Difference between revisions of "Efforts/Package Restructuring/Modelling"

Revision as of 12:29, 24 January 2008

Some Terms

Feature

Something usable for a target audience (e.g. user, administrator).

Module

Container for (source) code having similar semantics, see OOo Modules.

Part

May be a deliverable of a module. One or multiple parts compose a feature by completing each other.

Product

A set of features combined with a particular brand, basically abstract only.

Product Configuration

A concrete instance of a product, e.g. OOo for Linux.

Deliverable

One or multiple files to be delivered to an audience (users, developers), e.g. an ISO image or download installation set.

De-Composition

Looking at an installed OOo and its files, registry entries etc., we can see, that all these entities belong to one or multiple of the following categories, such that they provide or depend on it.

Brand

This is everything somehow related to a specific product, such as StarOffice or OpenOffice.org.

Operating System

This is everything somehow related to a specific operating system as Solaris, MS Windows or Linux.

Machine Architecture

This is everything related to a native (machine) interpreter, such as x86 or SPARC.

Localization

This is everything related to a human language, such as English or German.

Version

This is everything related to a particular version.

The consequences of this observation are, that entities unrelated to one or multiple categories, but shared by products differing in these categories, are identical and thus may be re-used for creating, changing or updating products respectively installed products.

Additionally every entity exactly belongs to one

• feature

as otherwise it would not be needed for any feature, thus it would not be needed at all. Certainly features may be related such that they require one another, even if this is not noticeable in the Product View.

Every product we define can be assembled out of parts (e.g. RPM packages), which are created along the above categories.

One Feature in Two Languages

In the simplest case, we have exactly one part for every product, e.g. for the Writer:

• The writer part.

This part may include all entities for the Writer for Linux x86, in English, applying the OpenOffice.org brand.

Later on, we would like to provide the Writer not only in English, but may be also in German, which leads to the following parts:

• writer_en - The Writer in English.
• writer_de - The Writer in German.

Obviously we now need to take care of any redundancy in these parts. We can do so by

• naming / placing this files differently, or
• by sharing them.

For obvious reasons, that sharing approach is better, leading to the following parts:

• writer - This is everything but the locale.
• writer_en - This is the English locale only.
• writer_en - This is the German locale only.

But wait, no we have another problem. What if the user only installs one of the parts? This would be an incomplete / inconsistent and useless installation.

On UNIX and alike platforms, parts are typically called packages. Package managers typically use dependencies and virtual packages (or provisions) to address the consistency problem. So, lets try to set the above parts into relationship using dependencies and provisions.

The writer ("all but the locale") part needs at least one writer locale part to be installed, to function correctly, but must on depend on any concrete locale. So we may want to model the dependencies and provisions as follows:

• writer depends on writer_lcl, provides writer
• writer_en depends on writer, provides writer_en, writer_lcl
• writer_de depends on writer, provides writer_de, writer_lcl

By now, we can only install any of the above parts, if we install a completing part as well, though we certainly can install all three parts.

Two Features/Two Layers

In the previous example, our only feature was the Writer. Lets now have two features, e.g.

• Writer, and
• Calc.

As you may know, OOo has a layered Architecture, therefore thus providing these two features as two parts:

• writer - The Writer product.
• calc - The Calc product.

leads to many same files to be installed twice. So, obviously we need again to take care of this redundancy. And again, we can do so by

• naming / placing this files differently, or
• by sharing them.

For obvious reasons, that sharing approach is better, leading to the following parts:

• writer - Everything needed for Writer, not being needed for Calc.
• calc - Everything needed for Calc, not being needed for Writer.
• basis - Everything needed for both, Calc and Writer.

Again, we now need to take care of completeness / consistency etc.

So, lets define some dependencies:

• writer - depends on basis, provides writer
• calc - depends on basis, provides calc
• basis - depends on nothing, provides basis

Now, the writer' part respectively the calc part can only if the basis part is going to be installed as well.

By now, we can only install a feature (Writer or Calc) in a complete way.

Two Features/Two Layers in Two Languages

So, lets see what happens, if we have two features, e.g. Writer and Calc, and two localizations, e.g. English and German.

Applying the above separation, we get

• writer
• calc
• basis
• writer_en
• writer_de
• calc_en
• calc_de
• basis_en
• basis_de

So, what do the dependencies should now look like? By combining the above orthogonally we get

• writer depends on writer_lcl, basis, provides writer, feature
• calc depends on calc_lcl, basis, provides calc, feature
• basis depends on basis_lcl, provides basis,
• writer_en depends on writer, basis_en, provides writer_lcl, writer_en,
• writer_de depends on writer, basis_de, provides writer_lcl, writer_de,
• calc_en depends on calc, basis_en, provides calc_lcl, calc_en,
• calc_de depends on calc, basis_de, provides calc_lcl, calc_de,
• basis_en depends on basis, optionally on feature_en, provides basis_lcl, basis_en,
• basis_de depends on basis, optionally on feature_de, provides basis_lcl, basis_de.

To ensure consistence among layers, we need to make feature locales (e.g. calc_en) dependent on the lower layer locale (e.g. basis_en).

Looking at it, we the get the following minimal (and consistent) installations:

1. writer, writer_en, basis, basis_en, == Writer English
2. writer, writer_de, basis, basis_de, == Writer German
3. calc, calc_en, basis, basis_en, == Calc English
4. calc, calc_de, basis, basis_de == Calc German

Which we can combine to:

1. 1+2: writer, writer_en, writer_de, basis, basis_en, basis_de == Writer English/German,
2. 1+3: writer, writer_en, calc, calc_en, basis, basis_en == Writer/Calc English
3. 1+4: writer, writer_en, calc, calc_de, basis, basis_en, basis_de == Writer English / Calc German
4. 2+3: writer, writer_de, calc, calc_en, basis, basis_de, basis_em == Writer German / Calc English
5. 2+4: writer, writer_de, calc, calc_de, basis, basis_de == Writer/Calc German
6. 3+4: calc, calc_en, calc_de, basis, basis_en, basis_de == Calc English/German
7. 1+2+3: writer, writer_en, writer_de, calc, calc_en, basis, basis_en, basis_de == Writer English/German, Calc English
8. 1+2+4: writer, writer_en, writer_de, calc, calc_de, basis, basis_en, basis_de == Writer English/German, Calc German
9. 1+3+4: writer, writer_en, calc, calc_en, calc_de, basis, basis_en, basis_de == Writer English, Calc English/German
10. 1+2+3+4: writer, writer_en, writer_de, calc, calc_en, calc_de, basis, basis_en, basis_de == Writer/Calc English/German
11. 2+3+4: writer, writer_de, calc, calc_en, calc_de, basis, basis_en, basis_de == Writer German, Calc English/German

Note: Obviously we need to be able, to deal in our implementations with the fact, that different features may only be available in different locales. This is unavoidable, if we want to separate features. Or to phrase it differently, all consistent (depends, provides) sets of parts should be valid.

Complex Example

Splitting the Writer along some the of above identified categories (brand, OS/Architecture, locale), we get the following parts:

• writer____ - This is everything of the writer, which does not belong to any of the categories.
• writer___en - This contains all English localization content not belonging to any other category.
• writer_OOo__en - This part contains all English localization content for the OOo brand.
• writer___de - This contains all German localization content not belonging to any other category.
• writer_OOo__de - This part contains all German localization content for the OOo brand.
• writer__linux_ - This contains all Linux specific stuff.

To ensure consistency, we may model the dependencies and provisions as follows:

• writer____ depends on writer___OS/Arch, writer_brnd__lcl, provides writer____
• writer___en depends on writer____, provides writer___en
• writer_OOo__en depends on writer___en, provides writer_brnd__lcl
• writer___de depends on writer____, provides writer___de
• writer_OOo__de depends on writer___de, provides writer_brnd__lcl
• writer__linux_ depends on writer____, provides writer__OS/Arch_

This is one may of modelling the dependencies, but is it optimal? It at least seems to be minimal. But what happens if we can get rid of any brand specific localization, thus the writer_OOo__de respectively the writer_OOo__en parts? Even we don't change anything in writer____ it seems that we need to update it, because of changing dependencies, writer____ would now directly depend on writer___lcl.

Rules of Thumb

• A part may depend on lower layer parts.
• A part may depend on more general parts (e.g. the writer_OOo__en part depends on the writer_OOo__ and the writer___en parts)!
• A more specific part implies any more general parts (e.g. the writer_OOo__en part implies the writer___en part)!
• A virtual part is more general than a concrete part.

Algorithms

• Minimal dependencies.
• Minimal provisions.
• Minimal updates because of changing dependencies / provisions.
• Optimal dependencies / provisions.
• Installation variants of one set of parts.

Composition

Model products by setting them into

• inheritance, respectively
• instantiation (template)

relationship.

Inheritance

Inheritance models a "is a" relationship. In practice that would mean, that a StarOffice 8 update 7 is an OOo 2.2.1 (respectively its basis) adding something.

Instantiation

Example

Template Product OOo-Standard {
  Features: writer, calc, impress, draw
}

Abstract Product OOo2.4 {
  Name: OOo 2.4
  Code-Base: SRC680m236
  Implements: OOo-Standard

}

Product OOo2.4-ISO : OOo2.4 {
  Name: OpenOffice 2.4
  Format: ISO-750
  Platform: Linux-x86, Windows-x86, Mac OS X x86
}

Product OOo2.4-download-linux-x86 : OOo2.4 {
  Name: OpenOffice 2.4
  Format: donwload
  Platform: Linux-x86
}

Product OOo2.4-download-windows-x86 : OOo2.4 {
  Name: OpenOffice 2.4
  Format: donwload
  Platform: windows-x86
}

Abstract Product FooOffice3u4 : OOo2.4 {
  Name: FooOffice 3 update 4
  Features: foo-templates, foo-fonts, foo-brand
  Updates: < FooOffice 3 u 4
}

Product FooOOo3u4-ISO : FooOffice3u4 {
  Name: FooOffice 3 update 4 ISO
  Format: ISO-750
  Platform: Linux-x86, Solaris-x86, Solaris-Sparc, Windows-x86, Mac OS X x86
}

Product FooOffice3u4-donwload : FooOffice3u4{
  Name: FooOffice 3 update 4 ISO
  Format: download
  Platform: Linux-x86
}

Abstract Product : OOo_2.4 {
  Name: BarOffice 5 update 6
  Features: bar-templates, bar-fonts, bar-brand
  Updates: < BarOffice 5 u 6
}

Tooling

• Comparison of Installation Sets
• Check for conflicts
• Creation of Installation Sets
• Visualization
• Deliverable Table
• Module Table

Product Pipeline

The modelling by inheritance and instantiation needs to lead to deliverables which are re-usable during productization and after installation, ideally leading to zero redundancy in case of the installation of many variants (OOo and derivatives).

Taking a look at how packages (like RPM) are typically organized, we can see, that the intermediate deliverables should express their needs and offers in terms of

• dependencies against a (virtual) package, as well as by listing the
• provisions.

Naming Schema

A name schema reflecting this approach may look like this:

<feature>_<brnd>_<pltfrm>_<lcl>
writer_foo_linux_en

or generalized

 <feature>[_<dimension>]*

Packages independent of a particular dimension just leave this position empty.