Provided by: libsolv-doc_0.6.35-2ubuntu1_all bug

NAME

       libsolv-history - how the libsolv library came into existence

HISTORY

       This project was started in May 2007 when the zypp folks decided to switch to a database
       to speed up installation. As I am not a big fan of databases, I (mls) wondered if there
       would be really some merit of using one for solving, as package dependencies of all
       packages have to be read in anyway.

       Back in 2002, I researched that using a dictionary approach for storing dependencies can
       reduce the packages file to 1/3 of its size. Extending this idea a bit more, I decided to
       store all strings and relations as unique 32-bit numbers. This has three big advantages:

       ·   because of the unification, testing whether two strings are equal is the same as
           testing the equality of two numbers, thus very fast

       ·   much space is saved, as numbers do not take up as much space as strings the internal
           memory representation does not take more space on a 64-bit system where a pointer is
           twice the size of a 32-bit number

       Thus, the solv format was created, which stores a repository as a string dictionary, a
       relation dictionary and then all packages dependencies. Tests showed that reading and
       merging multiple solv repositories takes just some milliseconds.

   Early solver experiments
       Having a new repository format was one big step, but the other area where libzypp needed
       improvement was the solver. Libzypp’s solver was a port from the Red Carpet solver, which
       was written to update packages in an already installed system. Using it for the complete
       installation progress brought it to its limits. Also, the added extensions like support
       for weak dependencies and patches made it fragile and unpredictable.

       As I was not very pleased with the way the solver worked, I looked at other solver
       algorithms. I checked smart, yum and apt, but could not find a convincing algorithm. My
       own experiments also were not very convincing, they worked fine for some problems but
       failed miserably for other corner cases.

   Using SAT for solving
       SUSE’s hack week at the end of June 2007 turned out to be a turning point for the solver.
       Googling for solver algorithms, I stumbled over some note saying that some people are
       trying to use SAT algorithms to improve solving on Debian. Looking at the SAT entry in
       Wikipedia, it was easy to see that this indeed was the missing piece: SAT algorithms are
       well researched and there are quite some open source implementations. I decided to look at
       the minisat code, as it is one of the fastest solvers while consisting of too many lines
       of code.

       Of course, directly using minisat would not work, as a package solver does not need to
       find just one correct solution, but it also has to optimize some metrics, i.e. keep as
       many packages installed as possible. Thus, I needed to write my own solver incorporation
       the ideas and algorithms used in minisat. This wasn’t very hard, and at the end of the
       hack week the solver calculated the first right solutions.

   Selling it to libzypp
       With those encouraging results, I went to Klaus Kaempf, the system management architect at
       SUSE. We spoke about how to convince the team to make libzypp switch to the new solver.
       Fortunately, libzypp comes with a plethora of solver test cases, so we decided to make the
       solver pass most of the test cases first. Klaus wrote a "deptestomatic" implementation to
       check the test cases. Together with Stephan Kulow, who is responsible for the openSUSE
       distribution, we tweaked and extended the solver until most of the test cases looked good.

       Duncan Mac-Vicar Prett, the team lead of the YaST team, also joined development by
       creating Ruby bindings for the solver. Later, Klaus improved the bindings and ported them
       to some other languages.

   The attribute store
       The progress with the repository format and the solver attracted another hacker to the
       project: Michael Matz from the compiler team. He started with improving the repository
       parsers so that patches and content files also generate solvables. After that, he
       concentrated on storing all of the other metadata of the repositories that are not used
       for solving, like the package summaries and descriptions. At the end of October, a first
       version of this "attribute store" was checked in. Its design goals were:

       ·   space efficient storage of attributes

       ·   paging/on demand loading of data

       ·   page compression

       The first version of the attribute store used a different format for storing information,
       we later merged this format with the solv file format.

   libzypp integration
       Integration of the sat-solver into libzypp also started in October 2007 by Stefan Schubert
       and Michael Andres from the YaST team. The first versions supported both the old solver
       and the new one by using the old repository read functions and converting the old package
       data in-memory into a sat solver pool. Solvers could be switched with the environment
       variable ZYPP_SAT_SOLVER. The final decision to move to the new solver was made in January
       of 2008, first just by making the new solver the default one, later by completely throwing
       out the old solver code. This had the advantage that the internal solvable storage could
       also be done by using the solver pool, something Michael Matz already played with in a
       proof of concept implementation showing some drastic speed gains. The last traces of the
       old database code were removed in February.

AUTHOR

       Michael Schroeder <mls@suse.de>