYYY-MM-DD 15:06:00 (XX minutes ago)
| |           Fifth commit - Christophe
| * 108b4c7 - YYYY-MM-DD 15:05:00 (XX minutes ago)
| |           Fourth commit - Christophe
* |   189de99 - YYYY-MM-DD 15:09:00 (XX minutes ago)
|\ \            Merge branch 'features/foo' - Christophe
| |/
| * 26835a0 - YYYY-MM-DD 15:04:00 (XX minutes ago) (features/foo)
| |           Third commit - Christophe
| * a61dd08 - YYYY-MM-DD 15:03:00 (XX minutes ago)
|/            Second commit - Christophe
* ae6f5fc - YYYY-MM-DD 15:00:00 (XX minutes ago)
            First commit - Christophe

From develop to a feature branch

git merge

Commands:

Time           Branch "develop"              Branch "features/foo"           Branch "features/bar"
------- -------------------------------- ------------------------------- -------------------------------
15:10   git merge --no-ff features/bar
15:09                                                                    git commit -m "Sixth commit"
15:08                                                                    git merge --no-ff develop
15:07   git merge --no-ff features/foo
15:06                                                                    git commit -m "Fifth commit"
15:05                                                                    git commit -m "Fourth commit"
15:04                                    git commit -m "Third commit"
15:03                                    git commit -m "Second commit"
15:02   git checkout -b features/bar
15:01   git checkout -b features/foo
15:00   git commit -m "First commit"

Result:

*   9e6311a - YYYY-MM-DD 15:10:00 (XX minutes ago) (HEAD -> develop)
|\            Merge branch 'features/bar' - Christophe
| * 3ce9128 - YYYY-MM-DD 15:09:00 (XX minutes ago) (features/bar)
| |           Sixth commit - Christophe
| *   d0cd244 - YYYY-MM-DD 15:08:00 (XX minutes ago)
| |\            Merge branch 'develop' into features/bar - Christophe
| |/
|/|
* |   5bd5f70 - YYYY-MM-DD 15:07:00 (XX minutes ago)
|\ \            Merge branch 'features/foo' - Christophe
| * | 4ef3853 - YYYY-MM-DD 15:04:00 (XX minutes ago) (features/foo)
| | |           Third commit - Christophe
| * | 3227253 - YYYY-MM-DD 15:03:00 (XX minutes ago)
|/ /            Second commit - Christophe
| * b5543a2 - YYYY-MM-DD 15:06:00 (XX minutes ago)
| |           Fifth commit - Christophe
| * 5e84b79 - YYYY-MM-DD 15:05:00 (XX minutes ago)
|/            Fourth commit - Christophe
* 2da6d8d - YYYY-MM-DD 15:00:00 (XX minutes ago)
            First commit - Christophe

git rebase

Commands:

Time           Branch "develop"              Branch "features/foo"           Branch "features/bar"
------- -------------------------------- ------------------------------- -------------------------------
15:10   git merge --no-ff features/bar
15:09                                                                    git commit -m "Sixth commit"
15:08                                                                    git rebase develop
15:07   git merge --no-ff features/foo
15:06                                                                    git commit -m "Fifth commit"
15:05                                                                    git commit -m "Fourth commit"
15:04                                    git commit -m "Third commit"
15:03                                    git commit -m "Second commit"
15:02   git checkout -b features/bar
15:01   git checkout -b features/foo
15:00   git commit -m "First commit"

Result:

*   b0f6752 - YYYY-MM-DD 15:10:00 (XX minutes ago) (HEAD -> develop)
|\            Merge branch 'features/bar' - Christophe
| * 621ad5b - YYYY-MM-DD 15:09:00 (XX minutes ago) (features/bar)
| |           Sixth commit - Christophe
| * 9cb1a16 - YYYY-MM-DD 15:06:00 (XX minutes ago)
| |           Fifth commit - Christophe
| * b8ddd19 - YYYY-MM-DD 15:05:00 (XX minutes ago)
|/            Fourth commit - Christophe
*   856433e - YYYY-MM-DD 15:07:00 (XX minutes ago)
|\            Merge branch 'features/foo' - Christophe
| * 694ac81 - YYYY-MM-DD 15:04:00 (XX minutes ago) (features/foo)
| |           Third commit - Christophe
| * 5fd94d3 - YYYY-MM-DD 15:03:00 (XX minutes ago)
|/            Second commit - Christophe
* d01d589 - YYYY-MM-DD 15:00:00 (XX minutes ago)
            First commit - Christophe

Side notes

git cherry-pick

When you just need one specific commit, git cherry-pick is a nice solution (the -x option appends a line that says "(cherry picked from commit...)" to the original commit message body, so it's usually a good idea to use it - git log <commit_sha1> to see it):

Commands:

Time           Branch "develop"              Branch "features/foo"                Branch "features/bar"
------- -------------------------------- ------------------------------- -----------------------------------------
15:10   git merge --no-ff features/bar
15:09   git merge --no-ff features/foo
15:08                                                                    git commit -m "Sixth commit"
15:07                                                                    git cherry-pick -x <second_commit_sha1>
15:06                                                                    git commit -m "Fifth commit"
15:05                                                                    git commit -m "Fourth commit"
15:04                                    git commit -m "Third commit"
15:03                                    git commit -m "Second commit"
15:02   git checkout -b features/bar
15:01   git checkout -b features/foo
15:00   git commit -m "First commit"

Result:

*   50839cd - YYYY-MM-DD 15:10:00 (XX minutes ago) (HEAD -> develop)
|\            Merge branch 'features/bar' - Christophe
| * 0cda99f - YYYY-MM-DD 15:08:00 (XX minutes ago) (features/bar)
| |           Sixth commit - Christophe
| * f7d6c47 - YYYY-MM-DD 15:03:00 (XX minutes ago)
| |           Second commit - Christophe
| * dd7d05a - YYYY-MM-DD 15:06:00 (XX minutes ago)
| |           Fifth commit - Christophe
| * d0d759b - YYYY-MM-DD 15:05:00 (XX minutes ago)
| |           Fourth commit - Christophe
* |   1a397c5 - YYYY-MM-DD 15:09:00 (XX minutes ago)
|\ \            Merge branch 'features/foo' - Christophe
| |/
|/|
| * 0600a72 - YYYY-MM-DD 15:04:00 (XX minutes ago) (features/foo)
| |           Third commit - Christophe
| * f4c127a - YYYY-MM-DD 15:03:00 (XX minutes ago)
|/            Second commit - Christophe
* 0cf894c - YYYY-MM-DD 15:00:00 (XX minutes ago)
            First commit - Christophe

git pull --rebase

I am not sure I can explain it better than Derek Gourlay... Basically, use git pull --rebase instead of git pull :) What's missing in the article though, is that you can enable it by default:

git config --global pull.rebase true

git rerere

Again, nicely explained here. But put simply, if you enable it, you won't have to resolve the same conflict multiple times anymore.


P
Peter Mortensen

This sentence gets it:

In general, the way to get the best of both worlds is to rebase local changes you’ve made, but haven’t shared yet, before you push them in order to clean up your story, but never rebase anything you’ve pushed somewhere.

Source: 3.6 Git Branching - Rebasing, Rebase vs. Merge


P
Peter Mortensen

The Pro Git book has a really good explanation on the rebasing page.

Basically a merge will take two commits and combine them.

A rebase will go to the common ancestor on the two and incrementally apply the changes on top of each other. This makes for a 'cleaner' and more linear history.

But when you rebase, you abandon previous commits and create new ones. So you should never rebase a repository that is public. The other people working on the repository will hate you.

For that reason alone I almost exclusively merge. 99% of the time my branches don’t differ that much, so if there are conflicts it's only in one or two places.


Merges do not combine commits - that would be rewriting history. Rebase does that.
I'm not sure why you can't rebase on a feature branch, then merge on a public branch.
P
Peter Mortensen

Git rebase is used to make the branching paths in history cleaner and repository structure linear.

It is also used to keep the branches created by you private, as after rebasing and pushing the changes to the server, if you delete your branch, there will be no evidence of branch you have worked on. So your branch is now your local concern.

After doing rebase we also get rid of an extra commit which we used to see if we do a normal merge.

And yes, one still needs to do merge after a successful rebase as the rebase command just puts your work on top of the branch you mentioned during rebase, say master, and makes the first commit of your branch as a direct descendant of the master branch. This means we can now do a fast forward merge to bring changes from this branch to the master branch.


P
Peter Mortensen

Some practical examples, somewhat connected to large scale development where Gerrit is used for review and delivery integration:

I merge when I uplift my feature branch to a fresh remote master. This gives minimal uplift work and it's easy to follow the history of the feature development in for example gitk.

git fetch
git checkout origin/my_feature
git merge origin/master
git commit
git push origin HEAD:refs/for/my_feature

I merge when I prepare a delivery commit.

git fetch
git checkout origin/master
git merge --squash origin/my_feature
git commit
git push origin HEAD:refs/for/master

I rebase when my delivery commit fails integration for whatever reason, and I need to update it towards a fresh remote master.

git fetch
git fetch <gerrit link>
git checkout FETCH_HEAD
git rebase origin/master
git push origin HEAD:refs/for/master

J
Jeremy Benks

It was explained many times what rebase and what merge is, but when should you use what?

When should you use rebase?

Rebase "lifts off" your changes and puts all the changes of the rebased branch into your current branch and then puts your changes on top of it. It therefore changes the history of your branch.

when you have not pushed the branch / no one else is working on it

you want your to see all your changes at one point together when merging back to the source branch

you want to avoid the auto-generated "merged .." commit messages

I said "you want to see all your changes at one place" because sometimes a merge operation puts all your changes together in one commit (some: merged from ... message). Rebase makes your change look like you made your commits all after each other with no one else doing something in between. This makes it easier to see, what you changed for your feature.

Make sure though, you use git merge feature-branch --ff-only to make sure there are no conflicts creating a single commit when you are merging your feature back to develop/master.

When should you use merge?

when you have pushed the branch / others are working on it too (rebase gets very complicated if others work on that branch too!)

you don't need the full history(*) / your feature doesn't have to have it's commits all in one place.

(*) you can avoid that your feature only gets one "merged .." commits by first merging the develop branch to your feature and then merging your feature back to develeop. This still gives you a "merged .." commit, but a least all the commits of your feature are still visible.


It is the other way round. If you rebase, your history is rewritten and therefore some information is lost. Merge doesn't alter nor loose any history, so your points are wrong. The one most important point you miss is that the rebase means that you have a linear history. Your answer misses the point of rebasing!
M
Marnen Laibow-Koser

When do I use git rebase? Almost never, because it rewrites history. git merge is almost always the preferable choice, because it respects what actually happened in your project.


@benjaminhull Thanks!—except I hope my answer is fact-based. IMHO opinion has little place in this sort of thing: it's a fact that losing your actual history makes life harder later.
Agree. Merge will never lead to corrupted history etc. (when you rebase your pushed commits)
I don't know why this answer has lots of downvotes. I had to upvote to minimize the damage. LOL. I partially agree, but I think we can rebase if we're the only one working in the branch to keep everything cleaner. The main problem is if for some unexpected reason others start working on it as well.