After having used p2's provisional API for building p2 related tools, we have found several issues regarding the category management in 3.5. This page is an attempt to describe the issue, propose requirements on a solution, and then a suggested implementation.
- 1 What are the problems
- 2 Missing Features
- 3 Requirements on a Solution
- 4 Proposed Solution
What are the problems
Categories are versioned
Versioning of categories is a complication since the categorization escapes the original repository (repo A) where it was declared. When a second repository (repo B) mirrors/aggregates the content, it adopts any categorization as its view of the categorization of its content (unless it always rewrites them with new IDs - i.e. block them from being mirrored). When a new version of a categorization is added in the original repository, and repositories are aggregated A + B to repo C, what is the categorization in C?
Resolution basics must be bypassed when dealing with category IU's. The "use the latest version" idiom no longer apply. The UI tooling have special queries to ensure that the category versions does not get in the way. The same is true for the Buckminster aggregator, and will be true for a vast amount of other applications in the future.
Not even in the only place where the categories are used (the UI) can versions be used. Instead, they just add to the complexity.
The fundamental issue is that versioning of categorization is an odd concept, there is no way for a user to select a particular version of categorization, and even if there was, it would just be very confusing. The categorisation is simply "this repositories categorization of its content" - it is always current - it is what it is.
When older update sites are transformed to p2, a new category IU is created on each visit. Each newly created IU has a unique version based on a timestamp.
This leads to:
- In order to understand if categorization has changed, the current categorization has to be compared.
- Since it is not possible to know if the category is originally from a old update site, comparisons are always needed
A Category is an Installable Unit, but is not installed
This leads to:
- Exceptional logic whenever a category is found in a result.
- All usage requires the exceptional handling, but the only use case where categories have a role is when presenting a categorized view of what to install to the user.
The categorization in 3.5 only solves the need to present things to install in a human friendly way.
- All users of a repository are presented with the same view irrespective of if they are developers or users of a published software.
- As a consequence, it is not possible to organize "technical" things for technical people
- There is no structure to categories (i.e. no subcategories) - which in very large repositories becomes a problem
Requirements on a Solution
- it is unlikely that it is possible to force all older update site owners to change to p2, or have them name/version categories.
- Blocking older update sites in p2 is not a solution.
- It should be possible to modify the categorization of IUs in a repository without having to modify the IUs.
- Categorisation should be mirrored
Treat categorization as a third type of p2 repository information - there is already IU (metadata), and artifacts. The new kind represents meta data at the repository level. There is already such metadata captured in properties in the repository itself.
The property mechanism has limitations in its capability to handle more complex information, but could otherwise be used for this purpose. A compromise could be to have a property refer to an artifact (in the artifact repository) that contains the current categorization.