Multi-threaded writing of output files

[gessulat]

For very large datasets, single-threaded IO operations are currently a speed bottleneck.
Pyarrow datasets natively support:

• partitioning a dataframe
• multi-threaded read
• multi-threaded write
• specifying number of cores used for parallel operations - personally, I prefer setting such values via environment variables

Tasks

• read partitioned input data
• write to partitioned output data

[gessulat]

This issue is related to the upgrade to polars #89
Optimizing reading and writing could be done independently, though.

[gessulat]

To motivate this @sambenfredj and I did some benchmarks with a 3M PSM Mokapot input file (tab/csv) and converted it to parquet with different reader and writer implementations. Note, there are several ways how to read and write parquet files. You can write parquet files with different compression algorithms and with different reader and writer implementations: pandas, pyarrow, and polars. Within pyarrow there are again multiple options to read and write parquet files. That's why the read_speed plot is confusing, but I didn't have the time to clean it up - sorry!

Speed is always in seconds.

TL;DR:

• using polars read and write implementation for parquet files with lz4 compression seems to be the optimal choice for performance
• if possible: directly read row groups

file sizes

• zstd offers best compression
• lz4 offers second best compression and is probably acceptable, given that lz4 yields better read/write performance.

read speed

• polars reader implementations are faster than the pyarrow implementation
• if we can read row groups directly (e.g. in a streaming setting) this would be the preferred way to do it.
• lz4 compression is offering the best read speed

write speed

• polars write implementation is faster than pyarrow
• surprisingly lz4 compressed writing is faster than non-compressed writing oO. This finding is consistent for both polars and pyarrow implementations.