To define dependency on a package or resource that's only needed after a process initialized

asked 2021-08-17 10:52:07 -0600

10937 ●203 ●339 ●330 http://www.linkedin.co...

updated 2021-08-18 13:52:54 -0600

Similar to #q264270but different question. In Catkin (or any modern package manifest formats defined in ros.org), what is the appropriate way to define a dependency on a resource that is NOT required to start and complete initializing a node/process (e.g. dynamic linking. Import in Python), but IS required later during runtime? ~~A good example in ROS (or any distributed system) is the package name of the publisher of a topic that your package subscribes to.~~

What I described above doesn't sound like to be in the scope of exec_depend (rep-0140).

I'm only using ROS1 but I assume this question is valid in ROS2 (not really sure though). To be more specific, this isn't even a ROS question. Ffrom ROS1 perspective, this is a Catkin package (which is technically ROS-agnostic) question.

Update: Elaborating an example per request. An example in ROS (or any distributed system) is when a node/process in your package subscribes to a topic/message that the process of another package/software publish. What are good and/or systematic ways to describe such dependency?. My primary usecase is in private company's product development so it can be narrow-sighted though, I'm seeing not defining this type of dependency put people in confusion. The issue I'm raising shouldn't be specific to ROS/Catkin though, I haven't found a good answer.

UPDATE2.

@Mike Scheutzow Thanks for deepening the conversation!

I think the user is abusing this dependency mechanism for the purpose they describe.

I dis/agree.

ros node A does not ordinarily declare a dependency on ros node B just because they have a pub/sub relationship.

Is your statement based on ROS practice? Or is that based even on the general software architecture broader than ROS? If the former I see why you say so, but it's never claried, hence my question on this page. If the latter, I don't have a clear answer but I'd be surprised if there's no software framework that allows to define this sort of dependency.

In general, there is no expectation that node B even runs on the same host as node A, so there should be no "requirement" that it be installed there.

My question is irrelevant to the location of nodes to be installed.

@gijs

Personally, I feel the dependency is on a type of information, not on any specific node or package.

Agreed. "Type of information" exactly (I used "resource" to mean that). That "information/resource" can be running nodes, publishers, or running processes (e.g. Docker engine), static files, etc. etc. I continue to be stuck with package name only because the framework I have an immediate need with, which is obviously Catkin, uses package name for dependency definition.

My initial motivation to ask this was that I've seen issues so many times found during runtime that turned out to be due to missing dependent resource that should be available during runtime but ... (more)

edit retag flag offensive close merge delete

Comments

It could help if you could give an example -- or describe what it is you actually want to do.

A good example in ROS (or any distributed system) is the package name of the publisher of a topic that your package subscribes to.

I'm not sure I understand this: publishers are bound by name, not by their location. What makes the package it is hosted by important in your system?

"packages" don't subscribe, nodes do.

gvdhoorn ( 2021-08-17 11:28:33 -0600 )edit

I added an update to elaborate my specific need, hope that helps

130s ( 2021-08-18 04:47:34 -0600 )edit

add a comment

2 Answers

Sort by » oldest newest most voted

answered 2021-08-17 11:42:26 -0600

gvdhoorn
86574 ●283 ●1432 ●1054 http://cor.tudelft.nl/

updated 2021-08-18 14:55:34 -0600

what is the appropriate way to define a dependency on a resource that is NOT required to start and complete initializing a node/process [..], but IS required later during runtime?

that would still be an exec_depend afaict.

From REP 140: Package Manifest Format Two Specification - Data Representation - <exec_depend> (multiple):

Declares a rosdep key or ROS package name that this package needs at execution-time

"execution-time" would be any time during execution.

Edit:

An example in ROS (or any distributed system) is when a node/process in your package subscribes to a topic/message that the process of another package/software publish. What are good and/or systematic ways to describe such dependency?.

there is currently no way to model this dependency in vanilla ROS 1. But that's mostly because plain ROS 1 does not come with any tools to model systems at this level. Tools like ros-model try to fill that gap (but there are more, mostly as research projects in the EU).

For ROS 2, this may improve somewhat. There is a (still open) PR (ros2/design#266) which adds something called the "Node Interface Definition Language" (sometimes coined NoDL), which allows a developer to describe the ROS API of a node in a more abstract way than reading source code, and to make it available offline (so no rosnode/rostopic/rosservice list needed) and machine readable.

Having written that: the fact a node subscribes to a specific topic with a specific type already seems to imply it has a dependency on something. Personally, I feel the dependency is on a type of information, not on any specific node or package. For services and action clients, the dependency would be on certain behaviour of the server. That does go into the direction of what the capabilities system was supposed to model.

In your latest edit, you still mention "package" and "software", so please clarify this: do you want to capture the requirement that specific software is present, or do you want to be able to express the requirement for certain topics, with certain types, to be available (ie: published)? Those are two different -- and in some way, disconnected -- things.

Edit 2:

My initial motivation to ask this was that I've seen issues so many times found during runtime that turned out to be due to missing dependent resource that should be available during runtime but somehow missed (by build, install, execution toolsets). In ROS toolkit, catkin, roslaunch-chk, roscompile are very helpful to capture missing deps, but even they can't catch the very type of dependency we're talking about.

So that's more the direction of ros-model I believe. Although that's really focused on modelling ROS systems.

Model driven engineering techniques, and Model Based system design would formally capture the kinds of requirements and constraints you seem to have in mind. They go much further than simple "this must be present" requirements though.

edit flag offensive delete link

Comments

I think the user is abusing this dependency mechanism for the purpose they describe. ros node A does not ordinarily declare a dependency on ros node B just because they have a pub/sub relationship. In general, there is no expectation that node B even runs on the same host as node A, so there should be no "requirement" that it be installed there.

Mike Scheutzow ( 2021-08-17 18:45:00 -0600 )edit

Yes, I'm assuming something like this as well. I sort-of tried to say this with:

publishers are bound by name, not by their location. What makes the package it is hosted by important in your system?

It sounds like the capabilities system would be better suited for this kind of dependency, but that was never really used very much.

gvdhoorn ( 2021-08-18 03:10:07 -0600 )edit

This sort of discussion is what I wanted to see I think. Though I marked this as an Answer, I interpret "don't do that unless you're desperate in defining what's ought to be loosely coupled dependency"...Clearly for public, general purpose library package, that's a no. And, good call for Capabilities! I like that concept. Unfortunately utilizing that architecture would result in re-design of our apps so cannot be done easily though.

130s ( 2021-08-18 05:07:23 -0600 )edit

add a comment

answered 2021-08-18 05:18:12 -0600

gvdhoorn
86574 ●283 ●1432 ●1054 http://cor.tudelft.nl/

updated 2021-08-18 14:55:46 -0600

I interpret "don't do that unless you're desperate in defining what's ought to be loosely coupled dependency"

I could see a use for an application package -- which would basically be a metapackage -- which essentially defines the set of packages (as exec_depends) which together contain the functionality required for a specific application to function correctly.

So plus_one_app1 could be a metapackage stating exec_depends on plus_one_secret_pkg1 and plus_one_generic_ui and moveit. If plus_one_app1 is then deployed, the requirements at the level you consider in your question would be satisfied (as in: the packages present in the workspace / on the machine are capable of satisfying the runtime requirements of all parts of the application, as deployed).

That's not completely the same as modelling "this node requires sensor_msgs/Image input at 10 Hz", which seems to be more like what you're after.

But by virtue of those specific packages being deployed together, the system designer / maintainer can make sure that level of requirements is implicitly satisfied as well. Even if we have no direct way of modelling such requirements right now in ROS proper.

edit flag offensive delete link

Comments

@130s: did you convert my comment to an (additional) answer?

gvdhoorn ( 2021-08-18 14:56:11 -0600 )edit

@gvdhoorn Nope, thought you did (also thought weird as I didn't think you'd do that).

130s ( 2021-08-18 16:28:28 -0600 )edit

No, I did not convert it to an answer.

gvdhoorn ( 2021-08-19 01:56:56 -0600 )edit

add a comment

To define dependency on a package or resource that's only needed after a process initialized

Comments

2 Answers

Comments

Comments

Question Tools

Stats

Related questions

To define dependency on a package or resource that's only needed after a process initialized edit

Comments

2 Answers

Comments

Comments

Question Tools

Stats

Related questions

To define dependency on a package or resource that's only needed after a process initialized