3 ms·
One notable omission from this piece is a technique to compress integer time series with both positive and negative values. If you naively apply bit-packing us
by naturalplasmoid 5y ago
One notable omission from this piece is a technique to compress integer time series with both positive and negative values.
If you naively apply bit-packing using the Simple8b algorithm, you'll find that negative integers are not compressed. This is due to how signed integers are represented in modern computers: negative integers will have their most significant bit set [1].
Zigzag encoding is a neat transform that circumvents this issue. It works by mapping signed integers to unsigned integers so that numbers with a small absolute value can be encoded using a small number of bits. Put another way, it encodes negative numbers using the least significant bit for sign. [2]
If you're looking for a quick way to experiment with various time series compression algorithm I highly recommend Daniel Lemire's FastPFor repository [3] (as linked in the article). I've used the Python bindings [4] to quickly evaluate various compression algorithms with great success.
Finally I'd like to humbly mention my own tiny contribution [5], an adaptation of Lemire's C++ Simple8b implementation (including basic methods for delta & zigzag encoding/decoding).
I used C++ templates to make the encoding and decoding routines generic over integer bit-width, which expands support up to 64 bit integers, and offers efficient usage with smaller integers (eg 16 bit). I made a couple other minor tweaks including support for arrays up to 2^64 in length, and tweaking the API/method signatures so they can be used in a more functional style. This implementation is slightly simpler to invoke via FFI, and I intend to add examples showing how to compile for usage via JS (WebAssembly), Python, and C#. I threw my code up quickly in order to share with you all, hopefully someone finds it useful. I intend to expand on usage examples/test cases/etc, and am looking forward to any comments or contributions.
[1] https://en.wikipedia.org/wiki/Signed_number_representation https://en.wikipedia.org/wiki/Signed_number_representation
[2] https://en.wikipedia.org/wiki/Variable-length_quantity#Zigzag_encoding https://en.wikipedia.org/wiki/Variable-length_quantity#Zigza...
[3] https://github.com/lemire/FastPFor https://github.com/lemire/FastPFor
[4] https://github.com/searchivarius/PyFastPFor https://github.com/searchivarius/PyFastPFor
[5] https://github.com/naturalplasmoid/simple8b-timeseries-compression https://github.com/naturalplasmoid/simple8b-timeseries-compr...
- jart 5y agoI love zigzag encoding. The way it uses shift and xor is beautiful. It gives it a performance edge of 3 cycles on my cpu versus signed-leb encoding. One thing I did once, to encode UNICODE lookup tables in a smaller form (to reduce Python binary size) is I found out it can be very profitable to (1) delta encode; then (2) zig zag encode it; and finally (3) apply deflate. For example: _PyUnicode_PhrasebookOffset2: size is 178,176 bytes _PyUnicode_PhrasebookOffset2: packed size is 100,224 bytes _PyUnicode_PhrasebookOffset2: rle size is 282,216 bytes _PyUnicode_PhrasebookOffset2: deflate size is 52,200 bytes _PyUnicode_PhrasebookOffset2: bz2 size is 76,876 bytes _PyUnicode_PhrasebookOffset2: δleb size is 47,198 bytes _PyUnicode_PhrasebookOffset2: δzd size is 12,748 bytes It's so cool.
- lucb1e 5y agoFor anyone else not familiar: δ is a lowercase delta (at least, I had to look it up)