There are some interesting meetups going on in Madrid, and as part of the engineering team at source{d}, we often attend and organize some of them; it is a nice way to blow off some steam after work, learn some cool new things and meet quite interesting people.

A few weeks ago, at a Papers we love Madrid meetup, I had the chance to give a talk about how git blame works.*

Blaming a file in git is an operation that identifies who modified each of the lines in a file, along with the modification date and other details. This comes in handy when you want to know which member of your team wrote a particular piece of code or who is to blame for a bug }:-).

For example, let’s blame src/bufio/bufio.go from the standard Go Distribution:

$ git blame src/bufio/bufio.go
64776da4 src/pkg/bufio/bufio.go (Rob Pike          2011-12-13 15:07:17 -0800  40) const minReadBufferSize = 16
4ffc7992 src/pkg/bufio/bufio.go (Rui Ueyama        2014-03-24 11:48:34 -0700  41) const maxConsecutiveEmptyReads = 100

According to this git blame output, it seems that the declaration of the constant minReadBufferSize was last modified (or created) by Rob Pike in 2011 while the last version of maxConsecutiveEmptyReads is attributed to Rui Ueyama in 2014.

The current git blame implementation is a highly optimized and quite powerful piece of code, but its core functionality is easy to understand once you grasp a few concepts:

  1. Levenshtein Distance
  2. Longest Common Subsequence Problem
  3. Diff
  4. Tracking Lines Across File Revisions

The goal of the talk was to understand these concepts and some of their more naive implementations. This blog post is not going to be a full transcription of my talk, but allow me to tease you with a brief introduction to these interesting topics:

Levenshtein distance

The Levenshtein distance is a popular measure of how (dis)similar two strings are. More precisely, it is the minimum number of edits you have to perform on one of them to turn it into the other, where by edits I mean: adding, removing or changing a single character.

As expected, the Levenshtein distance of two identical strings is 0:

a := "pain"
Levenshtein(a, a) // is 0

For strings that differ only in one edit, the Levenshtein distance should be 1:

a := "pain"
b := "plain"
Levenshtein(a, b)  // is 1, just insert 'l' at a[1]

c := "pan"
Levenshtein(a, c)  // is 1, just remove 'i' from a[2]

d := "pawn"
Levenshtein(a, d)  // is 1, just change 'i' to 'w'

When the strings differ in more than one edit, calculating the Levenshtein distance is no longer trivial since there are many different combinations of edits that will allow you to turn one string into another, and only some of them will have the minimum number of edits:

e := "Lost"
f := "plot"
Levenshtein(e, f)  // is 3, here is one possible
                   //       combination of edits:
                   //   - change 'L' at e[0] to 'p'
                   //   - insert 'l' at e[1]
                   //   - remove 's' from e[3]

Calculating the Levenshtein distance of two strings is a fun and interesting programming workout. In the case that you get stuck on it, you will find a recursive solution, as well as a dynamic programming one, in the slides from my talk.

Longest Common Subsequence problem

The Longest Common Subsequence Problem (LCS for shorts), is a classic computer science problem, consisting in finding the longest subsequence common to all sequences in a given set.

For example, the LCS of the strings "pain" and "plans" is "pan", as it has all the characters common to both strings, without messing up the character ordering.

Solving this problem for longer strings is not trivial, though:


Knowing the LCS of two strings is equivalent to knowing the actual set of edits you need to perform to turn one into the other. You will find an intuitive explanation of this important equivalence in the slides from the talk.


The Diff Algorithm is the basis of git blame and a venerable piece of code that has been laying around since 1970.

Given two files, the diff command will return the line edits you have to perform to one of them, to turn it into the other. For example: given the files a.txt and b.txt

$ cat a.txt

$ cat b.txt

The diff of both files would be:

$ diff a.txt b.txt
> bb
> bbb
< aaaaa
< aa

As you probably already know, this means:

  • 0a1,2: add at line 0 (at the beginning) of file a.txt, the lines 1 and 2 from file b.txt.
  • 2,3c4: substitute lines 2 and 3 of file a.txt with line number 4 of file b.txt.

The current version of diff is highly optimized, but, at its core, it can be easily understood as some hashing (turning lines into equivalent characters), and an LCS solver. You will find more details about this in the slides from the talk.

Tracking lines across file revisions

At the core of the git blame algorithm is the problem of tracking lines across file revisions; knowing at which revision each particular line was added or modified.

This problem is usually solved by creating a graph where:

  • vertexes represent each line of a file, for all revisions of the file
  • edges represent the same line across different revisions of the file

In the slides from the talk you can find examples of forward and backward versions of a graph traversal algorithm to solve the blaming problem, from the 2006 paper Mining Version Archives for Co-changed Lines(Zimmermann et al.).

*We would like to thank ShuttleCloud for hosting and organizing the event (and for the beers!).

This post was written by Alberto Cortés.