package docteur-unix

  1. Overview
  2. Docs
A simple read-only Key/Value from Git to MirageOS


Dune Dependency






An opiniated file-system for MirageOS

Published: 25 Mar 2022


Docteur - the simple way to load your Git repository into your unikernel

docteur is a little program which wants to provide an easy way to integrate a "file-system" into an unikernel. docteur provides a simple binary which make an image disk from a Git repository. Then, the user is able to "plug" this image into an unikernel as a read-only "file-system".


The distribution comes with a simple unikernel which show the given file from the given image disk. The example requires KVM.

$ git clone
$ cd docteur
$ opam pin add -y .
$ cd unikernel
$ docteur.make -b refs/heads/main disk.img
$ mirage configure -t hvt --disk docteur
$ make depends
$ mirage build
$ solo5-hvt --block:docteur=disk.img simple.hvt --filename /

NOTE: For mirage -t unix, the disk name is the filename:

$ mirage configure -t unix --disk disk.img
$ mirage build
$ make depends
$ ./simple --filename /

An image can be checked by docteur with docteur.verify`:

$ docteur.verify disk.img
commit	: 57d227d8f4808076646de35acf26dee885f2555b
author	: "Calascibetta Romain" <>
root	: 5886893922d57c1ff4871d9a6b7b2cfa48b9e9a6

Merge pull request #22 from dinosaure/without-c

Remove C code to be compatible with MirageOS

By this way, you can check the version of your snapshot and if the given disk.img is well formed for a MirageOS.

Docteur is able to save a remote Git repository, a local Git repository or a simple directory:

$ docteur.make disk.img
$ docteur.make disk.img
$ docteur.make disk.img
$ docteur.make git:// disk.img
$ docteur.make file://$(pwd)/ disk.img 
  ; assume that $(pwd) is a local Git repository
  ; $(pwd)/.git exists
$ docteur.make file://$(pwd)/ disk.img
  ; or it's a simple directory

NOTE: The last example can be less efficient (about compression) than others because we directly use our own way to generate a PACK file (which is less smart than git).

Docteur as a file-system

MirageOS does not have a file-system at the beginning. So we must implement one to get the idea of files and directories. Multiple designs exist and no one are perfect for any cases.

However, docteur exists as one possible "file-system" for MirageOS. It's not the only one but it deserves a special case. Indeed, you can look into irmin and ocaml-git for an other one.

Docteur provides only a read-only file-system and contents are not a part of the unikernel. Only meta-data are in the unikernel. Let me explain a bit the format.

The PACK file

In your Git repositories, most of your Git objects (files, directories, commits) are stored into a PACK file. It's an highly compressed representation of your Git repository (your history, your files, etc.). Indeed, the PACK file has 2 levels of compression:

  1. a zlib compression for each objects

  2. a compression between objects with a binary diff (libXdiff)

For example, 14 Go of contents (like a documentation) can fit into a PACK file of 280 Mo! It's mostly due to the fact that a documentation, for example, has several files which are pretty the same. According to the second level of the compression, we can store few objects as bases and compress the rest of the documentation with them.

So, docteur uses the same format as an image disk. Then, it re-uses the IDX file associated to the PACK file. By this way, we permit as fast access to the content.

Finally, contents of objects (files or directories) and where they are from their hashes into the PACK file are statically produced by docteur.make:

$ docteur.make <repository> [-b <refs>] <image>
$ docteur.make -b refs/heads/main disk.img

However, the indexation of objects is done by their hashes. It's not done by their locations in your system. Such information is calculated by the unikernel itself. At the beginning, it analyzes the PACK file and the IDX file to reconstruct the system's layout with filenames and directory names.

So, the more files there are, the longer this operation can take - and the more memory you use. Indeed, the system's layout is stored into memory with the art data-structure. Even if such data-structure is faster and smaller than an usual radix tree, if you take the example of a huge documentation, the unikernel needs ~650 Mo in memory.

docteur wants to solve 2 issues:

  • How to access to a huge file-system into an unikernel We can from a block-device (an external ressource of the unikernel)

  • How to fastly load a file We use a fast data-structure in-memory to get contents with art

Of course, in many ways, such layout can not fit in many cases. If you have multiple and small files, it's probably not the best solution. At least, it's one solution in the MirageOS eco-system!

Dependencies (12)

  1. mirage-kv >= "3.0.0" & < "5.0.0"
  2. lwt >= "5.4.0" & < "5.6.0"
  3. logs >= "0.7.0"
  4. git >= "3.7.0"
  5. digestif >= "1.0.0"
  6. carton >= "0.4.1"
  7. bigstringaf >= "0.7.0"
  8. art >= "0.1.1"
  9. mirage-unix >= "5.0.0"
  10. docteur = version
  11. dune >= "2.8.0"
  12. ocaml >= "4.07.0"

Dev Dependencies


Used by





Innovation. Community. Security.