PURPOSE:

Need to compute a reasonably-sized delta between two 10GB tarballs without requiring 2TB of memory?  This will do the job in a brutal, ugly, effective manner.


BUILD:


gcc -o chunker chunker.c

Everything else is in Python or Perl.  You'll need the sqlite module for Python.


USAGE:

bigdiffer.sh amalgamates all the utilities into one callable so that you can just use it.  Be aware that it will make a mess of your disk and memory depending on how big the files are, especially if you abort it early.

QUESTIONS:

Why aren't I using a rolling hash?

I tried it.  I couldn't get consistent results.  Because of the vast tracts of nulls in tar files that have lots of little files I
was getting things like good chunk sizes for the first half of the file, and then the second half of the file as a single chunk.
So I made it use the vast tracts of nulls as break points and called it a day.  It works great on tar files.  It works ok on other things.
It's modular enough that you can easily substitute your own chunker if a different algorithm will work better for the data in question.

If you want you can try chunker-rollingaverage.c  It's not the kind of hash most people use, but it gives me the results I want with the 
data I currently have.  Note that it increases the space and memory requirements for calculation considerably, but it does a reasonable job
of reusing data segments from files instead of mostly just wholesale replacing any file that changed.  Controlcompactor.pl will go through
afterward and consolidate sequential chunk references, which gets the size back down to something reasonable.

Doesn't this use a lot more memory than it needs to?

Yep.  If you're actually using it and it's actually a problem for you, file a bug report and I'll actually fix it.  
Otherwise, it works for what I need it for.