Re: Not cleaning up tempfiles for multipart?

[Eric Wong <normalperson@yhbt.net>]

Charles Oliver Nutter <headius@headius.com> wrote:
> On Sat, Mar 6, 2010 at 1:55 AM, Eric Wong <normalperson@yhbt.net> wrote:
> > Charles Oliver Nutter <headius@headius.com> wrote:
> >> If I'm correct, this is a bug. Tempfiles should not be relied upon to
> >> clean themselves up in response to GC, since you don't know when GC
> >> will fire...
> >
> > Why not?  Tempfiles are objects, too.  It's perfectly reasonable
> > to let GC clean them up like any other object.
> >
> >
> > I don't know why RewindableInput is monkeying with Tempfile internals,
> > though.
> 
> It's roundly considered by every programming language community in the
> world to be *really* bad form to let GC clean up IO. There's tons of
> reasons for this:

You're forgetting there are programming language communities that
consider GC to be bad form in the first place :)

Given that Ruby does have a GC, I'd prefer to be as lazy as possible and
let the runtime deal with it.

> * GC may not run as soon as you like, causing you to hold IO resources
> too long (and under heavy load, potentially use up all descriptors for
> a process)

MRI considers EMFILE/ENFILE to be analogous to ENOMEM, and would trigger
GC in that case.

> * On some systems, old objects get promoted to heaps that are only
> rarely collected, resulting in them potentially never GCing (or only
> GCing extremely rarely)

Those GCs could be made smarter about dealing with IO objects.

> * There's no reason you *shouldn't* be able to close resources when
> they're no longer used. In this case, close any files opened for a
> give request once the request has been processed (async libraries
> using those files should expect them to be closed and deal with that
> accordingly).

Of course being able to close IO objects explicitly is good.  In some
cases of socket/pipe IPC, it's the _only_ way to signal end-of-input.
HTTP/1.0 and HTTP/0.9 don't work otherwise.

> Bug-finding tools for several languages flag this sort of behavior as
> a high-priority bug. Languages like Go, C#, and Java have or are
> adding features to help guarantee you never do this. It's pretty much
> universally frowned upon.

As I understand it, C# and Java have to deal with operating/file
systems that don't work with unlinked files.  I guess JRuby is limited
to only being able to use things that work on those platforms, and can't
work with unlinked temporary files reliably, my condolences.

I haven't had a chance to look Go...

> When I tweeted about Rubyists leaving IO objects to be cleaned up or
> closed by GC, *everyone* agreed that it's a bug in the code, and that
> nobody should ever rely on GC to release File/IO resources.
> 
> What more can I say? :)

*everyone* meaning your followers on Twitter?  I suspect your followers
on Twitter are more inclined to agree with you :)

-- 
Eric Wong