# Image Format vs Fossil Repo Size
## The Problem
Fossil has a [delta compression][dc] feature which removes redundant
information from a file relative to its parent on check-in.[^delta-prgs]
That delta is then [zlib][zl]-compressed before being stored
in the Fossil repository database file.
Storing pre-compressed data files in a Fossil repository defeats both of
these space-saving measures:
1. Binary data compression algorithms turn the file data into
pseudorandom noise.[^prn]
Typical data compression algorithms are not [hash functions][hf],
where the goal is that a change to each bit in the input has a
statistically even chance of changing every bit in the output, but
because they do approach that pathological condition, pre-compressed
data tends to defeat Fossil’s delta compression algorithm, there
being so little correlation between two different outputs from the
binary data compression algorithm.
2. An ideal lossless binary data compression algorithm cannot be
applied more than once to make the data even smaller, since random
noise is incompressible. The consequence for our purposes here is
that pre-compressed data doesn’t benefit from Fossil’s zlib
compression.
You might then ask, what does it matter if the space savings comes from
the application file format (e.g. JPEG, DOCX, Zip, etc.) or from Fossil
itself? It really doesn’t, as far as point 2 above goes, but point 1
causes the Fossil repository to balloon out of proportion to the size of
the input data change on each checkin. This article will illustrate that
problem, quantify it, and give a solution to it.
[dc]: ./delta_format.wiki
[hf]: https://en.wikipedia.org/wiki/Hash_function
[zl]: http://www.zlib.net/
## <a id="formats"></a>Affected File Formats
In this article’s core experiment, we use 2D image file formats, but
this article’s advice also applies to many other file types. For just a
few examples out of what must be thousands:
* **Microsoft Office**: The [OOXML document format][oox] used from
Office 2003 onward (`.docx`, `.xlsx`, `.pptx`, etc.) are Zip files
containing an XML document file and several collateral files.
* **Libre Office**: For the purposes of this article, its
[OpenDocument Format][odf] is designed the same basic way as OOXML.
* **Java**: A Java [`.jar` file][jcl] is a Zip file containing JVM
`.class` files, manifest files, and more.
* **Windows Installer:** An [`*.msi` file][wi] is a proprietary
database format that contains, among other things, [Microsoft
Cabinet][cab]-compressed files, which in turn may hold Windows
executables, which [may themselves be compressed][exc].
* **SVG, PDF, TIFF, etc.**: Many file formats are available in both
compressed and uncompressed forms. You should use the uncompressed
form with Fossil wherever practical, as we will show below.
[cab]: https://en.wikipedia.org/wiki/Cabinet_(file_format)
[exc]: https://en.wikipedia.org/wiki/Executable_compression
[jcl]: https://en.wikipedia.org/wiki/Java_(programming_language)
[odf]: https://en.wikipedia.org/wiki/OpenDocument
[oox]: https://en.wikipedia.org/wiki/Office_Open_XML
[wi]: https://en.wikipedia.org/wiki/Windows_Installer
## <a id="demo"></a>Demonstration
The companion `image-format-vs-repo-size.ipynb` file ([download][nbd],
[preview][nbp]) is a [JupyterLab][jl] notebook implementing the following
experiment:
1. Create a new minimum-size Fossil repository. Save this initial size.
2. Use [ImageMagick][im] via [Wand][wp] to generate a JPEG file of a
particular size — currently 256 px² — filled with Gaussian noise to
make data compression more difficult than with a solid-color image.
3. Check that image into the new Fossil repo, and remember that size.
4. Change a random pixel in the image to a random RGB value, save that
image, check it in, and remember the new Fossil repo size.
5. Iterate on step 4 some number of times — currently 10 — and remember
the Fossil repo size at each step.
6. Repeat the above steps for BMP, PNG, and TIFF.[^tiff-cmp]
7. Create a bar chart showing how the Fossil repository size changes
with each checkin.
We chose to use JupyterLab for this because it makes it easy for you to
modify the notebook to try different things. Want to see how the
results change with a different image size? Easy, change the `size`
value in the second cell of the notebook. Want to try more image
formats? You can put anything ImageMagick can recognize into the
`formats` list. Want to find the break-even point for images like those
in your own repository? Easily done with a small amount of code.
[im]: https://www.imagemagick.org/
[jl]: https://jupyter.org/
[nbd]: ./image-format-vs-repo-size.ipynb
[nbp]: https://nbviewer.jupyter.org/urls/fossil-scm.org/fossil/doc/trunk/www/image-format-vs-repo-size.ipynb
[wp]: http://wand-py.org/
## <a id="results"></a>Results
Running the notebook gives a bar chart something like[^variance] this:
![results bar chart](./image-format-vs-repo-size.svg)
There are a few key things we want to draw your attention to in that
chart:
* BMP and uncompressed TIFF are nearly identical in size for all
checkins, and the repository growth rate is negligible past the
first commit.[^size-jump] We owe this economy to Fossil’s delta compression
feature: it is encoding each of those single-pixel changes in a very
small amount of repository space.
* The JPEG and PNG bars increase by large amounts on most checkins
even though each checkin *also* encodes only a *single-pixel change*.
* The size of the first checkin in the BMP and TIFF cases is roughly
the same as that for the PNG case, because both PNG and Fossil use
the zlib binary data compression algorithm. This shows that for
repos where the image files are committed only once, there is
virtually no penalty to using BMP or TIFF over PNG. The file sizes
likely differ only because of differences in zlib settings between
the cases.
* Because JPEG’s lossy nature allows it to start smaller and have
smaller size increases than PNG, the crossover point with
BMP/TIFF isn’t until 7-9 checkins in typical runs of this [Monte
Carlo experiment][mce]. Given a choice among these four file
formats and a willingness to use lossy image compression, a rational
tradeoff is to choose JPEG for repositories where each image will
change fewer than that number of times.
[mce]: https://en.wikipedia.org/wiki/Monte_Carlo_method
## <a id="makefile"></a>Automated Recompression
Since programs that produce and consume binary-compressed data files
often make it either difficult or impossible to work with the
uncompressed form, we want an automated method for producing the
uncompressed form to make Fossil happy while still having the compressed
form to keep our content creation applications happy. This `Makefile`
should[^makefile] do that for BMP, PNG, SVG, and XLSX files:
.SUFFIXES: .bmp .png .svg .svgz
.svgz.svg:
gzip -dc < $< > $@
.svg.svgz:
gzip -9c < $< > $@
.bmp.png:
convert -quality 95 $< $@
.png.bmp:
convert $< $@
SS_FILES := $(wildcard spreadsheet/*)
all: $(SS_FILES) illus.svg image.bmp doc-big.pdf
reconstitute: illus.svgz image.png
( cd spreadsheet ; zip -9 ../spreadsheet.xlsx) * )
qpdf doc-big.pdf doc-small.pdf
$(SS_FILES): spreadsheet.xlsx
unzip $@ -d $<
doc-big.pdf: doc-small.pdf
qpdf --stream-data=uncompress $@ $<
This `Makefile` allows you to treat the compressed version as the
process input, but to actually check in only the changes against the
uncompressed version by typing “`make`” before “`fossil ci`”. This is
not actually an extra step in practice, since if you’ve got a
`Makefile`-based project, you should be building — and testing — it
before checking each change in anyway!
Because this technique is based on dependency rules, only the necessary
files are generated on each `make` command.
You only have to run “`make reconstitute`” *once* after opening a fresh
Fossil checkout to produce those compressed sources. After that, you
work with the compressed files in your content creation programs. Your
build system might include some kind of bootstrapping or
auto-configuration step that you could attach this to, so that it
doesn’t need to be run by hand.
This `Makefile` illustrates two primary strategies:
### Input and Output File Formats Differ by Extension
In the case of SVG and the bitmap image formats, the file name extension
differs between the cases, so we can use `make` suffix rules to get the
behavior we want. The top half of the `Makefile` just tells `make` how
to map from `*.svg` to `*.svgz` and vice versa, and the same for `*.bmp`
to/from `*.png`.
### Input and Output Use the Same Extension
We don’t have that luxury for Excel and PDF files, each for a different
reason:
* **Excel:** Excel has no way to work with the unpacked Zip file
contents at all, so we have to unpack it into a subdirectory, which
is what we check into Fossil. On making a fresh Fossil checkout, we
have to pack that subdirectory’s contents back up into an `*.xlsx`
file with “`make reconstitute`” so we can edit it with Excel again.
* **PDF:** All PDF readers can display an uncompressed PDF file, but
many PDF-*producing* programs have no option for uncompressed
output. Since the file name extension is the same either way, we
treat the compressed PDF as the source to the process, yielding an
automatically-uncompressed PDF for the benefit of Fossil. Unlike
with the Excel case, there is no simple “file base name to directory
name” mapping, so we just created the `-big` to `-small` name scheme
here.
[^delta-prgs]:
This problem is not Fossil-specific. Several other programs also do
delta compression, so they’ll also be affected by this problem:
[rsync][rs], [Unison][us], [Git][git], etc. You should take this
article’s advice when using all such programs, not just Fossil.
When using file copying and synchronization programs *without* delta
compression, on the other hand, it’s best to use the most
highly-compressed file format you can tolerate, since they copy the
whole file any time any bit of it changes.
[^prn]:
In fact, a good way to gauge the effectiveness of a given
compression scheme is to run its output through the same sort of
tests we use to gauge how “random” a given [PRNG][prng] is. Another
way to look at it is that if there is a discernible pattern in the
output of a compression scheme, that constitutes *information* (in
[the technical sense of that word][ith]) that could be further
compressed.
[^tiff-cmp]:
We're using *uncompressed* TIFF here, not [LZW][lzw]- or
Zip-compressed TIFF, either of which would give similar results to
PNG, which is always zlib-compressed.
[^variance]:
The raw data changes somewhat from one run to the next due to the
use of random noise in the image to make the zlib/PNG compression
more difficult, and the random pixel changes. Those test design
choices make this a [Monte Carlo experiment][mce]. We’ve found that
the overall character of the results doesn’t change from one run to
the next.
[^size-jump]:
It’s not clear to me why there is a one-time jump in size for BMP
and TIFF past the first commit. I suspect it is due to the SQLite
indices being initialized for the first time.
Page size inflation might have something to do with it as well,
though we tried to control that by rebuilding the initial DB with a
minimal page size. If you re-run the program often enough, you will
sometimes see the BMP or TIFF bar jump higher than the other, again
likely due to one of the repos crossing a page boundary.
Another curious artifact in the data is that the BMP is slightly
larger than for the TIFF. This goes against expectation because a
low-tech format like BMP should have a small edge in this test
because TIFF metadata includes the option for multiple timestamps,
UUIDs, etc., which bloat the checkin size by creating many small
deltas.
[^makefile]:
The `Makefile` above is not battle-tested. Please report bugs and
needed extensions [on the forum][for].
[for]: https://fossil-scm.org/forum/forumpost/15e677f2c8
[git]: https://git-scm.com/
[ith]: https://en.wikipedia.org/wiki/Information_theory
[lzw]: https://en.wikipedia.org/wiki/Lempel%E2%80%93Ziv%E2%80%93Welch
[prng]: https://en.wikipedia.org/wiki/Pseudorandom_number_generator
[rs]: https://rsync.samba.org/
[us]: http://www.cis.upenn.edu/~bcpierce/unison/