Instead of converting all float arguments to ints, this
first attempts to 
convert the argument to a PyLong, and then the old
conversion code takes over.

The upshot is that the limit of floating point
arguments to seek() is raised from (2 ** 32) - 1 to
somewhere above (2 ** 62).

Perhaps a better way to handle this would be to not
accept floating point arguments to seek().

Logged In: YES 
user_id=23486

The patch works against HEAD; updated for line numbers (there's a very 
similar set of line elsewhere in the file):

http://issola.caltech.edu/~t/transfer/patch-1067760-seek.diff

rhettinger thought the current behavior was reasonable (see bug 
1067728).  I'm not qualified to comment ;) but it seems odd that floats 
would have drastically different limits than longs.  This looks like a 
reasonable compromise IMO.

Logged In: YES 
user_id=23486

Sorry, I should add: patch works against head on Linux RH 9.0 (and 
should work everywhere).  Behavior tested with a script.

Logged In: YES 
user_id=23486

Bob Ippolito pointed out that nowadays float-->int conversion of
arguments is deprecated; a DeprecationWarning wasn't being raised 
because conversion was done manually in seek.  Patch to fix this
at

http://issola.caltech.edu/~t/transfer/patch-1067760-seekwarn.diff

Linux RH 9.0 works, regression tests run w/o problem.

The patch is against the current HEAD but can probably be backported w/
o a problem.

Logged In: YES 
user_id=23486

Updated patch to check return value of PyErr_Warn in case 'python -W 
error'  is specified.  No other objections raised on python-dev...

revised patch still at
http://issola.caltech.edu/~t/transfer/patch-1067760-seekwarn.diff

Logged In: YES 
user_id=562624

The actual usable range of a float is 2**53, not 2**62.
Above this the resolution drops below 1 byte. It's enough
for some 9000 terabytes.

Logged In: YES 
user_id=21627

I'm attaching a different fix, which uses the index API to
obtain the offset.
As a consequence, passing a float becomes an outright error.

Logged In: YES 
user_id=21627

After discussion on python-dev, I committed r52738, which
raises a deprecation warning when a float is passed as seek
argument; the meaning of doing so remains unchanged. The
original patch is thus rejected.