Re: after_worker_exit on murder

[Eric Wong <e@80x24.org>]

Simon Eskildsen <simon.eskildsen@shopify.com> wrote:
> With Jeremy Evans' work on `after_worker_exit`, I was hoping I could
> replace our internal fork which has a `before_murder` hook to report
> to monitoring systems when workers are forcibly killed. However, the
> `after_worker_exit` is only called on reaping—not when murdering.

Hi Simon, it looks like Jeremy clarified after_worker_exit for you.

Anyways...  I remember rejecting patches to add more to timeout
support in unicorn over the years since I did not want it to be
a crutch for application developers, or worse; a reason for
people to feel locked into unicorn.  Instead I wrote things like
https://bogomips.org/unicorn/Application_Timeouts.html to
discourage relying on unicorn's built-in `timeout' feature.

But, it seems like there's still a reliance on the built-in
timeout...

Why is that?  (If you're allowed to disclose)

I don't mean to put you guys (Shopify) on the spot,
as I'm sure other folks do it, too; but you're here :)

Anyways, is this something that could or should be improved
in Rack or ruby itself?  Or are there buggy external libraries
or even external dependencies like NFS to blame?

(Or, perhaps "I don't know" is fine :)

The Ruby timeout handling in stdlib could be way better if
done natively, I think.  Perhaps `ensure` could be better
declared, too; but there's still the problem of getting
external code to work well with it...


Anyways I am of two minds; on one hand, buggy code is fine!
unicorn handles it by nuking a process.  But on the other
hand... it's irritatingly inefficient and shoving things under
the rug still rots and stinks eventually.