* Better introduction.

* Set notes in a different color than warnings.
This commit is contained in:
Eelco Dolstra 2004-10-13 15:08:35 +00:00
parent 1317242780
commit 2b20701f78
3 changed files with 122 additions and 119 deletions

View File

@ -5,51 +5,123 @@ has grown to 5, with more expected.</quote></para></epigraph>
<para>Nix is a system for the deployment of software. Software
deployment is concerned with the creation, distribution, and
management of software components (<quote>packages</quote>). There
are many tools for this, but they tend to ignore some important
requirements for deployment:
management of software components (<quote>packages</quote>). Its main
features are:
<itemizedlist>
<listitem><para><emphasis>Correctness</emphasis>. The basic goal of
software deployment is to transfer software from one machine (e.g.,
the developer's, where it presumably works) to another machine (e.g.,
the end user's). The software should work exactly the same on the
target machine as on the source machine. But this in practice turns
out to be rather difficult due to <emphasis>dependencies between
components</emphasis> and <emphasis>interference between
components</emphasis>. If we deploy a component that depends on other
components, then we should deploy those dependencies as well. If they
are missing on the target system, the component probably won't work.
If they <emphasis>are</emphasis> present but are not the right
version, the component might not work. And if even if they are the
right version, they may have been built with different flags or
options, which can cause incompatibilities. Interference occurs when
components <quote>collide</quote> with each other in the file system.
For instance, different versions of the same package tend to overwrite
each other, so they cannot be installed at the same time. But always
picking the latest version might break components that only work with
some older version.</para></listitem>
<listitem><para>It makes sure that dependency specifications are
complete. In general in a deployment system you have to specify for
each component what its dependencies are, but there are no guarantees
that this specification is complete. If you forget a dependency, then
the component will build and work correctly on
<emphasis>your</emphasis> machine if you have the dependency
installed, but not on the end user's machine if it's not
there.</para></listitem>
<listitem><para><emphasis>Variability</emphasis>. Many package
management tools have difficulty supporting the installation of
multiple versions or variants of the same component. This is bad
because as ...</para></listitem>
<listitem><para>It is possible to have <emphasis>multiple versions or
variants</emphasis> of a component installed at the same time. In
contrast, in systems such as RPM different versions of the same
package tend to install to the same location in the file system, so
you installing one version will remove the other. This is especially
important if you want to have use applications that have conflicting
requirements on different versions of a component (e.g., application A
requires version 1.0 of library X, while application B requires a
non-backwards compatible version 1.1).</para></listitem>
<listitem><para>Users can have different <quote>views</quote>
(<quote>profiles</quote> in Nix parlance) on the set of installed
applications in a system. For instance, one user can have version 1.0
of some package visible, while another is using version 1.1, and a
third doesn't use it at all.</para></listitem>
<listitem><para>It is possible to atomically
<emphasis>upgrade</emphasis> software. I.e., there is no time window
during an upgrade in which part of the old version and part of the new
version are simultaneously visible (which might well cause the
component to fail).</para></listitem>
<listitem><para>Likewise, it is possible to atomically roll-back after
an install, upgrade, or uninstall action. That is, in a fast (O(1))
operation the previous configuration of the system will be
restored. This is because upgrade or uninstall actions doesn't
actually remove components from the system.</para></listitem>
<listitem><para>Unused components can be
<emphasis>garbage-collected</emphasis> automatically and safely.
I.e., when you remove an application from a profile, its dependencies
will be deleted by the garbage collector if there are no other active
applications that are using it.</para></listitem>
<listitem><para>Nix supports both source-based deployment models
(where you distribute <emphasis>Nix expressions</emphasis> that tell
Nix how to build software from source) and binary-based deployment
models. The latter is more-or-less transparent: installation of
components is always based on Nix expressions, but if those
expressions have been built before and Nix knows that the resulting
binaries are available somewhere, it will use those
instead.</para></listitem>
<listitem><para>Nix is flexible in the deployment policies that it
supports. There is a clear separation between the tools that
implement basic Nix <emphasis>mechanisms</emphasis> (e.g., building
Nix expressions), and the tools that implement various deployment
<emphasis>policies</emphasis>. For instance, there is a concept of
<quote>Nix channels</quote> that can be used to keep software
installations up-to-date automatically from a network source. This is
a policy that is implemented by a fairly short Perl script, which can
be adapted easily to achieve similar policies.</para></listitem>
<listitem><para>Nix component builds aim to be <quote>pure</quote>;
that is, unaffected by anything other than the declared dependencies.
This means that if a component was built succesfully once, it can be
rebuilt again on another machine and the result will be the same. We
cannot <emphasis>guarantee</emphasis> this (e.g., if the build depends
on the time-of-day), but Nix (and the tools in the Nix Packages
collection) takes special measures to help achieve
this.</para></listitem>
<listitem><para>Nix expressions (the things that tell Nix how to build
components) are self-contained: they describe not just components but
complete compositions. In other words, Nix expressions also describe
how to build all the dependencies. This is contrast to component
specification languages like RPM spec files, which might say that a
component X depends on some other component Y, but since it does not
describe <emphasis>exactly</emphasis> what Y is, the result of
building or running X might be different on different machines.
Combined with purity, self-containedness ensures that a component that
<quote>works</quote> on one machine also works on another, when
deployed using Nix.</para></listitem>
<listitem><para>The Nix expression language makes it easy to describe
variability in components (e.g., optional features or
dependencies).</para></listitem>
<listitem><para>Nix is ideal for building build farms that do
continuous builds of software from a version management system, since
it can take care of building all the dependencies as well. Also, Nix
only rebuilds components that have changed, so there are no
unnecessary builds. In addition, Nix can transparently distribute
build jobs over different machines, including different
platforms.</para></listitem>
<listitem><para>Nix can be used not only for software deployment, but
also for <emphasis>service deployment</emphasis>, such as the
deployment of a complete web server with all its configuration files,
static pages, software dependencies, and so on. Nix's advantages for
software deployment also apply here, for instance, the ability
trivially to have multiple configurations at the same time, or the
ability to do roll-backs.</para></listitem>
</itemizedlist>
</para>
<para>Here are some of Nix's main features:
<itemizedlist>
<listitem><para>Nix can quite reliably figure out the dependencies
between components.</para></listitem>
</itemizedlist>
</para>
<para>This manual tells you how to install and use Nix and how to
write Nix expressions for software not already in the Nix Packages
collection. It also discusses some advanced topics, such as setting
up a Nix-based build farm, and doing service deployment using
Nix.</para>
<warning><para>This manual is a work in progress. It's quite likely
to be incomplete, inconsistent with the current implementation, or
@ -62,85 +134,10 @@ a Memory Management Discipline on Software
Deployment</citetitle></ulink> discusses the hashing mechanism used to
ensure reliable dependency identification and non-interference between
different versions and variants of packages. The LISA 2004 paper
<citetitle>Nix: A Safe and Policy-Free System for Software
Deployment</citetitle> gives a more general discussion of Nix from a
system-administration perspective.</para></note>
<para>
Nix solves some large problems that exist in most current deployment and
package management systems. <emphasis>Dependency determination</emphasis>
is a big one: the correct installation of a software component requires
that all dependencies of that component (i.e., other components used by it)
are also installed. Most systems have no way to verify that the specified
dependencies of a component are actually sufficient.
</para>
<para>
Another big problem is the lack of support for concurrent availability of
multiple <emphasis>variants</emphasis> of a component. It must be possible
to have several versions of a component installed at the same time, or
several instances of the same version built with different parameters.
Unfortunately, components are in general not properly isolated from each
other. For instance, upgrading a component that is a dependency for some
other component might break the latter.
</para>
<para>
Nix solves these problems by building and storing packages 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>
<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.
</para>
<para>
Other features:
</para>
<para>
<emphasis>Transparent source/binary deployment.</emphasis>
</para>
<para>
<emphasis>Unambiguous identification of configuration.</emphasis>
</para>
<para>
<emphasis>Automatic storage management.</emphasis>
</para>
<para>
<emphasis>Atomic upgrades and rollbacks.</emphasis>
</para>
<para>
<emphasis>Support for many simultaneous configurations.</emphasis>
</para>
<para>
<emphasis>Portability.</emphasis> Nix is quite portable. Contrary to
build systems like those in, e.g., Vesta and ClearCase, it does not rely on
operating system extensions.
</para>
<ulink
url='http://www.cs.uu.nl/~eelco/pubs/nspfssd-lisa2004-final.pdf'><citetitle>Nix:
A Safe and Policy-Free System for Software
Deployment</citetitle></ulink> gives a more general discussion of Nix
from a system-administration perspective.</para></note>
</chapter>

View File

@ -26,7 +26,7 @@
]>
<book>
<title>Nix: A System for Software Deployment</title>
<title>Nix User's Guide</title>
<subtitle>Draft (Version &version;)</subtitle>

View File

@ -112,10 +112,16 @@ div.note,div.warning
font-style: italic;
}
div.note h3,div.warning h3
div.warning h3
{
color: red;
text-decoration: underline;
font-size: 100%;
}
div.note h3
{
color: blue;
font-size: 100%;
}
div.navfooter *