[RCR] Thread.abort_on_exception default to true

[Roger Pack]

Background:
currently if you have a thread that "dies" without being cleaned up


Thread.new {
raise 'you never see this message'
}
sleep

The error messages are basically swallowed, which I always find
surprising--they aren't for the main thread but they are for internal
threads.

RCR is to default to abort_on_exception to true.

Thoughts?
=r

 

[Roger Pack]

What do you mean by "clean up" here? If you care about the thread
status, you can call join, that reports the exception inside of the
thread.

I agree that using #join here would result in cleanup that would allow
the exception to no longer be swallowed--however I will explain my side
of it. Typically one uses join to mean only "wait for
termination"--using it as cleanup has always felt surprising to me, and
whenever I start a new multi-thread application, I am almost always
surprised by the fact that some error messages are swallowed.

here's a (somewhat contrived) ex:

"deep within a some server: "

Thread.new { send an email }

# continue running, handle next request, etc.

In this instance, I don't want to join on the thread because I am not
concerned about when it finishes--it doesn't matter to me, and I don't
want to wait for it as it would slow up processing my next request.
Therefore the exception is forgotten.


I only mention it because this just happened to me (again--this time
while debugging someone else's code, but), it surprised me once more,
hence my noting it here--perhaps as a compromise threads could output
the uncaught backtrace to stderr unless #join has already been called on
the thread? I realize there are problems with that, too, just thinking
for options.

Thoughts?
Thanks.
=r

 

[Brian Candler]

In this instance, I don't want to join on the thread because I am not
concerned about when it finishes--it doesn't matter to me, and I don't
want to wait for it as it would slow up processing my next request.
Therefore the exception is forgotten.

You don't care if or when it finishes, but you also don't mind if it
crashes your main program by raising an asynchronous exception?

 

[Roger Pack]

abort_on_exception means asynchronous exceptions will be delivered to
the main thread. That means any exception may happen on any point in
the main thread. From my point of view, asynchronous exception is
like a nightmare, especially for production code. Most code is not
exception safe at that level. That is the reason abort_on_exception
to be false by default. That is only for debugging purpose.

Hmm. True (I've even thought before that Thread#raise should be
deprecated--too dangerous, can interrupt code in a finally block, etc).

That being said, could it perhaps at least output something to stderr by
default to satisfy both cases?

Thanks.
-=r

 

[Roger Pack]

Something to stderr is only useful in debugging, but no good for
production code. When debugging, put -d to the command line options.
It automatically enable abort_on_exception.

Sounds good. I suppose the -d option is underused by most people
including myself.
=r