> > Pco could technically work as a Parquet encoding, but people are wary of > its newness and weak FFI support. It seems there is no immediate action to > take, but would be worthwhile to consider this again further in the future.
I guess I'm more optimistic on the potential gaps. I think if there were a spec that allowed one to code it from scratch, I'd be willing to take a crack at seeing what it would take for another implementation in either Java or C++. (I looked at the links you provided but they were somewhat too high-level). I think having a spec would also guard against the "newness" concern. I can't say there wouldn't be other technical blockers but at least this would be someplace to start? Cheers, Micah On Thu, Jan 11, 2024 at 7:21 PM Martin Loncaric <m.w.lonca...@gmail.com> wrote: > (Oops, the repeating binary decimal is 1100... with period 4, so exactly 2 > bits of entropy for the 52 mantissa bits. The argument is the same though.) > > On Thu, Jan 11, 2024 at 10:02 PM Martin Loncaric <m.w.lonca...@gmail.com> > wrote: > > > To reach a conclusion on this thread, I understand the overall sentiment > > as: > > > > Pco could technically work as a Parquet encoding, but people are wary of > > its newness and weak FFI support. It seems there is no immediate action > to > > take, but would be worthwhile to consider this again further in the > future. > > > > On Thu, Jan 11, 2024 at 9:47 PM Martin Loncaric <m.w.lonca...@gmail.com> > > wrote: > > > >> I must admit I'm a bit surprised by these results. The first thing is > >>> that the Pcodec results were actually obtained using dictionary > >>> encoding. Then I don't understand what is Pcodec-encoded: the > dictionary > >>> values or the dictionary indices? > >> > >> > >> No, pco cannot be dictionary encoded; it only goes from vec<T> -> Bytes > >> and back. Some of Parquet's existing encodings are like this as well. > >> > >> The second thing is that the BYTE_STREAM_SPLIT + Zstd results are much > >>> worse than the PLAIN + Zstd results, which is unexpected (though not > >>> impossible). > >> > >> > >> I explained briefly in the blog post, but BYTE_STREAM_SPLIT does > terribly > >> for this data because there is high correlation among each number's > bytes. > >> For instance, if each double is a multiple of 0.1, then the 52 mantissa > >> bits will look like 011011011011011... (011 repeating). That means there > >> are only 3 possibilities (<2 bits of entropy) for the last 6+ bytes of > each > >> number. BYTE_STREAM_SPLIT throws this away, requiring 6+ times as many > bits > >> for them. > >> > >> On Mon, Jan 8, 2024 at 10:44 AM Antoine Pitrou <anto...@python.org> > >> wrote: > >> > >>> > >>> Hello Martin, > >>> > >>> On Sat, 6 Jan 2024 17:09:07 -0500 > >>> Martin Loncaric <m.w.lonca...@gmail.com> > >>> wrote: > >>> > > > >>> > > It would be very interesting to expand the comparison against > >>> > > BYTE_STREAM_SPLIT + compression. > >>> > > >>> > Antoine: I created one now, at the bottom of the post > >>> > <https://graphallthethings.com/posts/the-parquet-we-could-have>. In > >>> this > >>> > case, BYTE_STREAM_SPLIT did worse. > >>> > >>> I must admit I'm a bit surprised by these results. The first thing is > >>> that the Pcodec results were actually obtained using dictionary > >>> encoding. Then I don't understand what is Pcodec-encoded: the > dictionary > >>> values or the dictionary indices? > >>> > >>> The second thing is that the BYTE_STREAM_SPLIT + Zstd results are much > >>> worse than the PLAIN + Zstd results, which is unexpected (though not > >>> impossible). > >>> > >>> Regards > >>> > >>> Antoine. > >>> > >>> > >>> >