Introduction to debugging with git bisect

As a stupid developer who consistently makes mistakes, it happens all the time to discover some bugs only after shipping them to production. Determining how deep the bug is, it could be really hard to find out which commit introduced it. We often just trace code and try to find the exact line which caused the bug, but sometimes it's happening everywhere and we can do better.

I only just discovered git bisect recently, and I felt so dumb for knowing it so late. It's a powerful tool built-in in git for debugging commits using binary search.

Given a list of sorted elements, what is the most performant way to find a specific element? We can go through the list one by one from one end and try to find it, but it's gonna be linear time (O(N)), which is good but not ideal. Another approach is to use binary search, since we already have all the elements sorted, we can first check if the element in the middle matches the criteria, and then eliminate half of the elements in the list and so on. Finally, we can get the element we want in O(log N). Guess what, this technique also works in debugging with git.

#An example

Consider we have a bug in our application, we want to find the commit which introduced this bug. For simplicity, the bug is that one of these commits has accidentally removed the file index.js.

* e25b6d2 (HEAD) Remove b.js
* 92f7e81 Add c.js
* ffa3ea1 Remove a.js
* aec5b41 Add b.js
* 9b970bf Add a.js
* 301fd10 Add index.js

#Introducing git bisect

First, we start the bisection, by running git bisect start.

$ git bisect start

Then we want to find two commits as the starting and ending point. One should be marked as good which means the index.js file is still present, and another should be marked as bad which means the index.js file is missing. In this case, we just have to mark HEAD as bad and the first commit 301fd10 as good.

$ git bisect bad # Mark current HEAD `e25b6d2` as `bad`
$ git checkout 301fd10
$ git bisect good # Mark the first commit as `good`

After we have both a commit marked as good and a commit marked as bad, the bisection will start automatically. The output will display something like below.

Bisecting: 2 revisions left to test after this (roughly 1 step)
[aec5b41c8c441ee23a6c484812ce9d7061962813] Add b.js

Now that we are being checked out to the commit in the middle between our starting and ending point. All we have to do now is to determine if this commit can reproduce the bug we are trying the find. In this case, we can just list the files in the directory and find whether index.js exists. Or run a simple test command.

$ test index.js | echo "Not found"
# echo "Not found" if... not found

If the result is empty, then it means that we haven't found the commit yet, we can then mark this commit as good and then perform another search. If the result is Not found, then it means this commit can reproduce the bug, but not necessarily the first commit which introduces the bug, we can mark it as bad and go on.

$ git bisect good
Bisecting: 0 revisions left to test after this (roughly 1 step)
[92f7e815080a215b40e3b8b5fe4495b48d4d3f32] Add c.js

We repeat the above methods over and over again until we reach the point where there are no commits left to be tested.

$ test index.js | echo "Not found"
Not found
$ git bisect bad
Bisecting: 0 revisions left to test after this (roughly 0 steps)
[ffa3ea19589ed0672a03f651919e3557be65f33d] Remove a.js
$ test index.js | echo "Not found"
Not found
$ git bisect bad
ffa3ea19589ed0672a03f651919e3557be65f33d is the first bad commit
commit ffa3ea19589ed0672a03f651919e3557be65f33d
Author: Kai Hao <kevin830726@gmail.com>
Date: Sun Apr 28 19:23:02 2019 +0800
Remove a.js
:100644 000000 e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 0000000000000000000000000000000000000000 D a.js
:100644 000000 e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 0000000000000000000000000000000000000000 D index.js

Notice the line said ffa3e... is the first bad commit, we found the commit which introduced the bug! Furthermore, git also gives us the summary of the commit, which in fact deletes the file index.js.

Now that we've finished finding what we want, the last step is to checkout to the original HEAD we were working on. What you are going to do with that commit is up to you. You can just fix the bug, comment it in code review, or even amend it in the commit then rebase.

$ git bisect reset

#Tips & Tricks

#Mark without checkout

You can simply mark the commit without checking out to it. Useful when starting the bisection.

$ git bisect bad HEAD # Mark `HEAD` as `bad`
$ git bisect good 301fd10 # Mark `301fd10` as `good`

#Start alias

Going further, we can start the bisection in one line.

# Start bisection with `HEAD` marked as `bad` and `301fd10` marked as `good`
$ git bisect start HEAD 301fd10 --

#Automation

If you have already set up a command to automatically check whether the commit is good or bad, like in the previous example we use test index.js to check whether index.js exists. We can actually automate the above steps with a command called git bisect run

$ git bisect start HEAD 301fd10 --
$ git bisect run test index.js
$ git bisect reset

This means that we can actually use bisect in CI environment, for tasks like finding the commit which broke the tests and generates report, or even hotfix it when necessary.

#Summing up

git bisect is a powerful tool I wished I knew. There are more tips and details this introduction doesn't cover, I recommend to see the official doc for more usage like log, view, old/new etc. Happing debugging!

#References

  • Git official doc on bisect
githubEdit on GithubBack to posts list