-
Notifications
You must be signed in to change notification settings - Fork 422
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
Add support for spin_until_complete (take 2) #2475
base: rolling
Are you sure you want to change the base?
Add support for spin_until_complete (take 2) #2475
Conversation
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I'm sorry to just now be commenting on this change (even though it's the second iteration), but I wanted to revoice some concerns about the way that the change is being made.
rclcpp/include/rclcpp/executor.hpp
Outdated
if (condition()) { | ||
return FutureReturnCode::SUCCESS; | ||
} | ||
|
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This implementation is racy. If the condition switches after this line of code, and an infinite timeout is given, we got a dead lock.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think that's on the user since they provide the condition, but I can document it.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Currently the function is named spin_until_complete. Therefore I would expect, I can pass it a std::future or any condition, and it will do the correct thing, and protect me from this kind of races.
This is clearly not the case. I think it is not even possible with the current API.
To solve this, we would also need to pass a std::unique_lock, like done here:
https://en.cppreference.com/w/cpp/thread/condition_variable/wait
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Hm, as a second option, one could limit the timeout of spin_once_impl to a third optional argument 'condition_recheck_interval', so that we get a bounded wait time.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Wait, I don't see how this can cause an unwanted infinite wait/spin.
It is assumed that condition()
is initially false
. If it is true
from the get-go, then it just returns without spinning. Otherwise, it spins once before checking if condition()
is true
, otherwise it spins once again, and returns if it times out. If condition()
never changes to true
and timeout
is negative, then yes it will spin infinitely, but that's intentional.
Am I missing something?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@wjwwood to quote your here :
"The thing is that if we merge this just because Monday is coming up, then we have to come back and undo what we merged through a deprecation cycle, which is very undesirable."
I still haven't seen any example of how this API would be used in a real word application.
@SteveMacenski you want this API for "Make the APIs easier to use / understand without exposing the technical details under the hood for the Jr Eng / researcher / student that shouldn't be required to learn the details, and"
The currently proposed API does not reach this goal, as you need to be aware that you need to wake the wait, and to do this, you need to either have a known side effect of another API (which can change by any internal implementation change) or explicit manage a guard condition.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
That is true, and I stand by what I said, we don't want to merge things that we think will be obsolete immediately.
However, I believe that the situation where you're pulling my quote from (#2466 (comment)) and this one are quite different. First, it's not clear to me that what you've proposed here would actually replace the desire for the new API in this pr, so we may or may not deprecate what this pr adds even with the approach you suggested implemented and merged later. Whereas in the other situation, it's more of an A or B style design question. If we choose A now, we'll have a mess switching direction to B in the future, and it would never make sense to have both. Therefore it's more important to act when we're sure it's the right approach, unlike here, where adding a new function is by comparison low risk.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I still believe, that the currently proposed API is a to big pitfall for end users.
E.g.
spin_until_complete([end = std::chrono::system_clock::now() + std::duration::Seconds(1)] ()
{
return end < std::chrono::system_clock::now();
});
will not work as expected. I also don't really see how the API will make the live of a user easier in its current state.
I also proposed
spin_until_complete(
const std::function<bool(void)> & condition,
std::chrono::duration<TimeRepT, TimeT> timeout = std::chrono::duration<TimeRepT, TimeT>(-1),
std::chrono::duration<TimeRepT, TimeT> condition_check_interval = std::chrono::millisecons(10))
this approach, which is trivial to implements, and would make the situation way better.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I also don't really see how the API will make the live of a user easier in its current state.
I also don't think the callback approach is really that helpful (but I think it can result in less code), for that you should talk with @SteveMacenski. But I still think it's no better or worse than using futures.
this approach, which is trivial to implements, and would make the situation way better.
Having the check interval (essentially polling) is bandaid, really the condition should always be attached to some activity that causes the spin to interrupt in an event driven way. If you want something like the check interval, you can just use the timeout argument and loop the function, or you could use an empty timer, etc. etc.
Most of the time this is used to wait for a Service Client to receive a response to its request, or something similar, and for that this function will work fine, and in fact the future-based version of this worked fine.
@christophebedard @fujitatomoya can you guys comment on whether or not to delay this until after the release or not?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I also don't really see how the API will make the live of a user easier in its current state.
I also don't think the callback approach is really that helpful (but I think it can result in less code), for that you should talk with @SteveMacenski. But I still think it's no better or worse than using futures.
Just to be clear, I picked up this set of PRs mainly to finally push them across the finish line. I personally do not have a need for this API in itself, so I'd also like to hear what Steve (or any other potential user) has to say here.
@christophebedard @fujitatomoya can you guys comment on whether or not to delay this until after the release or not?
At this point I'd prefer to just push it to make sure to get it right.
22223f1
to
449c07a
Compare
Co-authored-by: Hubert Liberacki <[email protected]> Co-authored-by: Tomoya Fujita <[email protected]> Signed-off-by: Hubert Liberacki <[email protected]> Signed-off-by: Tomoya Fujita <[email protected]> Signed-off-by: Christophe Bedard <[email protected]>
Signed-off-by: Christophe Bedard <[email protected]>
449c07a
to
0fb8046
Compare
Signed-off-by: Christophe Bedard <[email protected]>
I had a few questions/suggestions/comments, but generally the approach lgtm. I think once you've addressed them to your satisfaction @christophebedard, then it should be good to go. |
Signed-off-by: Christophe Bedard <[email protected]>
Signed-off-by: Christophe Bedard <[email protected]>
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
lgtm, now that we've resolved the discussion about the API for now.
Should we merge this now, or wait until the |
We should really merge them together, because I don't think it's good if we have rclpy and rclcpp deviate from one another. |
My CI above did not include your changes for rclpy unfortunately. We'll have to run separate CI for that. |
Any news on this? |
Reverts #1956, so it un-reverts #1874 (with some major updates), see #1874 (comment)
Replaces #1957
Closes #1821
This adds
spin_until_complete(condition, timeout)
.This also adds
rclcpp::Executor::spin_for(duration)
. Original Issue: #1821 (Added by @fujitatomoya)This goes along with the following PRs: