ambers

Pure Rust SPSS .sav/.zsav reader and writer — Arrow-native, zero C dependencies.

Features

• Blazing fast read and write for SPSS .sav (bytecode) and .zsav (zlib) files
• Rich metadata: variable labels, value labels, missing values, MR sets, measure levels, and more
• Lazy reader via scan_sav() — Polars LazyFrame with projection and row limit pushdown
• Pure Rust with a native Python API — native Arrow integration, no C dependencies
• Benchmarked up to 3–10x faster reads and 4–20x faster writes compared to current popular SPSS I/O libraries

Installation

Python:

uv add ambers

Rust:

cargo add ambers

Python

import ambers as am
import polars as pl

# Eager read — data + metadata
df, meta = am.read_sav("survey.sav")

# Lazy read — returns Polars LazyFrame
lf, meta = am.scan_sav("survey.sav")
df = lf.select(["Q1", "Q2", "age"]).head(1000).collect()

# Explore metadata
meta.summary()
meta.describe("Q1")
meta.value("Q1")

# Read metadata only (fast, skips data)
meta = am.read_sav_metadata("survey.sav")

# Write back — roundtrip with full metadata
df = df.filter(pl.col("age") > 18)
am.write_sav(df, "filtered.sav", meta=meta)                        # bytecode (default for .sav)
am.write_sav(df, "compressed.zsav", meta=meta)                     # zlib (default for .zsav)
am.write_sav(df, "raw.sav", meta=meta, compression="uncompressed") # no compression
am.write_sav(df, "fast.zsav", meta=meta, compression_level=1)      # fast zlib

# From scratch — metadata is optional, inferred from DataFrame schema
am.write_sav(df, "new.sav")

.sav uses bytecode compression by default, .zsav uses zlib. Pass compression= to override ("uncompressed", "bytecode", "zlib"). Pass meta= to preserve all metadata from a prior read_sav(), or omit it to infer formats from the DataFrame.

Rust

use ambers::{read_sav, read_sav_metadata};

// Read data + metadata
let (batch, meta) = read_sav("survey.sav")?;
println!("{} rows, {} cols", batch.num_rows(), meta.number_columns);

// Read metadata only
let meta = read_sav_metadata("survey.sav")?;
println!("{}", meta.label("Q1").unwrap_or("(no label)"));

Metadata API (Python)

Method	Description
meta.summary()	Formatted overview: file info, type distribution, annotations
meta.describe("Q1")	Deep-dive into a single variable (or list of variables)
meta.diff(other)	Compare two metadata objects, returns MetaDiff
meta.label("Q1")	Variable label
meta.value("Q1")	Value labels dict
meta.format("Q1")	SPSS format string (e.g. "F8.2", "A50")
meta.measure("Q1")	Measurement level ("nominal", "ordinal", "scale")
meta.role("Q1")	Variable role ("input", "target", "both", "none", "partition", "split")
meta.attribute("Q1", "CustomNote")	Custom attribute values (list[str] or None)
meta.schema	Full metadata as a nested Python dict

All variable-name methods raise KeyError for unknown variables.

Metadata Fields

All fields returned by the reader. Fields marked Write are preserved when passed via meta= to write_sav(). Read-only fields are set automatically (encoding, timestamps, row/column counts, etc.).

Note: This is a first pass — field names and behavior may change without warning in future releases.

Field	Read	Write	Type
file_label	yes	yes	str
file_format	yes	—	str
file_encoding	yes	—	str
creation_time	yes	—	str
compression	yes	—	str
number_columns	yes	—	int
number_rows	yes	—	int | None
weight_variable	yes	yes	str | None
notes	yes	yes	list[str]
variable_names	yes	—	list[str]
variable_labels	yes	yes	dict[str, str]
variable_value_labels	yes	yes	dict[str, dict[float|str, str]]
variable_formats	yes	yes	dict[str, str]
variable_measures	yes	yes	dict[str, str]
variable_alignments	yes	yes	dict[str, str]
variable_storage_widths	yes	—	dict[str, int]
variable_display_widths	yes	yes	dict[str, int]
variable_roles	yes	yes	dict[str, str]
variable_missing_values	yes	yes	dict[str, dict]
variable_attributes	yes	yes	dict[str, dict[str, list[str]]]
mr_sets	yes	yes	dict[str, dict]
arrow_data_types	yes	—	dict[str, str]

Creating metadata from scratch:

meta = am.SpssMetadata(
    file_label="Customer Survey 2026",
    variable_labels={"Q1": "Satisfaction", "Q2": "Loyalty"},
    variable_value_labels={"Q1": {1: "Low", 5: "High"}},
    variable_measures={"Q1": "ordinal", "Q2": "nominal"},
)
am.write_sav(df, "output.sav", meta=meta)

Modifying existing metadata (from read_sav() or a previously created SpssMetadata):

# .update() — bulk update multiple fields at once, merges dicts, replaces scalars
meta2 = meta.update(
    file_label="Updated Survey",
    variable_labels={"Q3": "NPS"},        # Q1/Q2 labels preserved, Q3 added
    variable_measures={"Q3": "scale"},
)

# .with_*() — chainable single-field setters, with full IDE autocomplete and type hints
meta3 = (meta
    .with_file_label("Updated Survey")
    .with_variable_labels({"Q3": "NPS"})
    .with_variable_measures({"Q3": "scale"})
)

Immutability: SpssMetadata is immutable. .update() and .with_*() always return a new instance — the original is never modified. Assign to a new variable if you need to keep both copies.

Update logic:

• Dict fields (labels, formats, measures, etc.) merge as an overlay — new keys are added, existing keys are overwritten, all other keys are preserved. Pass {key: None} to remove a key.
• Scalar fields (file_label, weight_variable) and notes are replaced entirely.
• Column renames are not tracked. If you rename "Q1" to "Q1a" in your DataFrame, metadata for "Q1" does not carry over — you must explicitly provide metadata for "Q1a".

See metadata.md for the full API reference including update logic details, missing values, MR sets, and validation rules.

SPSS tip: Custom variable attributes are not shown in SPSS's Variable View by default. Go to View > Customize Variable View and click OK, or run DISPLAY ATTRIBUTES in SPSS syntax.

Streaming Reader (Rust)

let mut scanner = ambers::scan_sav("survey.sav")?;
scanner.select(&["age", "gender"])?;
scanner.limit(1000);

while let Some(batch) = scanner.next_batch()? {
    println!("Batch: {} rows", batch.num_rows());
}

Performance

Eager Read

All results return a Polars DataFrame. Best of 3–5 runs (with warmup) on Windows 11, Python 3.13, Intel Core Ultra 9 275HX (24C), 64 GB RAM (6400 MT/s).

File	Size	Rows	Cols	ambers	polars_readstat	pyreadstat	vs prs	vs pyreadstat
test_1 (bytecode)	0.2 MB	1,500	75	< 0.01s	< 0.01s	0.011s	—	—
test_2 (bytecode)	147 MB	22,070	677	0.286s	0.897s	3.524s	3.1x	12x
test_3 (uncompressed)	1.1 GB	79,066	915	0.322s	1.150s	4.918s	3.6x	15x
test_4 (uncompressed)	0.6 MB	201	158	0.002s	0.003s	0.012s	1.5x	6x
test_5 (uncompressed)	0.6 MB	203	136	0.002s	0.003s	0.016s	1.5x	8x
test_6 (uncompressed)	5.4 GB	395,330	916	1.600s	1.752s	25.214s	1.1x	16x

• Faster than polars_readstat on all tested files — 1.1–3.6x faster
• 6–16x faster than pyreadstat across all file sizes
• No PyArrow dependency — uses Arrow PyCapsule Interface for zero-copy transfer

Lazy Read with Pushdown

scan_sav() returns a Polars LazyFrame. Unlike eager reads, it only reads the data you ask for:

File (size)	Full collect	Select 5 cols	Head 1000 rows	Select 5 + head 1000
test_2 (147 MB, 22K × 677)	0.903s	0.363s (2.5x)	0.181s (5.0x)	0.157s (5.7x)
test_3 (1.1 GB, 79K × 915)	0.700s	0.554s (1.3x)	0.020s (35x)	0.012s (58x)
test_6 (5.4 GB, 395K × 916)	3.062s	2.343s (1.3x)	0.022s (139x)	0.013s (236x)

On the 5.4 GB file, selecting 5 columns and 1000 rows completes in 13ms — 236x faster than reading the full dataset.

Write

write_sav() writes a Polars DataFrame + metadata back to .sav (bytecode) or .zsav (zlib). Best of 5 runs on the same machine.

File	Size	Rows	Cols	Mode	ambers	pyreadstat	Speedup
test_1 (bytecode)	0.2 MB	1,500	75	.sav	0.001s	0.019s	13x
				.zsav	0.004s	0.025s	6x
test_2 (bytecode)	147 MB	22,070	677	.sav	0.539s	3.622s	7x
				.zsav	0.386s	4.174s	11x
test_3 (uncompressed)	1.1 GB	79,066	915	.sav	0.439s	13.963s	32x
				.zsav	0.436s	17.991s	41x
test_4 (uncompressed)	0.6 MB	201	158	.sav	0.002s	0.027s	16x
				.zsav	0.004s	0.035s	9x
test_5 (uncompressed)	0.6 MB	203	136	.sav	0.001s	0.023s	17x
				.zsav	0.003s	0.027s	9x
test_6 (uncompressed)	5.4 GB	395,330	916	.sav	2.511s	84.836s	34x
				.zsav	2.255s	90.499s	40x

• 6–41x faster than pyreadstat on writes across all files and compression modes
• Full metadata roundtrip: variable labels, value labels, missing values, MR sets, display properties
• Bytecode (.sav) and zlib (.zsav) compression

Roadmap

• Continued I/O performance optimization
• Expanded SPSS metadata field coverage
• Rich metadata manipulation — add, update, merge, and remove metadata programmatically
• Individual metadata field overrides in write_sav() — pass variable_labels=, variable_value_labels=, etc. alongside meta= to selectively override fields
• Currently supports read and write with Polars DataFrames (eager and lazy) — extending to pandas, Narwhals, DuckDB, and others

License

MIT