On Sun, Mar 08, 2015 at 11:44:18AM +0100, Tobias Pape wrote:
> Hi,
> 
> On 08.03.2015, at 04:18, Levente Uzonyi <leves at elte.hu> wrote:
> 
> > On Sat, 7 Mar 2015, David T. Lewis wrote:
> > 
> >> On Sat, Mar 07, 2015 at 08:28:55PM +0100, Tobias Pape wrote:
> >>> Hey,
> >>> On 07.03.2015, at 19:01, gettimothy <gettimothy at zoho.com> wrote:
> >>> 
> >>>> The image fixes the WriteStream>>NextChunkPut error we where getting when trying to modify comments which makes me happy (:
> >>> 
> >>> to which fix do you refer?
> >>> There is a hack out there calling flush every now and then,
> >>> but this all calls for a principled solution for streams that
> >>> read _and_ write.
> >> 
> >> You are right that it is a hack, but the actual problem is a defect
> >> in some C runtime libraries, notably on Ubuntu but possibly others
> >> as well.
> > 
> > Are you sure it's a bug in the OS? Isn't it just the allocate-on-flush behavior?https://en.wikipedia.org/wiki/Allocate-on-flush

No, I am not sure. I had convinced myself that it was a libc issue, but I may be wrong.

> 
> I too doubt an OS bug.
> Someone (i don't remember) said to me, that the behavior of the c calls
> make clear not to rely on certain thing.
> I plan to investigate the issue soonish.
> 
> best
> 	-tobias

That would be great.

The workaround that "fixes" the comment problem is in WriteStream>>nextChunkPut:,
so if you remove the #flush at the end of this method, the problem can be
reproduced. At the time, I was not able to come up with a unit test that would
reproduce the problem, but my best guess as to what was happening is in the
update comment:


  Name: Collections-dtl.568
  Author: dtl
  Time: 5 May 2014, 12:39:30.026 pm
  
  Add a flush to WriteStream>>nextChunkPut:
  
  This is a workaround for a bug in the runtime library of some versions of
  Ubuntu. The symptom is that creation of a class comment for a class that
  previously had no comment leads to a file size error in the new RemoteStream
  that points to the class comment. Actual file size and contents of the
  changes file are not affected by this bug, and the error occurs when reading
  contents of the changes file immediately following the initial save, Flushing
  the stream after writing a chunk to the changes file prevents the problem.

Dave