Inspecting history

Objectives

Be able find a line of code, find out why it was introduced and when.
Be able to quickly find the commit that changed a behavior.

Instructor note

30 min teaching/type-along
30 min exercise

Preparation

Please make sure that you do not clone repositories inside an already tracked folder:

$ git status

If you are inside an existing Git repository, step out of it. You need to find a different location since we will clone a new repository.

If you see this message, this is good in this case:

fatal: not a git repository (or any of the parent directories): .git

Our toolbox for history inspection

Instructor note

First the instructor demonstrates few commands on a real life example repository https://github.com/networkx/networkx (mentioned in the amazing site The Programming Historian). Later we will practice these in an archaeology exercise (below).

Warm-up: “Git History” browser 

As a warm-up we can try the “Git History” browser on the README.rst file of the networkx repository:

Visit and browse https://github.githistory.xyz/networkx/networkx/blob/main/README.rst (use left/right keys).
You can try this on some of your GitHub repositories, too!

Searching text patterns in the repository

With git grep you can find all lines in a repository which contain some string or regular expression. This is useful to find out where in the code some variable is used or some error message printed:

$ git grep TEXT
$ git grep "some text with spaces"

In the networkx repository you can try:

$ git clone https://github.com/networkx/networkx
$ cd networkx
$ git grep -i fixme

While git grep searches the current state of the repository, it is also possible to search through all changes with git log -S sometext which can be useful to find where something got removed.

Inspecting individual commits

We have seen this one before already. Using git show we can inspect an individual commit if we know its hash:

$ git show HASH

For instance:

$ git show 759d589bdfa61aff99e0535938f14f67b01c83f7

Line-by-line code annotation with metadata

With git annotate you can see line by line who and when the line was modified last. It also prints the precise hash of the last change which modified each line. Incredibly useful for reproducibility.

$ git annotate FILE

Example:

$ git annotate networkx/convert_matrix.py

If you annotate in a terminal and the file is longer than the screen, Git by default uses the program less to scroll the output. Use /sometext <ENTER> to find “sometext” and you can cycle through the results with n (next) and N (last). You can also use page up/down to scroll. You can quit with q.

Discussion

Discuss how these relatively trivial changes affect the annotation:

Wrapping long lines of text/code into shorter lines
Auto-formatting tools such as black
Editors that automatically remove trailing whitespace

Inspecting code in the past

We can create branches pointing to a commit in the past. This is the recommended mechanism to inspect old code:

$ git switch --create BRANCHNAME HASH

Example (lines starting with “#” are only comments):

$ # create branch called "older-code" from hash 347e6292419b
$ git switch --create older-code 347e6292419bd0e4bff077fe971f983932d7a0e9

$ # now you can navigate and inspect the code as it was back then
$ # ...

$ # after we are done we can switch back to "main"
$ git switch main

$ # if we like we can delete the "older-code" branch
$ git branch -d older-code

On old Git versions which do not know the switch command (before 2.23), you need to use this instead:

$ git checkout -b BRANCHNAME SOMEHASH

Exercise: Basic archaeology commands

History-1: Explore basic archaeology commands

Let us explore the value of these commands in an exercise. Future exercises do not depend on this, so it is OK if you do not complete it fully.

Exercise steps:

Clone this repository: https://github.com/networkx/networkx.git. Then step into the new directory and create an exercise branch from the networkx-2.6.3 tag/release:
```
$ git clone https://github.com/networkx/networkx.git
$ cd networkx
$ git switch --create exercise networkx-2.6.3
```
On old Git versions which do not know the switch command (before 2.23), you need to use this instead:
```
$ git checkout -b exercise networkx-2.6.3
```

Then using the above toolbox try to:

Find the code line which contains "Logic error in degree_correlation".
Find out when this line was last modified or added. Find the actual commit which modified that line.
Inspect that commit with git show.
Create a branch pointing to the past when that commit was created to be able to browse and use the code as it was back then.
How would you bring the code to the version of the code right before that line was last modified?

Solution

We provide here a solution for the command line but we also encourage you to try to solve this in the browser.

We use git grep:

$ git grep "Logic error in degree_correlation"

This gives the output:

networkx/algorithms/threshold.py:                print("Logic error in degree_correlation", i, rdi)

Maybe you also want to know the line number:

$ git grep -n "Logic error in degree_correlation"

We use git annotate:
```
$ git annotate networkx/algorithms/threshold.py
```
Then search for “Logic error” by typing “/Logic error” followed by Enter. The last commit that modified it was 90544b4fa (unless that line changed since).
We use git show:
```
$ git show 90544b4fa
```
Create a branch pointing to that commit (here we called the branch “past-code”):
```
$ git branch past-code 90544b4fa
```
This is a compact way to access the first parent of 90544b4fa (here we called the branch “just-before”):
```
$ git switch --create just-before 90544b4fa~1
```

Finding out when something broke/changed with `git bisect`

“But I am sure it used to work! Strange.”

Sometimes you realize that something broke. You know that it used to work. You do not know when it broke.

How would you solve this?

Before we go on first discuss how you would solve this problem: You know that it worked 500 commits ago but it does not work now.

How would you find the commit which changed it?
Why could it be useful to know the commit that changed it?

We will probably arrive at a solution which is similar to git bisect:

First find out a commit in past when it worked.

$ git bisect start
$ git bisect good f0ea950  # this is a commit that worked
$ git bisect bad main      # last commit is broken

Now compile and/or run and/or test and decide whether “good” or “bad”.
This is how you can tell Git that this was a working commit:
```
$ git bisect good
```
And this is how you can tell Git that this was not a working commit:
```
$ git bisect bad
```
Then bisect/iterate your way until you find the commit that broke it.
If you want to go back to start, type git bisect reset.
This can even be automatized with git bisect run SCRIPT. For this you write a script that returns zero/non-zero (success/failure).

Optional exercise: Git bisect

(optional) History-2: Use git bisect to find the bad commit

In this exercise, we use git bisect on an example repository. It is OK if you do not complete this exercise fully.

Begin by cloning https://github.com/coderefinery/git-bisect-exercise.

Motivation

The motivation for this exercise is to be able to do archaeology with Git on a source code where the bug is difficult to see visually. Finding the offending commit is often more than half the debugging.

Background

The script get_pi.py approximates pi using terms of the Nilakantha series. It should produce 3.14 but it does not. The script broke at some point and produces 3.57 using the last commit:

$ python get_pi.py

3.57

At some point within the 500 first commits, an error was introduced. The only thing we know is that the first commit worked correctly.

Your task

Clone this repository and use git bisect to find the commit which broke the computation (solution - spoiler alert!).
Once you have found the offending commit, also practice navigating to the last good commit.
Bonus exercise: Write a script that checks for a correct result and use git bisect run to find the offending commit automatically (solution - spoiler alert!).

Hints

Finding the first commit:

$ git log --oneline | tail -n 1

How to navigate to the parent of a commit with hash SOMEHASH:

$ git switch --create BRANCHNAME SOMEHASH~1

Instead of a tilde you can also use this:

$ git switch --create BRANCHNAME SOMEHASH^

Keypoints

git log/grep/annotate/show/bisect is a powerful combination when doing archaeology in a project.
git switch --create NAME HASH is the recommended mechanism to inspect old code.