aurora-reviews mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "John Sirois" <john.sir...@gmail.com>
Subject Re: Review Request 41915: Fixup TaskObserverTest to respect thread memory models.
Date Tue, 05 Jan 2016 05:08:09 GMT


> On Jan. 4, 2016, 8:25 p.m., Maxim Khutornenko wrote:
> > src/main/python/apache/thermos/observer/task_observer.py, line 72
> > <https://reviews.apache.org/r/41915/diff/1/?file=1181509#file1181509line72>
> >
> >     We usually put this default initialization directly into the arg list, e.g:'stop_event=threading.Event()'.
> 
> John Sirois wrote:
>     I'm leery of that for mutable objects like an Event.  Surprising things happen if/when
the containing object gets constructed a 2nd time and the single default Event has been mutated!
> 
> John Sirois wrote:
>     Maxim - your test originally so your call.  For a refresh - see here: https://reviews.apache.org/r/35527/
> 
> John Sirois wrote:
>     Disregard comment immediately above - now moved up under Bill's comment further above
where it belongs.
> 
> Maxim Khutornenko wrote:
>     Not sure how the current approach helps with anything. If a constructor is called
once with `stop_event` and another time without it, wouldn't we end up in the same state with
mutated field due to `stop_event` arg being None during the second call?

The issue is 2 or more calls without.  Say the interleaving goes like this:
```
to1 = TaskObserver(createPathDetector())
to1.stop()
to2 = TaskObserver(createPathDetector())
```

Since `to2` shares `to1`'s Event, its already `set` from the `stop` of `to1` and so never
runs.

Instead, by using `None` as the sentinel for the default, both `to1` and `to2` each get their
own fresh Event.


- John


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/41915/#review112750
-----------------------------------------------------------


On Jan. 4, 2016, 8:02 p.m., John Sirois wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/41915/
> -----------------------------------------------------------
> 
> (Updated Jan. 4, 2016, 8:02 p.m.)
> 
> 
> Review request for Aurora, Maxim Khutornenko and Bill Farner.
> 
> 
> Bugs: AURORA-1570
>     https://issues.apache.org/jira/browse/AURORA-1570
> 
> 
> Repository: aurora
> 
> 
> Description
> -------
> 
> Previously, a mock threading.Event was waited on in one thread
> and the count of waits was read in another thread.  Most thread
> memory models do not guaranty reads are fresh in this scenario
> unless there is a memory barrier of some sort forcing per-cpu
> caches to be flushed.
> 
> This change uses the underlying threading.Event as the memory
> barrier instead of mocking it and just wraps the event to record
> calls manually.
> 
>  src/main/python/apache/thermos/observer/task_observer.py      |  5 +++--
>  src/test/python/apache/thermos/observer/test_task_observer.py | 36 ++++++++++++++++++++++++------------
>  2 files changed, 27 insertions(+), 14 deletions(-)
> 
> 
> Diffs
> -----
> 
>   src/main/python/apache/thermos/observer/task_observer.py 1485de8faef52716f11b82a3556064de26c67427

>   src/test/python/apache/thermos/observer/test_task_observer.py ace15c5305e75fac3a82971f4d71b92bcb37bafc

> 
> Diff: https://reviews.apache.org/r/41915/diff/
> 
> 
> Testing
> -------
> 
> Before this change I got a failure between 1/5 and 1/10th of the
> time via:
> ```
> while true
> do
>   ./pants test src/test/python/apache/thermos/observer/ -- -kTaskObserverTest
> done
> ```
> 
> After the change I cannot trigger the failure.
> 
> 
> Thanks,
> 
> John Sirois
> 
>


Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message