Re: [h5md-user] increasing order of "time" dataset

[Felix Höfling]

Am 09.05.2014, 15:43 Uhr, schrieb Peter Colberg  
<address@hidden>:

> On Fri, May 09, 2014 at 09:57:44AM +0200, Felix Höfling wrote:
> > It does actually not break backwards compatibility as the difference is
> > relevant only for equality comparisons of floating-point numbers, which  
> > is
> > strictly discouraged. Thus I think the slightly relaxed condition would  
> > make
> > H5MD a bit more flexible.
>
> I noted that it breaks backwards compatibility to make you aware that
> this seemingly minor change requires a bump of the major version number.
> The values of a dataset can, e.g., be stored in a hash table with the
> times as keys, since H5MD 1.0 ensures that the times are unique.
>
> Peter

Uniqueness of floating-point numbers requires floating-point comparison.  
From your own experiences you may remember that such a thing should  
strictly be avoided.

Actually H5MD version 1.0.0 is flawed here: it assumes that floating-point  
numbers can be compared, which however depends on implementation-specific  
rounding modes etc. The truth is that a (sensible) software can not  
distinguish between increasing (= non-decreasing) and strictly increasing  
time datasets. It is not even guaranteed that two slightly different  
floating-point numbers end up as different values in the HDF5 file because  
of possible conversions going on.

I'm aware that the discussion has become a bit esoteric now ...

Felix