To be consistent with the system, following the introduction of the
fault management architecture, each package is named by an FMRI in the
pkg: scheme. That is, we have:
pkg://publisher/pkg_name@version
The publisher is generally expected to be a forward or reverse domain name identifying the publisher from which a package can be retrieved. Publishers which cannot be determined to be a domain name are legitimate, but optional functionality, like automatic server discovery for a particular publisher, may fail to work. In the examples that follow, we use "opensolaris.org" as a generic publisher.
Although RFC 2396 usage would suggest using the "authority" term, we instead call it the publisher name, as the role of this section of the FMRI is to identify the source for the package, which we consider a publication of one or more software components.
The pkg_name, like service names, can be split into a category,
subcategories, and a basename. This namespace might be populated
with "manifest" and other metadata endpoints, as well as the SHA-1
names of the package's included files. (Although the direct access
to properties of the svc FMRI scheme has been rarely used.)
A "group package" is a package that depends upon (minimum versions of) other packages, as well as optionally delivering files or other actions. An "incorporating package" is a group package that places forward constraints upon the versions of the packages it depends upon, which restricts the interval of valid package versions to one the author of the incorporation believes functional.
The primary design point of the package namespace is to allow multiple package producers to co-exist in a single namespace, so that images can switch between equivalent components from different producers.
At any point in the category hierarchy, a safe namespace can be created by using the forward or reverse domain name, either as a subcategory or as a comma-led prefix to a subcategory or package base name. (This scheme is similar to FMRI namespace escapes in smf(5), although we are eliminating use of stock symbol prefixes.) Generally, safe namespaces are only needed when the components delivered by a package need to be distinguished from those delivered by a package in the single namespace; reasons for distinguishing the components might include construction of a collection of components, insulated from changes in the wider system, or for legal reasons.
For instance, when example.com wishes to publish the "whiskers" package without reference to a larger namespace convention it can use any of the following examples:
pkg://opensolaris.org/.../com.example/whiskers pkg://opensolaris.org/.../com.example,whiskers pkg://opensolaris.org/.../com.example,software/whiskers
and so forth.
The top-level "site" category is reserved for use by the entity that administrates the server. Neither the organizations producing the operating system nor those providing additional software components may use the site category.
The top-level "vendor" category is reserved for use by organizations providing additional. The leading subcategory must be a domain. That is, if example.com wishes to publish the "whiskers" package in the vendor category, it would use a package name like:
pkg://opensolaris.org/vendor/example.com/whiskers
Use of the "vendor" category, and the vendor-escaped packages below, is likely only to be of interest to organizations that redistribute software from multiple source vendors.
ARC Notice
Some or all of these reservations may be eliminated or reduced when the single namespace convention reaches its final form.
The top-level "cluster", "feature", "group", and "metacluster" categories are all reserved for future use.
Packaging systems and distributions appear to have explicit categories, subcategories, and potentially larger groups; some distributions have explicit fields for these values, others use tagging or multi-valued fields, which allows them to classify certain packages multiply. For the FMRI namespace case, the system is similar to a packaging system with multiple, single-valued, category fields.
There appear to be two standard approaches to categorizing packages:
- By what primary class of thing a package delivers.
- By what area of functionality a package's contents address.
In the first approach, we get strict top-level categories like "application", "driver", "library", and "system" or "kernel", as well as potentially overlapping categories like "utility" and "tool". Use of the leading subcategory is limited, and generally given to the subsystem involved. A relatively detailed worked example of the X11 subsystem under this scheme is given in
http://mail.opensolaris.org/pipermail/pkg-discuss/2008-February/001838.html
XXX This reference is now wrong.
In the second, we would also see categories like these, but leading subcategory is much more likely to classify according to functionality, so that we would see "application/mail", "application/web", "application/compiler", and so forth. Most network packaging systems appear to classify in this fashion.
An appealing variation of the second form is to rotate all of the non-"application" packages under a "system" mega-category, such that all of the leaf packages (with the possible exception of device drivers) are exposed at the category level. Table 1 shows some example transformations under this scheme.
| FROM | TO |
|---|---|
| application/web/firefox | web/firefox |
| application/compiler/gcc4 | compiler/gcc4 |
| library/c | system/library/c |
| kernel/generic | system/kernel/generic |