Skip to content
Switch branches/tags
Go to file
Cannot retrieve contributors at this time


Rezip would like to be able to take a DEFLATE stream (e.g. a .gz/gzip archive, or a .zip file), and produce a (small) metadata file about how to recreate that exact file given the uncompressed data.

gzip and deflate are simple. However, it turns out, a lot of existing compressors do weird things.

Existing compressors

gzip(1)'s "lazy" optimisation works by omitting a match if it's immediately followed by a longer match, which can work well in practice, e.g. a123412f41234; when seeing the 412 following the f, you could insert a backreference, but, if you do, you'll miss the proceeding 1234 backreference.

gzip(1) has a number of odd cases:

  • Won't use the first byte of the file for a reference, so can't compress abcabc into (abc)\1, but can convert Sabcabc into S(abc)\1.
  • Sometimes skips length-3 runs. It's unclear to me exactly when it does this.
  • Frequently inserts (harmless) garbage at the end of files.
  • Doesn't maintain enough state to emit sensible guesses in some cases, so you can trick it into omitting a match for a crafted data file. My favourite is a3456789a23456a123412f4123456789, where it will omit the backreference for the 412 following the f.

In the real world, I've also seen Google Zopfli, which (intentionally) does some pretty weird things, and some Perl ecosystem compressor, which does some very weird things.

In Rust land, a number of compressors implement what gzip(1) calls the "fast" algorithm, where the lazy behaviour is not performed.


pristine-tar does this, but in a horrible, horrible way: it calls the system's gzip(1) with a whole range of flags, and works out which output has the smallest binary diff from the input file. This is ugly, and dependent on the system gzip compressor.

The aim for rezip is to have slow, but clear, code which can act as a reference compressor, then deal with weird zip files with less efficient (but still intelligent) formats.

I've tried to implement various techniques. None have really worked out so far:

  • emulate attempts to be a bug-compatible gzip 1.6 implementation. This fails as apparently our sets of bugs don't line up.
  • bestguess generates all the options for an encoding at each point, ranks them, then stores which the original encoder picked. As we're pretty good at ranking, this should result in mostly a stream of zeros, which can then be compressed. However, it currently needs updating for the lazy behaviour, as otherwise real gzip(1) files require huge metadata.