README
author Kristian H?gsberg <krh@redhat.com>
Mon Jun 09 12:47:37 2008 -0400 (2008-06-09)
changeset 230 c1e2aed8dd07
child 374 f71695220726
permissions -rw-r--r--
Rewrite depsolver to use a series of passes over all packages.

The big change is that we follow one step of the depedency chain for
each package to resolve in each iteration, and repeat until there are
no more possible moves. In contrast the old depsolver would try to
follow the dependency chain completely for one package at a time.

This new approach is simpler and faster, and at the same time more
roboust. Instead of knowing how one newly installed package may
affect other packages (obsoleting, pulling in new packages etc), the
new algorithm just looks at the total list of requires, provides,
obsoletes and conflicts after installing new packages.
krh@58
     1
Just random notes at this point:
krh@58
     2
krh@58
     3
 - Razor is a package management system replacing rpm and yum.  Razor
krh@58
     4
   implements management of packages installed on the system,
krh@58
     5
   dependency solving, and upgrading in a small compact code base with
krh@58
     6
   minimal dependencies.
krh@58
     7
krh@58
     8
 - Key points: one file format for package sets, specified and
krh@58
     9
   implemented by razor; no point in splitting dep-solver and package
krh@58
    10
   manager, there is too much implementation overlap.  By hand coding
krh@58
    11
   the on-disk format we have a solid foundation on which to
krh@58
    12
   implementing fast dependency solving, error recovery etc.  No
krh@58
    13
   complex dependencies in package management stack.  In other words,
krh@58
    14
   package management is *simple*, no need to overdesign it.
krh@58
    15
krh@58
    16
 - Use one simple file format as the core of the system.  The razor
krh@58
    17
   package set data structure represents a set of packages and is used
krh@58
    18
   for the current installed set, the sets available from upstream
krh@58
    19
   sources, and the set currently being install during a transaction.
krh@58
    20
   Tiered systems such as rpm+yum, deb+apt etc end up duplicating and
krh@58
    21
   reimplementing common operations on packages.  Other systems
krh@58
    22
   typically pull in external database libraries to manage the package
krh@58
    23
   meta data and in some cases each layer in the stack uses a
krh@58
    24
   different database dependency to represent essentially the same
krh@58
    25
   data.
krh@58
    26
krh@58
    27
 - Using external database libraries may seem a good idea up front,
krh@58
    28
   but monotone vs git is a good example of how you need to control
krh@58
    29
   the data structure and the on-disk format to really get excellent
krh@58
    30
   performance.  The razor package set data structure is fairly
krh@58
    31
   simple; it's an read-only, mmap()'able on-disk data structure.  The
krh@58
    32
   package set has a sorted index of the packages in the set, a sorted
krh@58
    33
   index of all dependencies and a sorted, compact directory of all
krh@58
    34
   files in all the packages.  Operations such as merging package
krh@58
    35
   sets, satisfying requires from one set against another can all be
krh@58
    36
   implemented in linear time.  Razor implements the package set
krh@58
    37
   representation down to the bytes on the disk and can optimize the
krh@58
    38
   representation and access methods to make the necessary operations
krh@58
    39
   fast and reliable.
krh@58
    40
krh@58
    41
 - The set of installed packages on a system is represented by just
krh@58
    42
   one package set file.  As we install a set of new packages, we
krh@58
    43
   prepare the new package set a temporary file and once the install
krh@58
    44
   has succeeded, we copy the package set file to the real filename.
krh@58
    45
   If there is a system crash, razor will notice the temporary file on
krh@58
    46
   boot and resume the update.  
krh@58
    47
krh@58
    48
 - We want to keep the numer of dependencies down in a system critical
krh@58
    49
   component such as the package manager.  Between rpm and yum we're
krh@58
    50
   pulling in berkeley db, sqlite, expat, python.
krh@58
    51
krh@58
    52
 - During update of several packages rpm+yum leaves the system in an
krh@58
    53
   inconsistent state for long periods of time.  Power failure in this
krh@58
    54
   window causes critical system corruption.  We want to minimize the
krh@58
    55
   window.