This is interesting, but it sounds like it corresponds more to the rle encoding that we do rather than the generic compression code.
Has anyone done a Java version of the library? It is faster to iterate on this kind of design in Java. On the other hand, I’ve heard that someone is thinking about doing a Rust ORC reader & writer, but it isn’t done yet. 😊 .. Owen > On May 19, 2022, at 17:16, Dongjoon Hyun <dongjoon.h...@gmail.com> wrote: > > Thank you for sharing, Martin. > > For codec one, you can take advantage of our benchmark suite > (NONE/ZLIB/SNAPPY/ZSTD) to show the benefits in ORC format. > > https://github.com/apache/orc/blob/main/java/bench/core/src/java/org/apache/orc/bench/core/CompressionKind.java#L34-L37 > > For the Spark connector one, I'd like to recommend you to send dev@spark > too. You will get attention in both parts (codec and connector). > > Then, I'm looking forward to seeing your benchmark result. > > Dongjoon. > > >> On Thu, May 19, 2022 at 4:12 PM Martin Loncaric <m.w.lonca...@gmail.com> >> wrote: >> >> I've developed a stable codec for numerical columns called Quantile >> Compression <https://github.com/mwlon/quantile-compression>. >> It has about 30% higher compression ratio than even Zstd for similar >> compression and decompression time. It achieves this by tailoring to the >> data type (floats, ints, timestamps, bools). >> >> I'm using it in my own projects, and a few others have adopted it, but it >> would also be perfect for ORC columns. Assuming a 50-50 split between >> text-like and numerical data, it could reduce the average ORC file size by >> over 10% with no extra compute cost. Incorporating it into ORC would be >> quite powerful since the codec by itself only works on a single flat column >> of non-nullable numbers. >> >> Would the ORC community be interested in this? How can we make this >> available to users? I've already built a Spark connector >> <https://github.com/pancake-db/spark-pancake-connector> for a project >> using >> this codec and gotten fast query times. >> >> Thanks, >> Martin >>
