Contents
Development Model
Following a discussion thread on pcl-users@code.ros.org (see here), we propose the following development model for PCL.
Proposed development model
PCL will follow the ROS distribution model, with a major release every 6 months. The first major release (1.0) is tightly coupled with the release of Diamondback (aka D-Turtle), the next major release is to be expected in E-Turtle, the next one in F-Turtle, etc.
In principle, the versioning system can be chosen as similar to:
ROS (http://wiki/StackVersionPolicy), the Linux kernel (http://en.wikipedia.org/wiki/Software_versioning#Odd-numbered_versions_for_development_releases) -- with odd/even cycles (e.g., 1.0 is stable, 1.1 contains new features but is marked unstable, 1.2 is stable, etc);
OGRE (_link goes here_) -- with each minor release going through a development/unstable/stable cycle (1.x -- e.g., 1.6 is marked as stable and 1.6.x contains only fixes, all features are added to a 1.7 development branch which will be marked as stable when ready, then 1.7.x will contain only fixes, and all new features are added to 1.8, etc);
VTK (http://www.vtk.org/Wiki/VTK_FAQ#What_is_the_release_schedule_for_VTK and http://www.vtk.org/Wiki/VTK_FAQ#What_is_the_policy_on_Changes_to_the_API): VTK has a formal release every eight to sixteen months. [...] Between patch releases maintain the API unless there is a really strong reason not to. Between regular releases maintain backwards compatibility to the API with prior releases of VTK when doing so does not increase the complexity or readability of the current VTK or when the benefits of breaking the API are negligible.
- ...
[add more versioning models here if they are relevant]
Versioning Policy
The best model for a library like PCL which needs to iterate fast at the moment, seems to be given by http://en.wikipedia.org/wiki/Software_versioning#Change_significance: In principle, in subsequent releases, the major number is increased when there are significant jumps in functionality, the minor number is incremented when only minor features or significant fixes have been added, and the revision number is incremented when minor bugs are fixed.
We want to offer our users the chance to work on:
a stable branch, so they can build solid applications on top. We can keep the API 100% fixed and do only patch releases (or additionally we can consider adding minor features that do not break the API -- need a discussion about what's considered minor) -- so all in all, API should be 100% backwards compatible;
a devel branch, where we add some of the latest and greatest features. Here we do our best to keep the API stable, but there are no guarantees. For examples, if we discover that the current sample consensus segmentation API is not really good, we need to go ahead and change it.
... and we want both these branches in the same ROS distribution. Obviously non-ROS users will not care too much, and they will download whatever they think is best suited for them (more iterated and stable but without all the features -- e.g. stable, or potentially unstable but with better functionality -- e.g., devel).
That being said, we will try to guarantee:
up to 100% API backwards compatibility between releases. New releases should build on top of the old releases and add new features, and attempt not to break the base API through any core modifications;
any new API modification decision has to be supported by a strong case/argument. The proposer needs to clearly explain why the current API cannot accommodate the existing design decisions.
The above two points will hopefully address the -- what it may seem -- rapid development pace, as shown in the next section.
Release Schedule
With this in mind, the versioning roadmap could look as follows:
- D-Turtle (6 months):
PCL 1.0 -- released in D-Turtle -- stable after API review;
PCL 2.0 -- released in D-Turtle -- devel, active development goes here;
- E-Turtle (6 months):
PCL 1.0 -- marked as deprecated;
PCL 2.0 -- marked as stable after API review;
PCL 3.0 -- released in E-Turtle -- devel, active development goes here;
- F-Turtle (6 months):
- PCL 1.0 -- removed;
PCL 2.0 -- marked as deprecated;
PCL 3.0 -- marked as stable after API review;
PCL 4.0 -- released in F-Turtle -- devel, active development goes here;
Therefore, throughout a major release, we guarantee not to break the API for the stable branch. Our users can decide if their application needs stable iterated code, or whether it can live with more experimental code. The fact that they can go back and forth (hopefully with no changes or minimal changes in their code) in the same ROS distribution is useful for us -- as they might decide that it's worth helping us out to stabilize the devel branch sooner, as it contains very useful features, etc.
API Changes between Versions
Every new version of PCL needs to have its own namespace, so as to not collide with previous PCL versions. Example:
namespace pcl -- will define the namespace for PCL 1.x;
namespace pcl2 -- will define the namespace for PCL 2.x;
namespace pcl3 -- will define the namespace for PCL 3.x;
- ...
In addition, the include subdirectory structure needs to be changed too. The header files will be stored in:
include/pcl -- in PCL 1.x;
include/pcl2 -- in PCL 2.x;
include/pcl3 -- in PCL 3.x;
- ...
In theory, this allows the user to mix and match algorithms from different versions of PCL, although (s)he will be discouraged to do so, through the use of a best practices document. However, we do not want to make this a hard constraint. This means that although methods that operate on pcl::PointCloud<T> cannot be used on pcl2::PointCloud<T>, there's nothing that prevents the user to write a bridge method/class that does the conversion either through casting (if no changes have been made in that class) or data copying. We could also provide such a class for our users, but we will keep it in an external package, outside of PCL.
ROS Packages
The ROS PCL packages will be named according to their version number. We therefore expect:
ROS D-Turtle to contain two packages: pcl and pcl2;
ROS E-Turtle to contain three packages: pcl, pcl2, and pcl3;
ROS F-Turtle to contain three packages: pcl2, pcl3, and pcl4;
ROS G-Turtle to contain three packages: pcl3, pcl4, and pcl5;
- ...
Packages that depend directly on PCL and are bundled in the same ROS stack will follow the same convention:
pcl_ros will be named pcl_ros2 if it depends on pcl2;
pcl_visualization will be named pcl_visualization2 if it depends on pcl2;
- ...
Potential (late) future changes to the development model
The reasons for the above scheme are obvious. PCL needs to iterate fast in order to find the best base classes, best base algorithms, etc. Therefore the structure of the base classes might change from pclX to pclY, etc.
HOWEVER, we anticipate that after a few major releases, we will reach a point where most of the basic things have already been solved. This means that when that point is reached, we might change to a different development model, that includes:
one pcl base package which contains the core which is rock-solid API stable and we don't expect it to change for a long long time;
one pcl_experimental add-on package which contains new features and algorithms which make use of the base classes but do not need any modifications for them in order to run.
All new classes will therefore be added to pcl_experimental, until they become stable (through usage and API review) so they can be migrated in the pcl package.
This might affect both ROS and non-ROS users, as we will provide a -lpcl and -lpcl_experimental in two TGZ archives separately. See http://wiki/pcl/standalone for more information.
Even further into the future, we can imagine changes at the methods level, where we can re-consider things such as __attribute__ ((unstable)) [doesn't exist, but maybe we'll make one until then ;)], or BOOST_STATIC_ASSERT (USE_PCL_EXPERIMENTAL), etc.