[RFC] try* interfaces in posix_mq 0.8.0pre

[RFC] try* interfaces in posix_mq 0.8.0pre

[Eric Wong]

I just pushed this out to RubyGems.org (and ruby_posix_mq.git).
Let us know what you think, thanks!

  posix_mq 0.8.0pre - kinder, gentler and less exceptional

  This adds trysend, tryreceive, and tryshift interfaces to avoid
  exceptions on common EAGAIN errors for non-blocking users.  EAGAIN
  during non-blocking messages is common when there are multiple
  readers/writer threads/processes working on the same queue.

  trysend is like send, except it returns true for success and false for
  EAGAIN.  send (still) returns nil, which I now consider a mistake but
  won't change until post-1.0...

  tryreceive and tryshift are like receive and shift respectively,
  but they return nil for EAGAIN and the same return values
  for their non-shift variants.

  None of these methods call mq_setattr() beforehand to set the
  non-blocking flag, it assumes the user set it once before they were ever
  called and never changes it.  Checking/setting the non-blocking flag
  every time is needless overhead and still subject to race conditions if
  multiple processes/queues keep flipping flag on the same queue
  descriptor.

  These interfaces are not yet final, feedback is appreciated
  at ruby.posix.mq@librelist.org.

-- 
Eric Wong

Re: [RFC] try* interfaces in posix_mq 0.8.0pre

[Iñaki Baz Castillo]

2011/2/27 Eric Wong <normalperson@yhbt.net>:
> I just pushed this out to RubyGems.org (and ruby_posix_mq.git).
> Let us know what you think, thanks!
>
>  posix_mq 0.8.0pre - kinder, gentler and less exceptional
>
>  This adds trysend, tryreceive, and tryshift interfaces to avoid
>  exceptions on common EAGAIN errors for non-blocking users.  EAGAIN
>  during non-blocking messages is common when there are multiple
>  readers/writer threads/processes working on the same queue.
>
>  trysend is like send, except it returns true for success and false for
>  EAGAIN.  send (still) returns nil, which I now consider a mistake but
>  won't change until post-1.0...
>
>  tryreceive and tryshift are like receive and shift respectively,
>  but they return nil for EAGAIN and the same return values
>  for their non-shift variants.
>
>  None of these methods call mq_setattr() beforehand to set the
>  non-blocking flag, it assumes the user set it once before they were ever
>  called and never changes it.  Checking/setting the non-blocking flag
>  every time is needless overhead and still subject to race conditions if
>  multiple processes/queues keep flipping flag on the same queue
>  descriptor.
>
>  These interfaces are not yet final, feedback is appreciated
>  at ruby.posix.mq@librelist.org.

Hi Eric, nice addition. In my benchmarks raising/rescuing exception is
really expensive. When a Ruby server gets congested and starts
raising/rescuing exceptions things become even worse. So I strongly
agree with your new features (similar to Gio in Unicorn, right?) :)

I will test it ASAP and comment here.



-- 
Iñaki Baz Castillo
<ibc@aliax.net>

Re: [RFC] try* interfaces in posix_mq 0.8.0pre

[Iñaki Baz Castillo]

2011/2/27 Iñaki Baz Castillo <ibc@aliax.net>:
> I will test it ASAP and comment here.

Hi Eric, I've cloned your repo and want to test it by building and
installing the gem but have found various issues:

- File .manifest doesn't exist (touch .manifest does the work).
- gemspec doesn't include lib/posix_mq.rb so such file is not
installed and the gem cannot be loaded.

-- 
Iñaki Baz Castillo
<ibc@aliax.net>

Re: [RFC] try* interfaces in posix_mq 0.8.0pre

[Iñaki Baz Castillo]

2011/2/27 Iñaki Baz Castillo <ibc@aliax.net>:
>> I will test it ASAP and comment here.
>
> Hi Eric, I've cloned your repo and want to test it by building and
> installing the gem but have found various issues:
>
> - File .manifest doesn't exist (touch .manifest does the work).
> - gemspec doesn't include lib/posix_mq.rb so such file is not
> installed and the gem cannot be loaded.

Apart from these temporal issues I've tested tryreceive and trysend
and they work well (as expected).

-- 
Iñaki Baz Castillo
<ibc@aliax.net>

[RFC] try* interfaces in posix_mq 0.8.0pre

[Iñaki Baz Castillo]

2011/2/27 Eric Wong <normalperson@yhbt.net>:
> Thanks!  Any comments/suggestions on the interface?

It looks just correct IMHO. Just one question: how about << method?
Currently it behaves as POSIX_MQ#send method (raises in non-blocking
mode id it cannot write) but it returns the POSIX_MQ instance itself
when success. I suggest that it could return true on success (as you
suggest for send method).
However it could also possible that << would behave as trysend
(returning true/false), but it would break backward compatiiblity.


> Thinking about it more, I really want to make POSIX_MQ#send return
> "true" on success now (and just raise on errors) for consistency with
> POSIX_MQ#trysend.

Sure.


> I doubt anyone actually checks the return value of POSIX_MQ#send...
> Maybe I'll change it and call the next release 1.0.0.

I don't check it as for now it doesn't make sense (nil says nothing,
and I already rescue the possible exceptions).



Regards.

--
Iñaki Baz Castillo
<ibc@aliax.net>

Re: [RFC] try* interfaces in posix_mq 0.8.0pre

[Eric Wong]

Iñaki Baz Castillo <ibc@aliax.net> wrote:
> 2011/2/27 Eric Wong <normalperson@yhbt.net>:
> > Thanks!  Any comments/suggestions on the interface?
> 
> It looks just correct IMHO. Just one question: how about << method?
> Currently it behaves as POSIX_MQ#send method (raises in non-blocking
> mode id it cannot write) but it returns the POSIX_MQ instance itself
> when success. I suggest that it could return true on success (as you
> suggest for send method).
> However it could also possible that << would behave as trysend
> (returning true/false), but it would break backward compatiiblity.

I actually considered doing something with POSIX_MQ#<<, but then I
realized people would expect to use it like an Array or Queue and
chain messages to it:

   mq << "FOO" << "bar"

They'd still get an exception if it returned anything other than the
mq object since the << fails on symbols/nil/false.

I considered an interface like this that would work like try/catch:

   mq.try do
     mq << "FOO" << "bar"
   end

But there's no telling at which point ("FOO" or "bar") the pushing
failed, either.   So I think << is really only good for people doing
blocking mq_send() (which I mainly use myself).

Thanks for the feedback!

Re: [RFC] try* interfaces in posix_mq 0.8.0pre

[Eric Wong]

Iñaki Baz Castillo <ibc@aliax.net> wrote:
> 2011/2/27 Iñaki Baz Castillo <ibc@aliax.net>:
> > I will test it ASAP and comment here.
> 
> Hi Eric, I've cloned your repo and want to test it by building and
> installing the gem but have found various issues:

The gem is installable via "gem install --pre posix_mq" from
RubyGemms.org

> - File .manifest doesn't exist (touch .manifest does the work).
> - gemspec doesn't include lib/posix_mq.rb so such file is not
> installed and the gem cannot be loaded.

I use "make gem" to generate teh manifest and gem in one go.

-- 
Eric Wong