polars_readstat

May 29, 2026 · View on GitHub

Polars plugin for SAS (.sas7bdat), Stata (.dta), and SPSS (.sav/.zsav) files.

The Python package wraps the Rust core in polars_readstat_rs and exposes a Polars-first API. The project includes cross-library parity tests and roundtrip checks to reduce regressions.

The Rust engine is generally faster for many workloads, but performance varies by file shape and options. If you need the legacy C/C++ engine, use version 0.11.1 (see the prior version).

Why use this?

In project benchmarks, the new Rust-backed engine is typically faster than pandas/pyreadstat on large SAS/Stata files, especially for subset/filter workloads.
It avoids the older C/C++ toolchain complexity and ships as standard Python wheels.
API is Polars-first (scan_readstat, read_readstat, write_readstat, write_sas_csv_import).
Because scan_readstat returns a Polars LazyFrame, column selection and row limits are pushed into the reader — only the data you actually need is read from disk.

Install

pip install polars-readstat

Core API

1) Lazy scan

import polars as pl
from polars_readstat import scan_readstat

lf = scan_readstat("/path/file.sas7bdat")
#   do something
df = lf.collect()

df = (
    scan_readstat("/path/file.sas7bdat")
    .select(["SERIALNO", "AGEP"])   # column pushdown — only these columns are read
    .head(1_000)                    # row limit is pushed down too
    .filter(pl.col("AGEP") >= 18)   # filters applied to streamed batches to avoid loading full file into memory
    .collect()
)

2) Getting metadata

from polars_readstat import ScanReadstat

reader = ScanReadstat(path="/path/file.sav")
schema = reader.schema           # polars.Schema
metadata = reader.metadata       # dict with file info and per-column details
lf = reader.df                   # LazyFrame — same as calling scan_readstat(path)

metadata is a dict with a variables (SPSS/Stata) or columns (SAS) list. Each entry includes:

"name" — column name
"label" — variable label (description), if present
"value_labels" — dict mapping coded values to label strings, if present

Polars lazy evaluation

scan_readstat returns a LazyFrame, so Polars can push operations into the reader before any data is loaded:

Read only specific columns — column selection is pushed into the reader; unselected columns are never read from disk:

lf = scan_readstat("file.sav")
df = lf.select(["id", "age", "income"]).collect()

Read the first N rows — head() / limit() stops the reader after N rows, so you never load the full file:

df = scan_readstat("file.sas7bdat").head(1000).collect()

Filter rows — filters are applied in Polars after reading, but still benefit from column pushdown if combined with .select():

df = scan_readstat("file.dta").select(["id", "age"]).filter(pl.col("age") >= 18).collect()

The benchmark numbers above reflect these optimizations — the large "Subset: True" speedups come from column pushdown.

3) Write (Experimental)

Writing support is experimental and compatibility varies across tools. Stata roundtrip tests are included; SPSS roundtrip coverage is limited. Please report issues.

from polars_readstat import write_readstat, write_sas_csv_import

write_readstat(df, "/path/out.dta")
write_readstat(df, "/path/out.sav")
write_sas_csv_import(df, "/path/out/sas_bundle", dataset_name="my_data")

write_readstat supports Stata (dta) and SPSS (sav).
Use write_sas_csv_import for SAS-ingestible output (.csv + .sas import script). Binary .sas7bdat writing is not currently supported.

Docs

View the docs at https://jrothbaum.github.io/polars_readstat/ for more information on the options you can pass to the scan and write functions.

Benchmarks compare four scenarios: 1) load the full file, 2) load a subset of columns (Subset:True), 3) filter to a subset of rows (Filter: True), 4) load a subset of columns and filter to a subset of rows (Subset:True, Filter: True).

Benchmark context:

Machine: AMD Ryzen 7 8845HS (16 cores), 14 GiB RAM, Linux Mint 22
Storage: external SSD
Last run: May 14, 2026 — polars-readstat v0.17.0 vs pandas and pyreadstat
Method: wall-clock timings via Python time.time()

Compared to Pandas and Pyreadstat (using read_file_multiprocessing for parallel processing in Pyreadstat)

SAS

all times in seconds (speedup relative to pandas in parenthesis below each)

Library	Full File	Subset: True	Filter: True	Subset: True, Filter: True
polars_readstat	0.55 (3.9×)	0.07 (28.4×)	1.46 (2.0×)	0.08 (39.4×)
pandas	2.16	1.99	2.93	3.15
pyreadstat	6.76 (0.3×)	1.64 (1.2×)	7.86 (0.4×)	2.18 (1.4×)

Stata

all times in seconds (speedup relative to pandas in parenthesis below each)

Library	Full File	Subset: True	Filter: True	Subset: True, Filter: True
polars_readstat	0.16 (7.3×)	0.10 (11.7×)	0.18 (7.3×)	0.09 (13.8×)
pandas	1.17	1.17	1.31	1.24
pyreadstat	5.48 (0.2×)	4.57 (0.3×)	5.67 (0.2×)	7.69 (0.2×)

SPSS

all times in seconds (speedup relative to pandas in parenthesis below each)

Library	Full File	Subset: True	Filter: True	Subset: True, Filter: True
polars_readstat	1.09 (62.5×)	0.15 (3.9×)	1.10 (62.4×)	0.15 (3.9×)
pandas	68.12	0.59	68.67	0.59
pyreadstat	3.06 (22.3×)	1.15 (0.5×)	7.09 (9.7×)	1.23 (0.5×)

zsav

all times in seconds (speedup relative to pandas in parenthesis below each)

Library	Full File	Subset: True	Filter: True	Subset: True, Filter: True
polars_readstat	3.97 (5.9×)	1.04 (2.1×)	4.77 (4.7×)	1.15 (2.0×)
pandas	23.47	2.20	22.40	2.29

Detailed benchmark notes and dataset descriptions are in BENCHMARKS.md.

Tests run

Test coverage includes:

Cross-library comparisons on the pyreadstat and pandas test data, checking results against polars-readstat==0.11.1, pyreadstat, and pandas.
Stata/SPSS read/write roundtrip tests.
Large-file read/write benchmark runs on real-world data (results below).

If you want to run the same checks locally, helper scripts and tests are in scripts/ and tests/.