Re: Review Request 61058: Added a lock-free event queue.

[Dario Rexin <da...@mesosphere.io>]

> On July 31, 2017, 9:05 p.m., Dario Rexin wrote:
> > 3rdparty/libprocess/src/event_queue.hpp
> > Lines 328 (patched)
> > <https://reviews.apache.org/r/61058/diff/3/?file=1785564#file1785564line331>
> >
> >     I don't think this queue matches the requirements for an actor system. This queue uses sub-queues as a performance optimization and the ordering between the different queues is not guaranteed. The assumption here is, that causal ordering only needs to be maintained per thread. In an actor system however, there is no guarantee on which thread an actor will run. Think of following scenario:
> >     
> >     Actor A receives a burst of messages and takes some time to process them.
> >     
> >     Actor B receives a message gets dispatched to Thread 1 and sends a message to Actor A. The queue of Actor B is now empty, so it donates the thread back to the pool. 
> >     
> >     Actor A is still busy processing the messages it received earlier and did not process the message that Actor B sent.
> >     
> >     Actor B receives another message and gets dispatched to Thread 2. It now sends another message to Actor A. 
> >     
> >     Now Actor A received 2 messages from Actor B from 2 different threads. The moodycamel queue does not guarantee the order of those messages. This violates the arrival ordering defined in the actor model and makes it impossible to reason about message ordering between two actors.
> 
> Benjamin Hindman wrote:
>     For posterity, to account for this "deficiency" in the concurrent queue we've forced the _single_ consumer semantics on the queue and let the consumer reorder any events when the read them out. See the lock-free implementation of `EventQueue::try_dequeue` for more details.
> 
> Dario Rexin wrote:
>     Clarified with benh and benm. There's additional code that restores ordering.

-.- I forgot to click publish


- Dario


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/61058/#review181840
-----------------------------------------------------------


On July 29, 2017, 8:52 p.m., Benjamin Hindman wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/61058/
> -----------------------------------------------------------
> 
> (Updated July 29, 2017, 8:52 p.m.)
> 
> 
> Review request for mesos and Benjamin Mahler.
> 
> 
> Bugs: MESOS-7798
>     https://issues.apache.org/jira/browse/MESOS-7798
> 
> 
> Repository: mesos
> 
> 
> Description
> -------
> 
> Added a lock-free event queue.
> 
> 
> Diffs
> -----
> 
>   3rdparty/libprocess/configure.ac cb2cf4f32be5cbdf9df1e32f9aaf2bbba0a5ae03 
>   3rdparty/libprocess/src/event_queue.hpp PRE-CREATION 
>   3rdparty/libprocess/src/process.cpp b268cdad776a3ca2a87cbe60eb098bde2a70667c 
> 
> 
> Diff: https://reviews.apache.org/r/61058/diff/3/
> 
> 
> Testing
> -------
> 
> make check
> 
> 
> Thanks,
> 
> Benjamin Hindman
> 
>