That sounds amazing.

On Mon, Aug 25, 2014 at 1:18 PM, Tim Holy <[email protected]> wrote:

> I'd like to announce the merger of a massive revamping of HDF5 & JLD. The
> bottom line is that now many immutables and types will be saved in a far
> more
> efficient fashion to disk, particularly for arrays of immutables. In such
> cases,
> the performance gains are quite extraordinary, often hundred-fold or
> larger in
> terms of time, and the resulting files are much smaller. For many
> situations,
> HDF5 is now comparable to the serializer (and sometimes faster) for reading
> from and writing to disk.
>
> In addition to the performance enhancement, the whole system in JLD for
> saving
> and loading Julia types has been given a facelift. Rather than barfing on
> broken or missing types, there's a sophisticated method to reconstruct
> types
> based on data in the disk file. You might appreciate this if you save data,
> change your type definitions in your code, and then want the data back
> again.
> While these types won't allow you to proceed entirely as if nothing is
> wrong,
> it does give you an easier path to recovery.
>
> Finally, all this has been done with an effort to preserve backwards
> compatibility. All tests pass with the new format (and many new tests have
> been added), and the JLDArchives tests pass, suggesting that at least for
> the
> files checked into JLDArchives, there's been no hit to the ability to read
> old
> files.
>
> Before tagging a new version, we're hoping that a few hardy souls will
> experiment by checking out master for the HDF5 package. Testing is needed
> because we're talking about a long-term data storage format here. Once this
> becomes "official" (by tagging a new version), I'll deposit a copy of a
> test file
> in JLDArchives, so that format becomes something we have to support going
> forward. It would be much nicer to uncover any problems quickly so we don't
> have to worry about ugly kludges :-).
>
> My role in this has exclusively been that of cheerleader: the work was
> done by
> Simon Kornblith and incorporated some great preliminary work by Matt
> Baumann.
> This is a 30-commit, 2,000 change, so it's a huge effort and very
> beautifully
> done. My sincere thanks go to them!
>
> --Tim
>
>

Reply via email to