pcodec VS ryg_rans

Context: I've been iterating on my side project pcodec (a codec for columns of numerical data) and have gradually improved decompression speed from ~150MB/s to ~1GB/s. Not everything here is novel or Rust-specific, but here's what I've learned in the process:

For context: I'm creating an experimental successor to my library Quantile Compression, which does good compression for numerical sequences and has several users. I have a variable number of symbols which may be as high as 212 in some cases, but is ~26 in most cases. The data is typically 216 to 224 tokens long.

I'm not a member, but you can use the CLI to try it out pretty easily: https://github.com/mwlon/quantile-compression/tree/main/q_compress_cli . Let me know how it does

You can try it out very easily with the CLI which works on CSV and Parquet columns now, e.g. cargo run --release compress --csv my.csv --col-name my_column out.qco

Right, please don't try to use it for general files. It looks like zpaq is kinda hard to set up except on windows, so I'm probably not going to, but I encourage you to try it out! There's an example you can use to generate a bunch of random numerical distributions, outputting binary files, .qco, and other formats.

Here's how you can generate benchmark data, including binary files: https://github.com/mwlon/quantile-compression/blob/main/q_compress/examples/primary.md

I made a simple CLI for compressing and inspecting .qco files. Not available on package managers yet, but it's still pretty easy to try out: https://github.com/mwlon/quantile-compression/blob/main/CLI.md

RANS uses compact table -- at the cost of speed. https://github.com/rygorous/ryg_rans Unfortunately, there are bugs in the repository, you should not use the code as is. It is better to rewrite the code -- it is quite simple when you understand it. Author's blog (http://fgiesen.wordpress.com/) contains a lot of details and is very interesting.