actor-framework/chat

CAF: The C++ Actor Framework

Yifei Yang

@Yifei-yang7

Hey I also updated from 0.17.6 to 0.18.5, and found caf::message_view is not an api anymore, may I know its alternative?

Dominik Charousset

@Neverlord

@Yifei-yang7 there isn't an analog in the new implementation since the entire type_erased_tuple-API has been cut out.

Yifei Yang

@Yifei-yang7

got it, thx! Is it possible to remote_spawn actors with detached option? Is it achievable by adding it into make_massage()?

Yifei Yang

@Yifei-yang7

or is it possible to change the api of add_actor_type to let spawn_option in? cause I found add_actor_type specifies no_spawn_option internally

Dominik Charousset

@Neverlord

The message only wraps the constructor arguments. Doesn't seem like there's an API for that at the moment.

Feel free to make a feature request on GitHub. :)

Yifei Yang

@Yifei-yang7

Thanks! I will make one for that. :)

Dominik Charousset

@Neverlord

👍

Rich Henry

@phaistos

im having the same issue as @ferjif above, i core dump whenever i try to use an atom defined with the new API, im following the pattern in the documentation

Dominik Charousset

@Neverlord

Are you passing your type IDs to CAF_MAIN or call init_global_meta_objects?

Here's a longer answer to a similar issue: https://stackoverflow.com/questions/70353846/segmentation-faults-when-testing-typed-actors-with-custom-atoms

Rich Henry

@phaistos

fixed, thanks a lot

i didnt see any discussion of that method, or the passing of arguments in the docs, but i didnt read them very closely either

and on closer inspection i see it

Dominik Lohmann

@dominiklohmann

If I have an actor behavior that takes a parameter by reference, would it theoretically be safe to store that reference elsewhere if I also manually keep a reference count to the current message as long as I keep the reference? How would I go about getting a strong reference to the current message from a "self" pointer?

Dominik Charousset

@Neverlord

self->current_mailbox_element()-> payload gives you the message in a handler. Keeping a reference/pointer into that message is only safe for read-only access because you otherwise break COW (changing a tuple with ref count > 1).

Dominik Lohmann

@dominiklohmann

It is for read-only access purposes, thanks! Not sure if you've seen my discussion comment regarding coroutines w/ CAF result (open to chat about it any time!), but I kind of need to extend the current message lifetime until the coroutine frame is destroyed because by-reference function parameters are only stored by-reference in the coroutine frame.

Dominik Charousset

@Neverlord

I've seen the notification, I'll get back to that if I can make the time. Are you still on 0.17? In that case it's a bit more complicated. Because in 0.17 CAF can fuse the content into the mailbox element itself. You could do something like self->current_mailbox_element()->move_content_to_message() to have something to hold onto.

But that of course invalidates the original reference.

Dominik Lohmann

@dominiklohmann

Yeah we sadly still are. The upgrade had to move back due to limited engineering resources. We still have some stuff serialized with the caf::binary_serializer, and the output of that changed between 0.17 and 0.18, so we're in the process of transitioning our database state in a forward-compatible way.

Invalidating the original reference is not a problem as I need to have a lifting function around every behavior callback anyways, I can move it in there. Thanks!

Dominik Charousset

@Neverlord

Depending on how long the migration goes, maybe you can skip 0.18 entirely and jump to 0.19. ;)
Good luck!

Dominik Lohmann

@dominiklohmann

Probably another 1-3 months, depending on customer needs :)

Dominik Lohmann

@dominiklohmann

Totally separate question:

self->set_down_handler(...);
auto handle = self->spawn<caf::monitored>(my_typed_actor, args...);

typed_actor<...>::behavior_type my_typed_actor(typed_actor<...>::pointer self, Args... args) {
  if (!validate(args...)) {
    self->quit(make_error(...));
    return typed_actor<...>::behavior_type::make_empty_behavior();
  }
  return {
    // ... 
  };
}

If the validation inside the my_typed_actor function fails the down handler will never be triggered, despite the actor shutting down correctly. Is this a bug, or are we simply abusing the API here?

Dominik Charousset

@Neverlord

I'd say it should trigger down messages. Do you have a minimal example that reproduces this?

Dominik Lohmann

@dominiklohmann

I'll open a PR with a unit test, assuming it still fails with 0.18 you can take it from there.

Dominik Charousset

@Neverlord

👍

Yifei Yang

@Yifei-yang7

Hi, may I know if message communication between cross-node actors is able to saturate the network bandwidth? For example if I try to transfer a large Arrow table between cross-node actors. Thank you!

Dominik Charousset

@Neverlord

The general advise we give is preferring small messages. Utilizing bandwidth shouldn't be an issue, but if you force the middleman to serialize KB or even MB worth of data it will 'block' him for quite a bit. So you could see latency spikes from such 'jumbo messages'. I'd say build a small prototype and measure how much latency you actually cause. The Prometheus metrics for the middleman should give you a good idea. Limiting the arrow tables to some reasonable size could also be a good idea if that's reasonable for your application.

Yifei Yang

@Yifei-yang7

Thanks for your advice. For serialization of middleman, as Arrow provide built in serialization methods, I call it inside inspect. So I guess the middleman can just transfer the serialized bytes from Arrow API, without incurring the serialization latency inside the middleman itself. May I know if this is true? But yeah for sure I will benchmark a bit to see the real effect. :)

Dominik Charousset

@Neverlord

Even if the serialization itself takes no time, you still have to copy data into the socket buffer. If that's a couple MB worth of data, it still can cause quite a bit of delay. Measurements are your friends. :)

Yifei Yang

@Yifei-yang7

got it, thanks! :-)

Kyle Klenk

@KyleKlenk

Hi, I would like to now if CAF can be configured to use green threads instead of hardware threads? We are running into issues where CAF is spawning more threads than cores on some of our systems and our admins are recommending that we consider using green threads.

Dominik Charousset

@Neverlord

CAF usually starts more threads than cores since slightly "overbooking" the system usually results in better performance. Some threads invariably have to wait at some point (e.g. due to memory access) and the extra threads can fill the gaps. The analogous abstraction to green threads in CAF are the actors, they all share a single thread pool for execution. The size of that thread pool can be configured with the parameter caf.scheduler.max-threads (https://actor-framework.readthedocs.io/en/stable/ConfiguringActorApplications.html).

Kyle Klenk

@KyleKlenk

Thanks for the explanation and pointing me to the configuration for specifying the number of threads.

Dominik Lohmann

@dominiklohmann

I've been looking at Flow recently, and was wondering how result types of observables integrate with typed actors? Is there an example for this, or are they just always lazily typed via caf::message?

Ah nvm, just came across caf::async::{consumer,producer}_resource<T>. That neatly solves what I wanted to do.

Dominik Charousset

@Neverlord

In general: an observable is private to an actor, so that wouldn't go into an actor interface. That's what the new stream<T> is going to be all about. The async building blocks only work for concurrency.

There's also some documentation available, btw.

Just do sphinx-build in the manual directory and you'll have the version that has some initial documentation for the new flow API. :)

html.zip

Or take the zip file.

Dominik Lohmann

@dominiklohmann

I see. I can see how a new stream<T> built on-top of data flows can work much better than the existing streaming API, which was a pain to debug and had lots of shutdown order issues.

If I understand data flows correctly they must be fully configured before their hosting actors launch. This is a limitation compared to the existing experimental streaming API which was fully configurable externally, and integrated into the typed actor API. Will this limitation also exist for the new stream<T>? I'm curious about your plans on the streaming front in general.

Dominik Charousset

@Neverlord

You can spin up flows at any time. You've probably seen the new spawn_inactive setups? You only need those if you're using observe_on, because the target actor for this operator must not run yet.

And, yeah. The old streams are going away for a reason. ;)

Dominik Lohmann

@dominiklohmann

I think there's a bug in the image for the concat operator: The concatenated stream can only start after the first stream is done, right?

Nvm, it actually can do the first one eagerly, that issue only occurs with at least three observables

Dominik Lohmann

@dominiklohmann

Currently just trying to get a feel of the API and what it's use cases and limitations are, so I started by comparing it to how we use the current streaming APIs. Kinda excited to toy around with it. I assume the new streaming stuff on top of it is all targeted for CAF 0.19?

Dominik Charousset

@Neverlord

Let me know if you do find bugs. The current flow branch is pretty much on its way back to master.

For CAF 0.19 we'll have flows and websockets (with caf_net), I'd like to see the new streams making the cut as well.

Dominik Lohmann

@dominiklohmann

New streams would likely be required for us from my experiments because flows cannot replace the old streams yet, unless that is staying for 0.19. The lack of dynamic setup with flows between actors after they've started will be a major hassle for us, because it requires rewriting how we start our ingestion pipeline. I'll let you know if I encounter any major issues, but so far I've seen none in the playground I've set up for my experiments.

If you need any design input for the new streams I'm always open to chat since we're really deep into the streaming internals by now. Would be really interested in some design doc if you have something like that.

Where communities thrive

People

Repo info

Activity