Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Parameter service behavior is inconsistent with the documentation of rcl_interfaces #2512

Open
mxgrey opened this issue Apr 23, 2024 · 9 comments
Assignees
Labels
bug Something isn't working

Comments

@mxgrey
Copy link
Contributor

mxgrey commented Apr 23, 2024

Bug report

Required Info:

  • Operating System:
    • Ubuntu 22.04
  • Installation type:
    • Any
  • Version or commit hash:
    • Iron / Humble
  • DDS implementation:
    • Any
  • Client library (if applicable):
    • rclcpp and rclpy

While implementing parameter services for rclrs, we noticed that the behavior of rclcpp and rclpy is inconsistent with the documentation of rcl_interfaces/srv/GetParameterTypes which says it returns a "List of types which is the same length and order as the provided names." and also "ParameterType.PARAMETER_NOT_SET indicates that the parameter is not currently set.".

However in the implementation of rclcpp and rclpy, if the node is set with allow_undeclared_parameters(false) and the service request includes an undeclared parameter, you get an empty response back:

$ ros2 service call /minimal_publisher/get_parameter_types rcl_interfaces/srv/GetParameterTypes "{names: [use_sim_time, hello]}"
requester: making request: rcl_interfaces.srv.GetParameterTypes_Request(names=['use_sim_time', 'hello'])

response:
rcl_interfaces.srv.GetParameterTypes_Response(types=[])

This contradicts the documentation for the service, which does not indicate any situation in which the response array would be a different size from the request array.

It seems the current behavior of rclcpp and rclpy was an intentional choice with the rationale that asking about an undeclared parameter should be seen as an error. However the mismatch between the behavior of the client libraries and the service documentation means either:

  • the client libraries must be changed to match the service documentation, or
  • the service documentation must be changed to match the client library behavior.

I would argue that the current service documentation makes more sense than the client library behavior, so the client libraries should change their behavior to comply with it for the following reasons:

  1. Returning an empty array is unhelpful whether the client is a human or a program. It provides the least amount of information possible in response to the service request and forces clients to always make multiple service requests in order to reliably detect the values of parameters.
  2. The concept of declared / undeclared does not even exist in rcl_interfaces. I see parameter declaration as an implementation detail for parameter management that's internal to a node. Clients of parameter services should not generally need to be concerned about parameter declaration settings within a node that they are querying, but the current behavior is leaking that abstraction into the behavior of the parameter services.
  3. An undeclared parameter in an allow_undeclared_parameters(false) node can be thought of as unsettable, and therefore its value is just permanently unset. Naturally this means any client asking for the value of that parameter should be told that the value of the parameter is unset.
  4. If we guarantee that the response array is always the same size as the request array, we reduce the likelihood that clients will write an accidental buffer overflow while iterating through the request and response arrays simultaneously because they weren't aware of the service behavior diverging due to parameter declaration.

Furthermore if a client calls rcl_interfaces/srv/SetParameters with an undeclared parameter on a node with allow_undeclared_parameters(false) then the client libraries should return something like { successful: false, reason: "Parameter is not declared by the node, which only allows declared parameters" } where the reason value is a fixed and consistent string literal across all client libraries so that parameter service clients can do a simple string comparison against. I would recommend defining the string literal inside of rcl_interfaces/msg/SetParametersResult with a name like REASON_UNDECLARED_PARAMETER. Currently rclcpp and rclpy produce different reason values for this same situation. Admittedly the service documentation says that reason should only be read by humans, but I think this particular case warrants a fixed and consistent value.

I'm happy to submit PRs for these recommended changes, but I thought I'd open a discussion on this first in case there are any strong objections. Tagging @fujitatomoya @ivanpauno and @wjwwood as the main people who participated in the original discussion.

@fujitatomoya
Copy link
Collaborator

@mxgrey

i may be mistaken something, but here are my comments.

are you suggesting we should remove allow_undeclared_parameters flag and rclcpp::exceptions::ParameterNotDeclaredException at all?
either allow_undeclared_parameters is true or false, we can get, type, set, list and describe parameters via service with default elements for those undeclared parameters? so the remote node can set any undeclared parameter via parameter service?

or allow_undeclared_parameters is only effective in set_parameter request? but the question is how we can know that parameter is not actually declared or just set the default values as dynamic typed parameter. this eventually causes the ParameterNotDeclared result for the application if that is undeclared.

i think that application can call list service 1st to know what parameters are declared and then call other services.

While implementing parameter services for rclrs, we noticed that the behavior of rclcpp and rclpy is inconsistent with the documentation of rcl_interfaces/srv/GetParameterTypes which says it returns a "List of types which is the same length and order as the provided names." and also "ParameterType.PARAMETER_NOT_SET indicates that the parameter is not currently set.".

this needs to be fixed and consistent, either way we go. thank you for pointing this out.

Returning an empty array is unhelpful whether the client is a human or a program. It provides the least amount of information possible in response to the service request and forces clients to always make multiple service requests in order to reliably detect the values of parameters.

true.

but server can response quick to protect itself. if undeclared parameter request is included, it can return immediately with empty result, because that is not permitted operation for the parameters that this node possesses. otherwise, server needs to deal with many default elements for each request from multiple clients.

The concept of declared / undeclared does not even exist in rcl_interfaces. I see parameter declaration as an implementation detail for parameter management that's internal to a node. Clients of parameter services should not generally need to be concerned about parameter declaration settings within a node that they are querying, but the current behavior is leaking that abstraction into the behavior of the parameter services.

based on this description, remote nodes can set any undeclared parameters via service request. i am not sure if this is good idea.

An undeclared parameter in an allow_undeclared_parameters(false) node can be thought of as unsettable, and therefore its value is just permanently unset. Naturally this means any client asking for the value of that parameter should be told that the value of the parameter is unset.

i am not really following this, can you rephrase this a bit?

what if that is declared parameter with dynamic typing? at the node initialization it cannot specify the type, but later on it sets the type and value to align with connected sensor mode and types.

If we guarantee that the response array is always the same size as the request array, we reduce the likelihood that clients will write an accidental buffer overflow while iterating through the request and response arrays simultaneously because they weren't aware of the service behavior diverging due to parameter declaration.

i understand this, but technically this is just a buffer overflow bug in the application...

Furthermore if a client calls rcl_interfaces/srv/SetParameters with an undeclared parameter on a node with allow_undeclared_parameters(false) then the client libraries should return something like { successful: false, reason: "Parameter is not declared by the node, which only allows declared parameters" } where the reason value is a fixed and consistent string literal across all client libraries so that parameter service clients can do a simple string comparison against. I would recommend defining the string literal inside of rcl_interfaces/msg/SetParametersResult with a name like REASON_UNDECLARED_PARAMETER. Currently rclcpp and rclpy produce different reason values for this same situation. Admittedly the service documentation says that reason should only be read by humans, but I think this particular case warrants a fixed and consistent value.

totally agree on this.

@mxgrey
Copy link
Contributor Author

mxgrey commented Apr 26, 2024

are you suggesting we should remove allow_undeclared_parameters flag and rclcpp::exceptions::ParameterNotDeclaredException at all?

This issue report is only concerned with getting the behavior of the parameter services of the client libraries to conform to the documentation of the services in node_interfaces. The implementation and API of the client libraries are a separate matter.

To make this discussion easier, I've gone ahead and implemented the changes I recommend in a draft PR #2515

There are only two things I left out of the draft PR that I think could make the situation better, but they would be somewhat disruptive, so I think there should be some buy-in before trying to do it:

  • The error message given for reason when trying to set an undeclared parameter should be a constant value in rcl_interfaces/msg/SetParametersResult. In the draft it's simply a magic string literal.
  • The use of node_params->describe_parameters is clunky, feeding in one parameter name at a time. This could be improved if we do one of the following:
    • Change the behavior of NodeParametersInterface::describe_parameters to not care about whether undeclared parameters are allowed.
    • Change the API of NodeParameterInterface::describe_parameters by adding a bool argument to ignore whether or not declared parameters are allowed.
    • Add a describe_parameter method to NodeParameterInterface.

how we can know that parameter is not actually declared or just set the default values as dynamic typed parameter

If "we" are the client that's calling _/get_parameters, then I would say the difference shouldn't matter. Either way the value is simply unset. On the other hand if "we" are the client that's calling _/set_parameters then you'll find out the value cannot be set by the { sucessful: false, ... } response that you receive. There are many reasons that a _/set_parameters request might fail to set a value, and the allow_undeclared_parameters(false) situation is only one of them.

otherwise, server needs to deal with many default elements for each request from multiple clients.

You'll see in my draft PR that we actually cut down on complexity and lines of code with my suggestion, except for describe_parameters because of some rigidity in the NodeParameterInterface::describe_parameters API.

based on this description, remote nodes can set any undeclared parameters via service request. i am not sure if this is good idea.

Users can set parameter handler callbacks that can reject an attempt to set the value of a parameter for any reason. allow_undeclared_parameters(false) would simply be one reason to reject an attempt to set a parameter to a certain value. I'm not actually proposing any change here except to standardize the reason given in this situation.

what if that is declared parameter with dynamic typing? at the node initialization it cannot specify the type, but later on it sets the type and value to align with connected sensor mode and types.

From the perspective of rcl_interfaces parameter services, there is no difference between the value or type of an unset dynamic parameter versus an undeclared parameter. You would only see a difference in two places:

  • The unset dynamic parameter will appear during _/list_parameters whereas an undeclared parameter will not
  • The unset dynamic parameter can have its value changed during _/set_parameters whereas an undeclared parameter in an allow_undeclared_parameters(false) node cannot

i understand this, but technically this is just a buffer overflow bug in the application...

I would argue that if the application is complying with the documentation of rcl_interfaces (which says the size of the response vector will always match the size of the request vector) while rclcpp and rclpy are not complying with the documentation, then the bug is in rclcpp and rclpy.

@fujitatomoya
Copy link
Collaborator

To make this discussion easier, I've gone ahead and implemented the changes I recommend in a draft PR #2515

thanks, i will take a look!

The error message given for reason when trying to set an undeclared parameter should be a constant value in rcl_interfaces/msg/SetParametersResult.

agree, so that application can check against it.

The use of node_params->describe_parameters is clunky, feeding in one parameter name at a time. This could be improved if we do one of the following

sounds good.

The unset dynamic parameter will appear during _/list_parameters whereas an undeclared parameter will not

that is why originally i was thinking this comes down to call list_parameter service to know declared parameters, and then call other services.
I am not really sure returning PARAMETER_NOT_SET values for the undeclared parameters via service would help this situation, that is my original question.

Just idea, what if we have PARAMETER_NOT_DECLARED state? and return the same size of array requested via service? does that help the application?

I would argue that if the application is complying with the documentation of rcl_interfaces (which says the size of the response vector will always match the size of the request vector) while rclcpp and rclpy are not complying with the documentation, then the bug is in rclcpp and rclpy.

yes, this is inconsistent and bug in the system. we need to fix for sure. what i meant is after fixing this inconsistency...

@mxgrey
Copy link
Contributor Author

mxgrey commented Apr 28, 2024

Just idea, what if we have PARAMETER_NOT_DECLARED state?

I don't think it makes sense to consider "Undeclared" a different type from "Unset". Consider a node that has allow_undeclared_parameters(true) instead of (false). Would that node return NOT_DECLARED or would it return UNSET when a user queries for an undeclared parameter? It seems to me like you're trying to avoid returning "Unset" because you want to indicate to the caller that the parameter cannot be given a value, but if the node has allow_undeclared_parameters(true) then any undeclared parameter can be set, which would then make it misleading or at least ambiguous to return NOT_DECLARED.

Rather than thinking about "undeclared" as its own type I would recommend thinking in terms of constraints on the values. Suppose I write a node and declare a string-type parameter called DEVICE_NAME. I could choose to add a custom parameter event handler that will only allow fully-uppercase string values for DEVICE_NAME. If a _/set_parameters client tries to give the DEVICE_NAME parameter a value with any lowercase characters in it, they would receive a response of { successful: false, reason: "ONLY UPPERCASE PLEASE" }. I am allowed to do this and parameter clients are expected to handle it by examining the reason given when an attempt to set the value fails. We would not suggest adding a new type to rcl_interfaces called UPPERCASE_STRING, because there is nothing special about the data type for my parameter, there are simply some constraints on what values it can have.

Similarly, there is nothing special about the data type of an undeclared: The data type of any undeclared parameter is the null set, also called the unit type, which rcl_interfaces expresses as "Unset". The only thing that is special about undeclared parameters is that when a node has the allow_undeclared_parameters(false) setting, the undeclared parameters are constrained to remain null. Nothing is different about their actual data type, and whether or not they behave any differently than a regular "Unset" type depends on the allow_undeclared_parameters node setting which isn't generally visible outside of the node.

In general I don't think there's a point to treating the allow_undeclared_parameters(false) constraints as a special case when ultimately users can apply any constraints they want to any parameters they want, declared or undeclared, and parameter clients need to be prepared to recognize and handle any arbitrary constraints anyway.

Undeclared parameters can be thought of as dynamic-type parameters which currently have a value type of "Unset", and the allow_undeclared_parameters node setting determines whether those undeclared parameters are constrained to remain unset or can be given a new value type by parameter clients. The { successful: false, reason: ... } response of _/set_parameters is the mechanism in rcl_interfaces for communicating the constraints on parameter values, and I think that's sufficient for undeclared parameters in allow_undeclared_parameters(false) nodes.

@mxgrey
Copy link
Contributor Author

mxgrey commented Apr 28, 2024

that is why originally i was thinking this comes down to call list_parameter service to know declared parameters, and then call other services.
I am not really sure returning PARAMETER_NOT_SET values for the undeclared parameters via service would help this situation, that is my original question.

The responsibilities of the different parameter services are pretty clear in my mind:

  • If a client just wants to know the values of some parameters, they check _/get_parameters. If they don't care about why some parameters may be unset, then this is sufficient.
  • If a client just wants to know which parameters are declared, they check _/list_parameters.
  • If a client wants to know whether an unset parameter has that value because it's undeclared, then they check both _/get_parameters and _/list_parameters.
  • If a client wants to set the value of a parameter (which is the only reason they should really care about declared/undeclared) then they will call _/set_parameters when the parameter does not have the value that they want, and then they will discover that it cannot be set based on the response. In this case it would be enough for them to just follow this sequence:
    1. Call _/get_parameters and find out that the parameter they care about is not the desired value because it is currently unset
    2. Call _/set_parameters to try to set the desired value and find out from the service response that it cannot be set because it is undeclared on a node that has allow_undeclared_parameters(false)
    3. Handle this in whatever way is appropriate for their application

@fujitatomoya
Copy link
Collaborator

@mxgrey

thank you very much iterating and sharing thoughts, i was checking myself what could be missing here. but now it is clear to me!

Consider a node that has allow_undeclared_parameters(true) instead of (false). Would that node return NOT_DECLARED or would it return UNSET when a user queries for an undeclared parameter?

i was thinking that, NOT_DECLARED is assigned to undeclared parameters regardless of allow_undeclared_parameters flag. so the user application can know what is declared or undeclared at least.
but as you mentioned, this can be more complication for user application. user application perspective, there is no difference between NOT_SET and NOT_DECLARED...

In general I don't think there's a point to treating the allow_undeclared_parameters(false) constraints as a special case when ultimately users can apply any constraints they want to any parameters they want

i think this is originally developed exclude any parameters to be set that do not belong to node. but true, we can do this with default add_on_set_parameters_callback callback, and reject all undeclared parameters to be set.

The responsibilities of the different parameter services are pretty clear in my mind:

sounds reasonable. i am now inclined to take this consistent behavior along with set_parameters sevice to return the same size of requested array as interface describes.

you already opened a couple of draft PRs, #2515 and ros2/rcl_interfaces#164, thanks. (i think we need to update rclpy aligned with #2515)

before moving to the implementation, can we keep this open for a bit to get more feedbacks?

@wjwwood @mjcarroll @alsora @clalancette what do you think?

@fujitatomoya fujitatomoya self-assigned this Apr 29, 2024
@fujitatomoya fujitatomoya added the bug Something isn't working label Apr 29, 2024
@fujitatomoya
Copy link
Collaborator

@mxgrey i added this to Client WG meeting. hopefully we can get some feedback from there.

@alsora
Copy link
Collaborator

alsora commented May 4, 2024

Reporting here my thoughts (discussed in the client library WG).

TL;DR: I think we should change the implementation to always match the documentation.

I agree that the current behavior is not intuitive, and indeed we stumbled on it also in our application.
In particular, having different results depending on the value of allow_undeclared_parameters.

As an additional point, the current implementation (which returns an empty vector if something is not present) doesn't work well with "optional parameters".
For example, if I'm interested in a bunch of parameters and I don't know which ones are declared or not, the only option is currently to request them one by one, because if requesting more than on at the same time I would get an empty vector if any was missing.

I also agree on having an explicit success field in the service response.
I think that it's good practice for all services to have that (and also a message field).

Note that this change will break the user applications (for example in my application we currently check for an empty vector to know if the service failed...), but I think it's ok and worth to do it.

@mxgrey
Copy link
Contributor Author

mxgrey commented May 17, 2024

My main blocker right now for finishing up the PRs for this is how to improve the description fetching.

The essential problem is that NodeParametersInterface::describe_parameters has the exception-based behavior baked in pretty strongly, which makes it awkward to implement the service with it correctly. I have three proposals for how to fix this, and would appreciate feedback from maintainers about which would be least controversial:

(1) Add a new argument to describe_parameters to override the behavior when allow_undeclared_parameters is false:

std::vector<rcl_interfaces::msg::ParameterDescriptor>
describe_parameters(
  const std::vector<std::string> & names, 
  bool override_allow_undeclared_false=false) const = 0;

(2) have a function that gets parameter descriptions one at a time, returning std::nullopt when the parameter is undeclared:

std::optional<rcl_interfaces::msg::ParameterDescriptor>
describe_parameter(const std::string & name) const = 0;

(3) reimplement NodeParameters::describe_parameter to just not behave differently under allow_undeclared_parameters(false), i.e. have it return blank unassigned-type descriptions just like the service describes, instead of throwing an exception.

Likely controversies:

  • For (1) the argument name is clunky and confusing, and I can't think of a way to make it less confusing. Also this suggestion breaks the abstract interface API.
  • For (2) there would be a lot of redundancy between describe_parameter and describe_parameters. Also this forces all implementers of the interface to add this new function.
  • (3) would be the easiest and cleanest, but it would be inconsistent with the exception-driven behavior that the other rclcpp APIs currently have for undeclared parameters.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
bug Something isn't working
Projects
None yet
Development

No branches or pull requests

3 participants