[Python-Dev] small PATCH to fix os.fsync on OS X

[Richard Boulton]

James Y Knight wrote:
> On Aug 7, 2008, at 12:36 AM, Ian Charnas wrote:
>> While some people might suggest that fullfsync should be exposed as a
>> method on the 'os' module for platforms that support it, I believe
>> this is not the best approach.  Because the docstring for os.fsync
>> says that it forces the file to be written to disk, I think it's up to
>> python to ensure that functionality is really there on all platforms.
>> Currently on OS X, fullfsync is needed to provide this functionality,
>> so I think python's os.fsync should use fullfsync on OS X.
> 
> Please don't make that change. fsync on OSX does the same thing as fsync 
> traditionally does on other OSes. That is: it flushes the data to 
> "disk", but doesn't make any effort to ensure that the disk drive 
> actually flushes the data from its own write cache to the platter.
> 
> Other OSes just don't expose that functionality *at all*.
> 
> See e.g. this thread:
> http://lists.apple.com/archives/Darwin-dev/2005/Feb/msg00072.html
> 
> You should instead propose a patch to add F_FULLSYNC to the fcntl module 
> when present.

I can't think of a situation where it would be useful to do an fsync() 
in which you don't want the data to be flushed as far as possible.  If 
you're bothering to call fsync(), it's presumably because you need to 
guarantee that the data will be in a consistent state in the event of a 
failure (like a power cut).  So I think it would be helpful for an fsync 
call in a high-level language to handle the details of this.

However, my perspective is largely tied to databases, so there may be 
other situations that I can't think of where it does make sense to want 
to push changes as far as the on-disk buffers, but no further.

If the change suggested isn't to be made to os.fsync(), I think it would 
at least be helpful to add a note to the os.fsync docstring mentioning 
this limitation, and pointing to the F_FULLSYNC in the fcntl module when 
this is implemented.  And perhaps add an implementation of fsync which 
_does_ also do F_FULLSYNC elsewhere (shutil.fsync()?), to save all those 
users of fsync() who also want F_FULLSYNC from having to implement code 
which calls F_FULLSYNC when available and os.fsync() otherwise.

-- 
Richard