Polypress

A lossless compressor for data tables: CSV, TSV and other delimited text, JSON, JSON Lines, Parquet, and Stata, SPSS and SAS files.

The table comes back exactly: the same columns in the same order, the same rows in the same order, and every cell as the exact text it was. Before writing an archive, Polypress decodes it and compares every cell.

Why it is smaller

General-purpose compressors see a table as a stream of bytes. Columnar formats compress each column on its own. Polypress uses the fact that the columns of a real table depend on each other.

• Rows are reordered by a parent column. City is nearly determined by Postal Code. Sorting the rows by postal code puts equal cities next to each other, and the city column collapses into long runs. The order costs nothing to store: the decoder rebuilds the parent column first and repeats the same sort. Rows come back in their original order.
• Numbers are predicted from the values above them and from comparable neighbouring columns, and only the error is stored.
• Copied values are stored once. A number that also appears inside another column's text (latitude inside location) is not stored twice.
• The modelled columns are finished with xz.

The core idea is not new. US 8,312,026 B2 (Vo, AT&T, filed 2009) describes it. It was arrived at here independently, and the patent has expired.

Results

Measured on 1,000 government open-data tables chosen by a rule, not by hand: the Socrata catalog's datasets in order of page views, keeping the first 1,000 that download as CSV and have at least 2 columns, 20 rows and 50 KB, with no duplicates. That is 7.2 GB of CSV and 27.9 million rows. The 189 largest tables were cut at a fixed size, and every tool got the same cut file.

Polypress 1.00 was compared with 22 other tools: xz, zstd, gzip, bzip2, brotli, lz4, zip, 7-Zip (LZMA2 and PPMd), and Parquet, ORC and Feather with their codecs.

Restored exactly	1,000 of 1,000
Smallest of all 23 tools	781 of 1,000 (78%)
All 1,000 together, against the best other tool on each table	418 MB vs 486 MB (1.16x smaller)
Compression of the CSV	median 16x

Head to head, Polypress is smaller than xz -9e on 990 of 1,000 tables, zstd -19 on 997, gzip -9 on all 1,000, and parquet+zstd on 899 of 900.

Where it loses. It is not the smallest on 219 tables, mostly text-heavy ones, and mostly to 7-Zip's PPMd. The worst result is 54% bigger than orc+zstd on one table.

Speed. On an Apple M4 MacBook Air, it compresses at about 18 MB/s and decompresses at about 150 MB/s (xz -9e: 4 and 199). That is one timing pass, so read it as an order of magnitude.

Parquet and exact text. Parquet, read with type inference, changed the printed text on 702 of the tables (1.50 became 1.5). Part of a Parquet size win there is lost formatting, not compression.

Every table in the benchmark is government open data, so there is no reason to assume the results carry over to other kinds of data. Nobody outside the project has used it yet.

Install

pip install polypress               # the polypress command
pip install "polypress[parquet]"    # also installs pyarrow, for Parquet files

Ready-built for macOS on Apple silicon and for Linux (x86-64 and ARM). Anywhere else, pip compiles it from source, which needs a C compiler. The package is the C program and a small launcher; there is no second copy of the compressor in Python.

Build from source

You need a C compiler. The xz and ReadStat sources are included, so nothing else has to be installed.

sh work/csrc/build.sh                              # -> work/csrc/polypress
PREFIX=/usr/local sh work/csrc/build.sh install    # put it on your PATH

• Mac app: sh work/app/build_app.sh builds Polypress.app into ~/Applications (macOS 11 or later, Xcode command line tools).
• Parquet goes through a converter that needs Python 3 and pyarrow: polypress-parquet when installed with pip, work/py/parquet.py from source.

Use

polypress compress data.csv              # -> data.csv.ppz
polypress restore  data.csv.ppz          # -> data.csv
polypress info     data.csv.ppz          # what is inside
polypress convert  data.csv data.json    # one table format to another
polypress stream-compress big.csv        # for files bigger than memory

polypress-parquet to-csv data.parquet | polypress compress - -o data.ppz
polypress restore data.ppz -o - | polypress-parquet from-csv - data.parquet

• Exactly means the table. Column names, order, rows and every cell come back exactly. A CSV file may not come back byte for byte, because quoting and line endings can be written more than one way.
• Stata, SPSS and SAS files are kept byte for byte, labels included. When one is converted to another format, Polypress says what the conversion leaves out.
• Text encoding is never guessed. UTF-8, or a file with a byte-order mark, is read as such. Anything else is refused with the bad byte named. Name the encoding with --encoding latin-1 (or cp1252, utf-16, ...).

Licence

MIT. The bundled xz (0BSD) and ReadStat (MIT) keep their own licences, in their folders under work/csrc/.