Improve renames in git blog post

main
Gaël Depreeuw 4 years ago
parent 2d2b9afe6c
commit 2d13360bd7
Signed by: Mithror
GPG Key ID: 8AB218ABA4867F78
  1. 273
      content/post/renames-in-git-explained.md

@ -14,168 +14,176 @@ handles file and/or directory renames. The short answer to this is: **It**
**doesn't**. **doesn't**.
The slightly longer answer is: **It does, but probably not in the way you** The slightly longer answer is: **It does, but probably not in the way you**
**envision it**. **envision it?**.
To help you understand this topic a bit more, we first have to go back to the Let's first take a look at how Git works internally. If you don't quite
basics: What actually is a file or directory name? The answer to this question understand everything which follows, I can recommend reading chapter 10 of
is highly dependent on the underlying file system, but in general it can be the [Git Pro Book 2nd. Edition](https://git-scm.com/book/en/v2/Git-Internals-Plumbing-and-Porcelain).
boiled down to this:
> A file (or directory) name is an index used by the file system to look up the
> contents of the file. (Note: from now on I will only refer to file names, but
> the same applies to directory names as well.)
What you should note from this is that a filename is actually not a property of
the file content itself, but part of the meta-data regarding the content. In
Linux, for instance, the filename of a file is stored in the directory, which
is basically a associative array which maps filenames to inodes (the object
which stores the meta-data of a file).
When renaming a file, what you are actually doing is updating a look up table.
In Linux, this would be updating the associative array of the directory. If you
move a file, then you remove the element from one directory and add it to
another directory.
How this all works internally depends on the OS and the underlying file system,
but more importantly is seldom related to the content of a file. Which brings us
to the next chapter.
## Git stores content, not files ## Git stores content, not files
When you commit to a Git repository it basically does the following: When you commit to a Git repository it basically does the following:
For each directory (including the top one), create a **tree** object. This is Create a **blob** object for every file in the index (a.k.a. the staging area).
done by looking at every file and directory to be commited and create **blob** A blob object is created by taking the content of the file, prepending a header
objects for the files and tree objects for the directories. The hash of each and compressing the result. A SHA-1 hash is then calculated for this object
such object is added to this tree object together with the filename if the which will be used to identify the object. The object is stored in the aptly
object type is blob and the directory name if the object type is tree. This is named object store (found in `.git/objects`). The first 2 characters of the
then prepended with a header and compressed. The SHA-1 hash is calculated and hash (in hex format) are used as a directory within this store, while the
the object is stored in the object store (.git/objects), using the first two remaining characters are the filename of the blob object.
characters as a directory and the rest as filename.
It then creates a commit object which points to the top level tree's hash. Let's look at an example. Create a git repo somewhere and create a file.
> Note: it of course only really does this for files which were part of the
> staging area. That's the most efficient. Of course if the content of a file
> was changed, it hash will change and thus the tree object it was part of will
> change and its hash will also change and so on until the top level tree
> object.
As an example, suppose you have the following structure:
```bash ```bash
. git init foo
├── README.md cd foo
├── bar echo "foo" >> foo.txt
   ├── bar.md
   └── baz
   └── baz.md
└── foo
└── foo.md
``` ```
If you were to commit this structure to git, you will have (simplified): If you look into your `.git/objects` directory, it will be empty, aside from
two empty subdirectories. Let's create a blob out of this file now.
- 4 blob objects (README.md, bar.md, foo.md, baz.md) ```bash
- 4 tree objects (., ./foo, ./bar and ./bar/baz) $ git hash-object -w foo.txt
- 1 commit object 257cc5642cb1a054f08cc83f2d943e56fd3ebe99
```
In my case: You will now find an object in the store at:
```bash ```bash
gael@Aviendha:~/git/tmp$ git commit -m "First commit" $ find .git/objects -type f
[master (root-commit) 8be3cf0] First commit
4 files changed, 4 insertions(+)
create mode 100644 README.md
create mode 100644 bar/bar.md
create mode 100644 bar/baz/baz.md
create mode 100644 foo/foo.md
gael@Aviendha:~/git/tmp$ find .git/objects/ -type f
.git/objects/52/01cdd884658a103819d66f910ea25ba1dad2e0
.git/objects/be/e527307ae70706c20eb89f205f444c3bb385e9
.git/objects/6b/dd34e3e9ab26062ab881adb1024923923b5f8e
.git/objects/8b/e3cf05d01320a124991a8e7c10fe83ec9cd5e3
.git/objects/25/7cc5642cb1a054f08cc83f2d943e56fd3ebe99 .git/objects/25/7cc5642cb1a054f08cc83f2d943e56fd3ebe99
.git/objects/57/16ca5987cbf97d6bb54920bea6adde242d87e6
.git/objects/f9/07d059fcdc9b594c6e14dc0c3826f26ab47832
.git/objects/e8/45566c06f9bf557d35e8292c37cf05d97a9769
.git/objects/0c/7d27db1f575263efdcab3dc650f4502a2dbcbf
``` ```
To get the top level tree object, just look at the commit: Here's an interesting exercise: what happens if you rename the file and create
the blob with the renamed file?
```bash ```bash
gael@Aviendha:~/git/tmp$ git cat-file -p 8be3cf0 $ mv foo.txt bar.txt
tree 5201cdd884658a103819d66f910ea25ba1dad2e0 $ git hash-object -w bar.txt
author Gaël Depreeuw <gael@depreeuw.dev> 1606569688 +0100 257cc5642cb1a054f08cc83f2d943e56fd3ebe99
committer Gaël Depreeuw <gael@depreeuw.dev> 1606569688 +0100
First commit
``` ```
And if we look at the tree: That's right, nothing changed, this makes sense as we're only adding the content
to the object store! So how does Git remember the file names?
```bash ## Filenames are part of tree objects
gael@Aviendha:~/git/tmp$ git cat-file -p 5201cdd
100644 blob e845566c06f9bf557d35e8292c37cf05d97a9769 README.md Aside from **blob** objects, Git also creates **tree** objects. You can sort of
040000 tree f907d059fcdc9b594c6e14dc0c3826f26ab47832 bar compare it to the directories in your worktree, i.e. for each directory in your
040000 tree 0c7d27db1f575263efdcab3dc650f4502a2dbcbf foo worktree, you will have a tree object. A tree object's content looks like:
```code
<mode> <type> <hash> <name>
``` ```
The contents of `README.md` is: You can create this one for yourself by doing:
```bash ```bash
gael@Aviendha:~/git/tmp$ git cat-file -p e845566 $ git update-index --add --cacheinfo 100644 \
README 257cc5642cb1a054f08cc83f2d943e56fd3ebe99 foo.txt
$ git write-tree
fcf0be4d7e45f0ef9592682ad68e42270b0366b4
$ git cat-file -p fcf0be4d7e45f0ef9592682ad68e42270b0366b4
100644 blob 257cc5642cb1a054f08cc83f2d943e56fd3ebe99 foo.txt
``` ```
So what does this all mean, when we rename a file? There are 3 different types (that I know of) which can be referred to in a tree
object: blob, tree, commit. Blobs represent file content, tree represent other
tree (i.e. subdirectories) and commits represent submodules (i.e the commit
at which they are included). A commit is a type of object which is also present
outside the tree objects. They contain the top tree object (representing the
top level of your repository), a link to one or more parent commits and some
meta data (author, commit msg, date, ...). Finally and for completion's sake,
there is also an object for annotated tags, which contain the commit it is
pointing too as well as some meta data.
## Renaming file ## Renaming
If we're just looking at renaming a file, then the contents of the file will Armed with the knowledge about trees and blobs, it should be fairly easy to
not change. This means the blob object representing the file does not change. understand what happens if you rename a file. To make not make it easier to
What does change is: understand, consider a simple example: we just rename a file at the top level.
1. The old file's name is removed from the tree object it belong to. > Note: more complex examples are just more time consuming to explain, but
2. The new file's name is added to the tree object it belongs to (with the same > not to understand. The same principles apply.
hash in this case).
As such Git is not aware of any name changes. This is why the short answer is: In case of such a rename, when you commit this rename, your repository will
Git doesn't handle file renames. The repository itself has no notion of this be impacted as follows:
action. It's just has content and a structure for that content.
- The blob representing the file remains unchanged.
- The top level tree object changes as it now has a different file name.
- The commit object will point to the new tree. (It's parent will point to the
old tree.)
Nowhere is there any special mention of a rename occuring. Remember, we're just
storing content! As such Git is not aware of any name changes. This is why the
short answer was: Git doesn't handle file renames. The repository itself has no
notion of this action. It's just has content and a structure for that content.
However, that does not mean you lose your history when you rename a file. However, that does not mean you lose your history when you rename a file.
### How to see history of a renamed file ### How to see history of a renamed file
When you remove and add a file (which is what a rename is for Git), Git will Git might not store information on renames in it repository but it does come
analyze this and when the files are X% alike (with X being defaulted to 50), packed with an algorithm that detects file renames. For every add/delete pair
it will assume a rename occured. You can show the log of a file including added to the index, it determines how alike the paired files are. If they are
renames using: at least 50% alike, it considered the pair to have been a rename. If there
are multiple possibilities it takes the highest percentage one. If multipe files
have the same percentage, it picks one depending on the implementation.
```bash > **Note**: I believe, but am not sure, it basicaly takes the first
git log --follow -- <file> > alphabeticaly match in the last case.
```
If you want to adjust the treshold you can use the `-MX%` option, where X is the By default `git log -- <file>` does not track accross renames. If you want to
percentage you want (0-100). do see the history across renames, you will need to add the `--follow` option.
Because there is a percentage treshold, the recommendation is that you do not You can also define the treshold percentage to be different from 50%. This is
combine renaming a file, with modifying a file. If files are 100% identical when done via the `-M<n>` or `--find-renames=<n>` option. See the git documentation
adding/removing it makes it much easier to see them as renames. If on the other for the correct syntax.
hand, you rename a file and start modifying it heavily, Git might not detect
this as a rename, unless you lower the treshold.
You can also turn off rename detection by doing `--no-renames` You can also turn off rename detection by doing `--no-renames`
### Rename best practice
Because of the treshold and the cheapness of commits, it is recommended that
when you rename a file/directory. You commit those renames first, before you
continue working on the renamed file. This basically makes it so you can use
a treshold of 100% all the time.
### Why did Git do it this way
This is pure speculation, but here's my thoughts on it:
Filenames are actually part of the underlying file systems, so for a version
controls system to support multiple file system they have to handle filenames
in their own way. This includes renames. If you think about what this would
require for Git, it would not be very straightforward: Git could have chosen to
provide a command to store rename data, let's say: `git rename fileA fileB`, but
what should this command do? We can image it could create new '**rename**
object, which would hold the blob hash and the name of the previous file. Now,
every time you would walk through history, when you encounter this object type,
you would need to remember this redirections. There's probably a lot of little
nuances which are not immediately apparent though and it does not deal with one
of the major drawbacks of this new command: What happens if the user forgets it
and just does `mv fileA fileB`?
Well, we'd actually want to have some mechanism to detect this as a rename as
once this is commited it becomes more difficult to undo this change
(especially if we already pushed the commit!). So it sure would be nice if Git
could somehow figure out that it was a rename. Which is exactly what they did.
But now that we have this functionality, what actually is the point of the
new command we wanted to implement? This is probably highly subjective, but to
me it seems completely irrelevant now. Instead of having a command which can be
forgotten and for which we need contigency, just use the contigency as the
solution! It makes the behaviour a lot more consistent!
### Can I fix my commit if I did change a lot of content after renaming ### Can I fix my commit if I did change a lot of content after renaming
First, to prevent this: always check using `git status` whether are not the First, to prevent this: always check using `git status` whether are not the
rename is being detected. Now, how to solve it? rename is being detected. Now, how to solve it?
It depends. If your commit is local only and it is the last commit, then you can It depends. If your commit is local only and it is the last commit, then you can
fix this easily. There are many ways to to it, but a couple options are: fix this easily. There are many ways to to it, but one option is:
```bash ```bash
git mv <newname> <oldname> # Undo the file rename git mv <newname> <oldname> # Undo the file rename
@ -184,37 +192,40 @@ git mv <oldname> <newname> # Rename the file
git commit # Commit the rename git commit # Commit the rename
``` ```
If you want to rename first and the changes second you can also do this, but
it is a bit more complex:
```bash
git reset --soft HEAD~ # Go back one commit, but keep the changes
git restore --staged <oldname> <newname> # unstage the deletion and addition
git restore <oldname> # undelete the old file
mv <newname> <newname.tmp> # make a temp backup of the new file
git mv <oldname> <newname> # Rename the old file
git commit # commit the rename
cp <newname.tmp> <newname> # apply the new changes
git commit -a # Commit the changes
```
If the commit is already a couple of commits ago, you can do the same with an If the commit is already a couple of commits ago, you can do the same with an
interactive rebase and amending the commit at the right time. interactive rebase and doing either of the above at the correct time.
If you already pushed your commits you will have to check with the team if you If you already pushed your commits you will have to check with the team if you
can rewrite the history and push it. If this is not possible, you might need to can rewrite the history and push it. If this is not possible, you might need to
find the right treshold to have Git mark it as a rename. find the right treshold to have Git mark it as a rename.
### Why did Git do it this way
This is pure speculation but dealing with renames is not as easy as it first
looks. For instance you could add a git command to do a rename (like subversion
has), which could create a new type of object a rename object which links two
objects (old and new). But what if the user forgets to do this and just uses
`mv fileA fileB` and commits this? Should Git automatically assume this is a
rename? It could use the same treshold discused earlier to determine so. That
would make it easier. But then what is the point of having a dedicated rename
command? I think for easy of use, they just decided not to add such a command,
because it is not a solution for all instances. Instead, the rename detection
works good enough for everything and they leave it up to the commiter to make
sure his renames are detected properly.
## Summary ## Summary
So in summary: no, Git does not store renames in its repository. Instead, it So in summary: no, Git does not store renames in its repository. Instead, for
for every add/delete pair part of a commit, Git will do a likeness analysis and every add/delete pair in a commit, Git will do an similarity analysis and
when they are X% alike (default 50%), it will assume a rename occured. when they are X% alike (default 50%), it will assume a rename occured.
Some commands influenced by this are: git log, git diff and git merge. Options Some commands influenced by this are: `git log`, `git diff` and `git merge`.
related to renames are: Options related to renames are:
```txt ```txt
-M=<n>, --find-renames=<n> # where n is the treshold percentage. -M=<n>, --find-renames=<n> # where n is the treshold percentage.
--no-renames # don't do any rename detection --no-renames # don't do any rename detection
``` ```
It is best practise to handle renames in their own commits. Try to avoid
renaming and modifying a file within the same commit.

Loading…
Cancel
Save