[ros-users] catkin_make convenience improvements

Wed Jan 30 20:18:21 UTC 2013

On Wed, Jan 30, 2013 at 11:59 AM, Thibault Kruse <kruset at in.tum.de> wrote:

> I raised some concerns over catkin_make convenience in a meeting, here are
> the results of the brainstorming. catkin_make is mostly a convenience
> wrapper to cmake/make, for usage with the catkin build system starting with
> ROS Groovy.
>
> The current catkin_make design was guided by the desire to have a
> consistent and self-explaining workspace layout for tutorials and novice
> users, a quick way to create workspaces, and a wrapper around cmake to
> reduce cumbersome typing of common -D options.
>
> In it's current state, catkin_make may not as convenient as it could be.
> Current quirks are:
> - the workspace/src folder in itself is not popular with everybody
> - it is inconvenient to have to cd into the workspace folder to invoke
> catkin_make
>
+1

> - it is inconvenient that wstool commands work in the src folder and below
> only, currently, meaning another "cd" step
>
+1

> - invoking catkin_make once with custom options (for build/src locations),
> and then again without options does not remember the last options, and
> retyping the options is also cumbersome.
> - cleaning up builds (make clean, make distclean) does not work too well
> and is cumbersome, more support would be nice
>
> Since REP128 (http://ros.org/reps/rep-0128.**html<http://ros.org/reps/rep-0128.html>)
> does not make a strong effort to justify the design it defines, we have to
> also look at some use cases and find justifications.
>
> - We can consider 3 types of user groups who may want to use catkin_make
> -- total noobs: Those need one source space, one build configuration and
> that's it

+100

> -- advanced programmers: Those may typically have 2 build configurations,
> one with debug settings, one with performance optimized settings
> -- power users: Those may have a large number of build configurations for
> the same workspace, for cross compilation experiments and such
> I assume the toplevel "src" folder allows to have some cleanliness in
> particular for power users, having multiple build, devel and install spaces
> at the workspace root level, and packages one level below. On the other
> hand, I assume it would be similarly clean if all build, devel and install
> spaces were stored away in workspace/builds, and packages else reside in
> the workspace root.
>
> - catkin workspaces are used by several people using symbolic linking (for
> convenience, and maybe due to known limitations of catkin overlaying)
>
> - catkin also does env_caching, which needs to be taken into account for
> sequences of buildspace creation actions
>
> - there is also catkin_make_isolated, which currently creates a different
> folder layouts, maybe some consolidation could happen there as well
>
>
>
> So some design strategies that we came up with regarding more flexibility
> of invoking catkin_make and setting up workspace layouts according to
> personal preferences.
>
>
> Alternative "Marker-file":
> This idea involves placing a marker file at the workspace root, which can
> be used by catkin_make as point of reference when invoked in a subfolder
> (and maybe also by roscd without arguments). To support invoking
> catkin_make without arguments with custom paths, the marker file could have
> content that defines build configurations, e.g. the last one that was
> specified. For users that typically have multiple build configurations,
> this could be extended to have differently named configurations. Probably
> cmake already caches options for us, so really what could be stored in the
> marker file would be the path to a build space. However build-spaces get
> nuked regularly  (since we have no great "clean" target), so duplication of
> information might still make sense in the marker file (unless we find a
> nice "clean" command).
> Such a markerfile might also be made compliant with wstool, one way or the
> other. The disadvantage here is that some additional infrastructure is
> require, to be maintained.
>
> Alternative "Build Env Switching":
> This idea involves switching build configurations (rather than
> workspaces), meaning every build folder gets it's own file to source, like
> "setup.sh", though a different name might be better. Such a file would
> declare an env variable pointing to the build folder, and invoking
> catkin_make from anywhere would invoke cmake/Make in the folder of the env
> var (if set). Disadvantages are that this is less transparent to the user,
> requires the user to know about this setup file as well, and encourages the
> bad habit of reusing the same shell for different build environments.
>
> So this is just brainstorming, and anyone who feels passionately enough
> about catkin_make is welcome to contribute ideas, opinions, (informal)
> votes. The next steps could be to write prototypes and a REP. Once we start
> doing that, we'll probably take the discussion to the ROS Buildsystem SIG,
> but I'd be happy to get some feedback by mere catkin users.
>

Thibault,

Thanks for bringing up this design discussion! I'd also like to see the
catkin workflow made simpler and more intuitive for those not familiar with
FHS and the nitty-gritty of CMake. I have a bunch of thoughts / opinions /
feedback from using rosbuild and catkin, as well as explaining catkin to my
colleagues who are used to rosbuild.

Whenever I'm using catkin, I feel like it's been designed primarily for the
people who want to release and/or install code, and focuses on solving
those problems more than it does make it convenient for the student or
researcher who only ever builds his or her code from source.

What if, given a workspace, we could just have a bunch of source packages
and then we could put all of the "build", "devel", and "install"
directories under a hidden directory called ".catkin".  Then, inside of the
".catkin" directory we could have a config.yaml file. This combination of
".catkin/config.yaml" will act as the marker file, and keep the workspace
focused on the source code instead of the buildsystem.

Then a workspace approaches the simplicity of a rosbuild workspace managed
with rosws.

 - my_project_workspace/
   - .catkin/
     - build
     - devel
     - install
   - my_package_a...
   - my_package_b...
   - some_repo_x
     - some_package_c
     - some_package_d

Then have commands like "catkin loadws" to source .catkin/devel/setup.sh
(this also becomes more cross-platform than manually sourcing one of the
several shell scripts in your "devel" directory. I feel like if users are
going to be required to maintain this structure in their workspaces, then
it should be done with tooling like the philosophy behind rosws/wstool.

The main effect of this is that it brings the focus back onto the source
code, as opposed to the whole catkin infrastructure. I definitely
appreciate the huge amount of work that has gone into catkin, but I don't
want to have to be messing around with these additional layers of
directories. I want to be focusing on robotics code, and I just want
whatever build system I'm using to stay in the shadows and do its thing.

Rosbuild excels in this regard: as long as packages are under the
$ROS_PACKAGE_PATH, you can call "rosmake some_package" from anywhere, and
it just builds the code, and then to launch something you use rosrun or
roslaunch, and you don't care about the location of the built targets. I
feel like ROS already has an abstraction layer in place that catkin could
be dropped into so that by I???turtle, you can just call "rosmake" to
invoke catkin from an arbitrary place like directory, and you wouldn't even
know it's using catkin.

I understand that there are a ton of more advanced things that we'd like to
do (and new patterns that catkin enables) but I think it would be
unfortunate it we lost the ease-of-use of rosbuild.

What do others think?

-j

-- 
Jonathan Bohren
 PhD Student
Dynamical Systems and Control Laboratory
Laboratory for Computational Sensing and Robotics
The Johns Hopkins University

(707) 520-4736
jbo at jhu.edu
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.ros.org/pipermail/ros-users/attachments/20130130/cf6133cb/attachment-0004.html>