When I was drowning in the sea of new information at NIPS 2016, colleagues pointed me at the excellent blog post by Erik Bernhardsson: The half-life of code & the ship of Theseus. It was very inspiring; Erik actually implemented something which was in my plans, too. The idea is simple: see how many lines in the source code remain unchanged over time. If you’ve not read that article yet, please do - I am moving further.


The problem with the tool Erik used, git-of-theseus, is that all the mainstream git libraries are not suitable for efficient analysis we want. Particularly, we need git blame information at every consecutive revision, and the only way to obtain it using those libraries is to apparently execute git blame at every revision. Since git blame takes the linear time O(N), N is the number of commits in the history, running it on all the history takes the quadratic time O(N2). Alternatively, we could cache the results from the previous git blame to reuse them during the next run, effectively reducing the complexity back to the linear time.

Why speed matters to us? We would like to conduct that sort of analysis on every repository in the world. Our data processing pipeline leverages src-d/go-git, our own Git client (and soon server) implementation. We constantly harness the fruits from maintaining the special, fine-tailored Git library in the production, and creating src-d/hercules is just an example.

There is another problem with the tool related to the history itself. How do you define the linear history in a Git repository? The taken approach seems valid at first: we pick HEAD, go back to the parent, then go back again and again until we reach the root commit. Sssh, here comes Mr. Merge:

      A---B---C topic
     /         \
D---E---F---G---H---I master
     ----------> time
 way <----------

Which direction shall we continue traversing the commit graph at H, to C or to G? What about more complex cases, “polymerges” which are N>2 branches converging at the single point? Mr. Merge laughs out loud and cries, “There is no right answer!”. “But wait, what about all the past, don’t we know it?”, you may ask. “Let’s pick all the commits belonging to master and sort them in time!” You cannot. A branch in git is only a pointer to some commit. As time goes by, pointers move and we lose the previous states completely. That is, according to how git works, there is no “master” branch in the past, there is only master which exists at this very moment in the present. In other words, git commits do not have branch associations except branch HEAD-s. Luckily, we merge secondary branches to the major branches and not vice versa in most sane git workflows, so the merge commit’s first parent is usually what we want. At the same time, execute git help rev-list and surprise how many ways there are to extract the linear history.

Mr. Merge is not willing to leave us yet and points his dirty finger at git diff, grinning. Indeed, git diffon a merge commit is not going to show us the expected changes. Instead, it exclusively demonstrates how we solved the conflicts during the merge; in the case of no conflicts, it is empty at all. Now the funny thing, imagine the following situation: commit B changes line L and commit G changes the same line L exactly the same way. The merge will treat the lines equal and exclude them from the conflicts. Who is the L’s author? When was it changed? Mr. Merge laughs again and repeats his dreadful words, “There is no right answer!”

If you are interested in git diff and git blame internals, please read the awesome presentation by Alberto Cortés, source{d}’s employee who implemented those functions in go-git.

For now, since there is no correct solution to the history problem, the only way to deal with it is to support taking any commit sequence as input. As for the diff problem, I treat secondary branches as if they appear at the merge time. This is far from ideal but at least better than ignoring them completely.

incremental blame

Here “incremental blame” is different from git blame --incremental. The former reduces the complexity of doing many sequential blames. The latter streams the output in a machine readable format.

Let me remind you of red-black trees. RB tree is the algorithm to maintain a balanced binary tree so that the height variance is lower than 2x. Let’s store our changed line intervals in an RB tree with line numbers as keys:

Random go-git blame.
The corresponding interval tree.

As can be seen, each line interval corresponding to the same commit is encoded by two nodes, the beginning and the end, and the end node is in turn the beginning of the next interval.

We start from the first element in the commit sequence aka “history root” and iterate. We create the separate tree for every new file. Given the diff for every commit, we maintain the RB tree structure efficiently. Every atomic edit in a file is either an insertion or deletion of lines. Let’s consider them separately.


Suppose that we insert 2 new lines at position 314 which is in the middle of 635c77e0 (brown, Alberto Cortés).

  1. Insert new node A into the tree at 314 pointing to our commit. This leads to rebalancing the tree in O(log(N)) at worst.
  2. Insert new node B into the tree at 314 + 2 = 316 pointing to 635c77e0. Rebalancing the tree again.
  3. For all the nodes greater than B, we add the same delta +2 to the key. No need to rebalance the tree.The complexity is O(N).


Suppose that we delete 2 lines at position 314. For all the nodes greater than 635c77e0 as 312, we add the same delta -2 to the key. No need to rebalance the tree. The complexity is O(N).

Thus we’ve got the linear complexity for both operations. Still, let’s look at a typical commit on GitHub:


Insertions and deletions are often coupled, and often their lengths are equal, thus deltas neutralize and we do not have to update the subsequent keys at all, getting O(log(N)), which is really O(1) amortized.

An alternative implementation would be using single-linked lists, where each node carries the corresponding line interval length:

Single-linked list, folded into a circle to fit.

That structure makes insertions and deletions constant time but seeking for the right line number is always linear. In other words, disregarding our luck, we always spend O(N) on updating a file.

The only way to choose the best data structure is to conduct real world experiments, however, the blame performance itself is far from being the bottleneck at this moment.


Our blame algorithm requires to have the diff on each of the changed files. The funny thing is that extracting diffs at git “porcelain” level of abstraction is impossible. Git exposes each commit as a snapshot of the repository, not a difference between adjacent revisions (refer to the book). Thus the only way to obtain a commit’s diff is by fair diff-ing the corresponding files, and this is what git blame does repetitively. Internally, git may store deltas between blobs to optimize the space, but it is hard to use them and has several limitations.

Diff-ing the file trees is another problem. Suppose that you’ve got two snapshots of a deep directory with millions of files each. How to you decide which files were changed, deleted or added? Without breaking the git abstraction, the only way left is to do a linear scan over the whole tree which may take ages to complete.

Internally, git assigns hashes to each directory based on it’s contents. We could eliminate traversing the majority of unchanged directories by comparing the corresponding hashes. Again, “leaky abstraction” diff-tree algorithm is not included into go-git, but Alberto will finish it soon. It should speedup Hercules, too.

The last problem is how we work with the repository. Cgit, libgit2 and the rest are designed to work with the file system in the first place, they cache everything they can. go-git has backends, including the file system and fully in-memory storage. If we load a git repository completely in memory, all operations will fly, and it is the key point why Hercules still works pretty fast. If we use the current file system backend which does not cache anything and we’ll get 100x slowdown. Analysing git/git in memory takes about 9 gigs, and I am scared to play that trick with the Linux kernel.


The overall performance of Hercules compared to Theseus is about 6x better on small repos and 20% better on large ones (like git/git). As for now, it is mostly usable with moderately sized repos. The plots look pretty much the same as they should.

The λ metric is going to be taken into account in one of our repositories clustering schemes. Most likely, we will create several perks on top of it, like clustering directories in a repository based on the local λ value or finding weakly designed modules. Who knows?

This post was written by Vadim Markovtsev. Follow him on Twitter: @vadimlearning.