Re-hardlink directories created using rsync --link-dest.

rsync --link-dest

This is a nice way to make backups which are trivial to restore (cp -a) and use little disk space. Low-tech enough that I trust it when I often have new files rather than modified large files.

The problem

Sometimes, because of a user mistake or a bug, you get a new full backup happening. As a result, a lot of disk space is suddenly allocated for a backup where little changed. This program is a shot at curing this situation after it happened.

The algorithm

Let's take the following simple backup content:

/some/place/2017-01-01/home/foo/bar
/some/place/2017-01-01/home/foo/boo
/some/place/2017-01-02/home/foo/bar
/some/place/2017-01-02/home/foo/baz
/some/place/2017-01-03/home/foo/bar
/some/place/2017-01-03/home/foo/baz
/some/place/2017-01-03/home/foo/boo

It can be seen as a sparse 2d grid where cells in each row are strongly related (they are more likely to be the same file than accross rows):

This tool takes the heads as arguments (ex: one folder per day). It resurses through the first head, and, for each non-directory path found, iterates over the heads on its right, building mapping between inodes and the heads each inode is in:

# relative_path = "/home/foo/bar"
{
  1: ["2017-01-01", "2017-01-02"],
  3: ["2017-03-03"],
}

Then, each inode is compared to the others (by picking an arbitrary file in the list), and each inode identical to a given one gets unlinked and relinked to that inode.

Optionally (enabled by default), the code then remembers the (relative) path it just analysed, so that it does not analyse it again when iterating over next head. This is very efficient, but makes memory usage proportional to the number of unique relative paths and their length. In my experience, with pypy, this is under 200MB for over 300k unique relative paths.

Then it goes to the next relative path in current head. Then it moves on to the next head to process files which did not yet exist in current head.

As a result, execution time is dominated by the number of unique relative paths instead of the total number of files. For the example above, it will read these files, in this order, this number of times:

/some/place/2017-01-01/home/foo/bar
/some/place/2017-01-03/home/foo/bar
/some/place/2017-01-01/home/foo/boo
/some/place/2017-01-03/home/foo/boo
/some/place/2017-01-02/home/foo/baz

Assuming inodes 1 and 2 are identical, and 4 and 5 are identical, optimal rsync --link-dest (so without users & code issues) would have produced:

And this script will produce:

An advantage of this method compared to other hard-linkers is that it freezes space regularly while running by unlinking all homonym duplicates at once.

A disadvantage of this method compared to other hard-linkers is that it will not link together identical files if their name differs. But neither would rsync --link-dest, so it would not be our backup too of choice.