tvl-depot/doc/manual/introduction.xml
Eelco Dolstra 2a4bac5459 * Refactoring.
* Convert tabs to spaces.
2003-11-26 11:24:13 +00:00

291 lines
11 KiB
XML

<chapter>
<title>Introduction</title>
<sect1>
<title>The problem space</title>
<para>
Nix is a system for controlling the automatic creation and distribution
of data, such as computer programs and other software artifacts. This is
a very general problem, and there are many applications that fall under
this description.
</para>
<sect2>
<title>Build management</title>
<para>
Build management tools are used to perform <emphasis>software
builds</emphasis>, that is, the construction of derived products
(<emphasis>derivates)</emphasis>) such as executable programs from
source code. A commonly used build tool is Make, which is a standard
tool on Unix systems. These tools have to deal with several issues:
<itemizedlist>
<listitem>
<para>
<emphasis>Efficiency</emphasis>. Since building large systems
can take a substantial amount of time, it is desirable that build
steps that have been performed in the past are not repeated
unnecessarily, i.e., if a new build differs from a previous build
only with respect to certain sources, then only the build steps
that (directly or indirectly) <emphasis>depend</emphasis> on
those sources should be redone.
</para>
</listitem>
<listitem>
<para>
<emphasis>Correctness</emphasis> is this context means that the
derivates produced by a build are always consistent with the
sources, that is, they are equal to what we would get if we were
to build the derivates from those sources. This requirement is
trivially met when we do a full, unconditional build, but is far
from trivial under the requirement of efficiency, since it is not
easy to determine which derivates are affected by a change to a
source.
</para>
</listitem>
<listitem>
<para>
<emphasis>Variability</emphasis> is the property that a software
system can be built in a (potentially large) number of variants.
Variation exists both in <emphasis>time</emphasis>---the
evolution of different versions of an artifact---and in
<emphasis>space</emphasis>---the artifact might have
configuration options that lead to variants that differ in the
features they support (for example, a system might be built with
or without debugging information).
</para>
<para>
Build managers historically have had good support for variation
in time (rebuilding the system in an intelligent way when sources
change is one of the primary reasons to use a build manager), but
not always for variation in space. For example,
<command>make</command> will not automatically ensure that
variant builds are properly isolated from each other (they will
in fact overwrite each other unless special precautions are
taken).
</para>
</listitem>
<listitem>
<para>
<emphasis>High-level system modelling language</emphasis>. The
language in which one describes what and how derivates are to be
produced should have sufficient abstraction facilities to make it
easy to specify the derivation of even very large systems. Also,
the language should be <emphasis>modular</emphasis> to enable
components from possible different sources to be easily combined.
</para>
</listitem>
</itemizedlist>
</para>
</sect2>
<sect2>
<title>Package management</title>
<para>
After software has been built, is must also be
<emphasis>deployed</emphasis> in the intended target environment, e.g.,
the user's workstation. Examples include the Red Hat package manager
(RPM), Microsoft's MSI, and so on. Here also we have several issues to
contend with:
<itemizedlist>
<listitem>
<para>
The <emphasis>creation</emphasis> of packages from some formal
description of what artifacts should be distributed in the
package.
</para>
</listitem>
<listitem>
<para>
The <emphasis>deployment</emphasis> of packages, that is, the
mechanism by which we get them onto the intended target
environment. This can be as simple as copying a file, but
complexity comes from the wide range of possible installation
media (such as a network install), and the scalability of the
process (if a program must be installed on a thousand systems, we
do not want to visit each system and perform some manual steps to
install the program on that system; that is, the complexity for
the system administrator should be constant, not linear).
</para>
</listitem>
</itemizedlist>
</para>
</sect2>
</sect1>
<!--######################################################################-->
<sect1>
<title>What Nix provides</title>
<para>
Here is a summary of Nix's main features:
</para>
<itemizedlist>
<listitem>
<para>
<emphasis>Reliable dependencies.</emphasis> Builds of file system
objects depend on other file system object, such as source files,
tools, and so on. We would like to ensure that a build does not
refer to any objects that have not been declared as inputs for that
build. This is important for several reasons. First, if any of the
inputs change, we need to rebuild the things that depend on them to
maintain consistency between sources and derivates. Second, when we
<emphasis>deploy</emphasis> file system objects (that is, copy them
to a different system), we want to be certain that we copy everything
that we need.
</para>
<para>
Nix ensures this by building and storing file system objects in paths
that are infeasible to predict in advance. For example, the
artifacts of a package <literal>X</literal> might be stored in
<filename>/nix/store/d58a0606ed616820de291d594602665d-X</filename>,
rather than in, say, <filename>/usr/lib</filename>. The path
component <filename>d58a...</filename> is actually a cryptographic
hash of all the inputs (i.e., sources, requisites, and build flags)
used in building <literal>X</literal>, and as such is very fragile:
any change to the inputs will change the hash. Therefore it is not
sensible to <emphasis>hard-code</emphasis> such a path into the build
scripts of a package <literal>Y</literal> that uses
<literal>X</literal> (as does happen with <quote>fixed</quote> paths
such as <filename>/usr/lib</filename>). Rather, the build script of
package <literal>Y</literal> is parameterised with the actual
location of <literal>X</literal>, which is supplied by the Nix
system.
</para>
</listitem>
<listitem>
<para>
<emphasis>Support for variability.</emphasis>
</para>
<para>
As stated above, the path name of a file system object contain a
cryptographic hash of all inputs involved in building it. A change to
any of the inputs will cause the hash to change--and by extension,
the path name. These inputs include both sources (variation in time)
and configuration options (variation in space). Therefore variants
of the same package don't clash---they can co-exist peacefully within
the same file system. So thanks to Nix's mechanism for reliably
dealing with dependencies, we obtain management of variants for free
(or, to quote Simon Peyton-Jone, it's not free, but it has already
been paid for).
</para>
</listitem>
<listitem>
<para>
<emphasis>Transparent source/binary deployment.</emphasis>
</para>
</listitem>
<listitem>
<para>
<emphasis>Easy configuration duplication.</emphasis>
</para>
</listitem>
<listitem>
<para>
<emphasis>Automatic storage management.</emphasis>
</para>
</listitem>
<listitem>
<para>
<emphasis>Atomic upgrades and rollbacks.</emphasis>
</para>
</listitem>
<listitem>
<para>
<emphasis>Support for many simultaneous configurations.</emphasis>
</para>
</listitem>
<listitem>
<para>
<emphasis>Portability.</emphasis> Nix is quite portable. Contrary
to build systems like those in, e.g., Vesta and ClearCase [sic?], it
does not rely on operating system extensions.
</para>
</listitem>
</itemizedlist>
<para>
Here is what Nix doesn't yet provide, but will:
</para>
<itemizedlist>
<listitem>
<para>
<emphasis>Build management.</emphasis> In principle it is already
possible to do build management using Fix (by writing builders that
perform appropriate build steps), but the Fix language is not yet
powerful enough to make this pleasant. The <ulink
url='http://www.cs.uu.nl/~eelco/maak/'>Maak build manager</ulink>
should be retargeted to produce Nix expressions, or alternatively,
extend Fix with Maak's semantics and concrete syntax (since Fix needs
a concrete syntax anyway). Another interesting idea is to write a
<command>make</command> implementation that uses Nix as a back-end to
support <ulink
url='http://www.research.att.com/~bs/bs_faq.html#legacy'>legacy</ulink>
build files.
</para>
</listitem>
</itemizedlist>
</sect1>
<!--######################################################################-->
<sect1>
<title>The Nix system</title>
<para>
...
</para>
<para>
Existing tools in this field generally both a underlying model (such as
the derivation graph of build tools, or the versioning scheme that
determines when two packages are <quote>compatible</quote> in a package
management system) and a formalism that allows ...
</para>
<para>
Following the principle of separation of mechanism and policy, the Nix
system separates the <emphasis>low-level aspect</emphasis> of file system
object management form the <emphasis>high-level aspect</emphasis> of the
...
</para>
</sect1>
</chapter>
<!--
local variables:
sgml-parent-document: ("book.xml" "chapter")
end:
-->