Proposal

NOTE: This proposal is obsolete. See the new Eclipse Plug-in Versioning proposal that has been adopted in Eclipse 3.2.

Eclipse Plug-in Versioning

Summary
The Eclipse team currently changes their plug-in version numbers to match each major release of Eclipse (e.g., 2.1, 3.0). This is a convenient way of understanding the origin of plug-ins but do not capture the actual semantics of the changes which occurred. For example, the change from 2.1 to 3.0 indicates a major and incompatible change. Of course, this is generally not what was intended as most 3.0 plug-ins are in fact compatible with their 2.1 versions. Here we outline the current use of version numbers and propose a new process for using plug-in version numbering to better indicate levels of compatibility.

Last Modified: November 21, 2004

Problem Definition

Current Eclipse practice calls for the teams to increment plug-in version numbers to match the upcoming release of Eclipse. For example, some months ago we changed our plug-in version numbers from 3.0.0 to 3.1.0. In May 2005 we will ship Eclipse 3.1 and all plug-ins (save a few third party ones) will be version 3.1.0. This is convenient because it allows people to look at the plug-in version number and immediatly understand what version of Eclipse that plug-in is from.

This approach has several drawbacks however

Incrementing the major (first) version segment renders the plug-in "incompatible" with all previous versions of that plug-in. Dependent plug-ins which specify a version number (e.g., <import plugin="org.eclipse.foo" version="2.1.0"/>) will not resolve against version 3.0.0 of the prerequisite plug-in. Unresolved plug-ins cannot participate in a running Eclipse environment. Even though we have worked hard to make the content of the 3.0 plug-ins binary backward compatible, 2.1 based third party plug-ins (which specify prereq version numbers) require changes to work in 3.0.
This approach does not set a good example. People building on top of Eclipse look to us to understand how they should structure their world. Following this model, plug-ins like EMF and GEF should increment their version numbers to match the Eclipse version numbers regardless of whether they changed at all. This has the nasty side effect of rendering all of their users incompatible as well. It also does not scale well as a little change at the bottom then ripples all the way up.
Eclipse is getting away with this because a) we are at the bottom of the stack, b) we ship a single wad that is built, packaged and distributed in unison and c) we do not use version numbered prerequisites ourselves. It is really this last point which hurts the most. By not using versioned prereqs, we are not eating our own dog food. Clearly we do not mean that the resources plug-in from 3.0 will run on ANY version of the Eclipse runtime. People who build and distribute plug-ins are not generally able to play it quite so fast and loose.
Features are not the cure. To address some of this, we use features and their abilily to group plug-ins. The assumption is that if the feature constraints are met, then the underlying plug-ins will all be happy and thus versioned plug-in dependencies are not needed. This works but assumes that the feature writer (may not be related to the plug-in authors) can understand the plug-in constraints well enough to define the feature. In practice this means that the plug-ins in the feature all have similar constraints such that they can be distilled to a smaller set of inter feature constraints. Features are a packaging mechanism for plug-ins. Pushing constraint satisfaction up to the packaging layer puts a significant load both on the people creating the packages (e.g., features) as well as the packaging (e.g., feature) mechanism. And finally, features are an update mechanism and update is not required to be in a running configuration.

Proposal

To address these concerns, we propose that the Eclipse SDK team start using the version number semantics defined by Eclipse for all their plug-ins. The numbering process works as follows:- new plug-ins start as 0.0.0 and go through various version numbers (<1.0.0) until they are ready for their first release. At that time, developers may choose to set the version number to 1.0.0. This is mostly cosmetic and is not required. Prior to first release the numbering process depends mostly on how widely availability and use of the early versions. For example, a beta at 0.9 sets a good tone and might encourage release at 1.0.0.

Immediately after a release, plug-ins continue with the same number as they had in the release (e.g., 1.0.0).
As the plug-in is developed (whether for a new release or for maintenance), the first three segments of the version number is changed to accurately reflect those changes. For example, if some minor fixes are done, the service (third) segment should be incremented (e.g., 1.0.1). If significant but compatible rework is done and/or new function added, the minor (second) segment should be incremented. (e.g., 1.1.0). Note that in this example, the significant work overshadows the service fixes so the service (third) segment is set to 0. If major rework is done such that the new version is incompatible with previous versions then the major (first) segment should be incremented (e.g., 2.0.0) Note that the minor and service numbers are 0 here as well.
A plug-in's version number need only be incremented to distinguish one release from another. For example, if we release 1.0.0 of a plug-in and then do some bug fixes for a milestone (say M3), we increment the version to 1.0.1. If that plug-in were eventually released, people could look at the version number and see that a) it is different from 1.0.0 and b) that some bug fixes were done. If instead, further bug fixes are done for M5, there is no need to further increment in the version number (i.e., to 1.0.2). Leaving the version at 1.0.1 still conveys the important information to users of the next release. Of course, teams could increment the version further if needed/desired.
Plug-in version number changes are solely at the discretion of the plug-in development team. They are in the best position to understand their changes and the scope of impact.
Features are used to group together plug-ins and label them with a particular "product" release version (e.g., 3.0). The org.eclipse.platform 3.0.0 feature contains a complete and precise specification of the plug-ins in the 3.0 platform.

Use of the qualifier

The current versioning scheme is effectively highlighting release families. A release family is a set of releases which form a substantial and separable effort. For example, Eclipse 2.0, 2.1, 3.0 and 3.1 are all different release families. The life of release family continues after its initial release (e.g., 3.0) to include maintenance releases (e.g., 3.0.1, 3.0.2, ...). Using family numbers as the base for plugin version numbers is convenient and easy to relate to but, as we have seen, conflicts with the Eclipse versioning semantics. The proposal outlined above summarily dumps release family notions in favour of semantically correct version numbers. This is has its own problems.

For example, assume we have implemented the proposal above.

On June 25, 2004 Eclipse 3.0 is released. org.eclipse.foo is version 1.2.0
The next day a maintenance stream is created in CVS to continue development for the 3.0.x family of releases.
On June 29, 2004 someone does a minor bug fix to org.eclipse.foo in HEAD and ups the plugin version number to 1.2.1 (correctly capturing the nature of the change)
On July 7, 2004 someone does a different minor bug fix to org.eclipse.foo in the maintenance stream and ups the plugin version number to 1.2.1 (correctly capturing the nature of the change)
nothing else happens to org.eclipse.help before both Eclipse 3.0.1 and Eclipse 3.1 ship.

In this case we have to different sets of content for org.eclipse.foo but the version number is the same. In the current model the plugins would have version numbers based on their release family and so their intended target would be clear. The question then is how to mix-in the release family progression with the plugin content progression. The first three segments of the version number have clearly defined semantics which do not match the needs for identifying releases or release families. The qualifier (fourth segment) however has no particular semantics and so is free for use.

We propose the use of a convention where the release family is the first two digits of the qualifier. In the above example this gives the following version numbers (respectively)

1.2.1.31
1.2.1.30

These are both semantically correct (i.e., numbering indicates the degree of change) and easily understood as being related to 3.1 and 3.0. This story is sufficient if we are only concerned about comparing version numbers from release to release. The other way in which the current (and above) scheme fall down is that during the development cycle, the version number of a plugin never changes even though its content may change dramatically. This makes it nearly impossible for the Update technology to be used successfully on milestone or integration builds. A refinement to the use of the qualifier resolves this.

Noticing that plugin version numbers are intended to relate to content, the qualifier should include the version tag from the map file used for the build. Eclipse is built by reading a map file which lists a CVS repository location and a version tag to use when fetching the content. If a team has no input for a build, they do not update the map file and the version (and thus the plugin) remains unchanged and identical plug-ins will not be duplicated on the update site and consumers will not have to download them if they already have them.

Using this approach the qualifier looks like <release family>_<tag>. In the example from above this approach give the following version numbers respectively:

1.2.1.31_v20040629
1.2.1.30_v20040707

This approach gives version numbers which correctly reflect the progression of the content and respects the release family progression.

Issues

Generating the warm fuzzies

The main issue that arises from this proposal is that some people have been getting a warm fuzzy feeling from the current state where plugin version numbers do "match" the release version number. Unfortunately this sense of happiness is false.

Between releases the plugin contents change but the numbers stay the same. This apparent matching is only relevant for final GA releases.
Some plugins (e.g., 3rd party) do not have matching numbers (e.g., lucene)
No warmth is generated regarding additions or deletions of plugins or features. Sure they are all numbered nicely but are the right things there?

The true warmth generator is a relatively simple tool that checks the installed set of pluigns/features against a bill of materials generated as part of the build. The output of this tools shows up in the Configuration Details dialog to warm service personel and in the feature/plugins views of the About dialog to warm users. For example, in the configuration details, any discrepencies are highlighted while in the About dialog the matchin is rendered, for example, as a green check mark beside plugins that match the bill of materials.

Existing uses of the qualifier

In some cases we currently use the qualifier to indicate emergency fixes or patches. Version numbers such as 3.0.1.1 indicate an emergency change made for 3.0.1 while it was in the field. This proposal does not conflict with that approach though it does indicate a different approach. 3.0.1.1 as a version number does generate some warmth as a minor fix to 3.0.1. Under our proposal the version numbering would look more like:

1.2.1.30_v20040911 = base version in the 3.0.1 release
1.2.1.30_v20041004 = patch for the plugin in the 3.0.1 release

Again, rather than using the version numbers themselves to generate (a somewhat false) warmth, the simple reporting tool would correctly identify the configuration as having been patched.

Action

The main action required is to decide we want to do this and refine/document the details of the structure and process. PDE build would need some very minor modifications to accomodate the addition of the release family number but otherwise there are no technological hurdles. This approach dovetails nicely with a proposal to enhance Eclipse's use of the update technology. See Update Site Proposal for more details.

Going forward teams would be responsible for updating the plugin version just as they are responsible for updating their copyright dates, etc. Note however that these changes are done very infrequently and will not be a particular burden.

Transitioning

If we decide to follow this new policy we can retrofit the 3.1 plug-ins with "correct" version numbers or start the new approach immediately following the release of Eclipes 3.1. Starting now avoids situations where Eclipse 3.0 based plug-ins do not work on the released 3.1 because of resolution problems (rather than because of actual incompatibilities). However, it may adversely affect people in the community who have already come to know, like and use the 3.1.0 version numbers.

A concrete proposal for how to do this follows:

Assume that the baseline version number is 3.0.0
All teams do a quick review of their plugin changes and update their plugin version numbers correctly reflect the degree of change since 3.0.0.
All version numbers are updated to have a ".qualifier" suffix (e.g., 3.0.1.qualifier).
The line "qualifier = context" is added to all build.properties files
PDE build is setup to add "31_" to the qualifier context for current builds