maskops 0.1.5

High-speed PII masking as a Polars plugin — powered by Rust

MaskOps

High-speed PII masking as a native Polars plugin — powered by Rust.

MaskOps extends Polars with zero-overhead PII detection and masking expressions. No NLP models. No intermediate files. Just regex + Rust running directly on Arrow buffers.

How It Works

flowchart LR
    A[🐍 Python\nPolars DataFrame] -->|mask_pii / contains_pii| B[Polars\nExpression Engine]
    B -->|Arrow buffer\nzero-copy| C[🦀 Rust Core\nmaskops]
    C -->|IBAN / VAT| D[Masked\nSeries]
    C -->|Email / Phone| D
    C -->|IP / Credit Card| D
    C -->|DNI / NIE / NIN| D
    C -->|Personalausweis| D
    C -->|RUT / CPF / CURP| D
    D -->|back to Python| A
    style A fill:#306998,color:#fff
    style C fill:#CE422B,color:#fff
    style B fill:#2E2E2E,color:#fff
    style D fill:#2E7D32,color:#fff

No Python objects created per row. No NLP model loaded. No intermediate files.

• Presidio is heavy — it spins up NLP models for structured CSV data that doesn't need them.
• Pure Python regex on large DataFrames is slow.
• MaskOps compiles to a native .so that Polars calls directly — same speed as built-in expressions.

Architecture

maskops/
├── Cargo.toml               # Rust dependencies
├── pyproject.toml           # maturin build backend + PyPI metadata
├── src/
│   ├── lib.rs               # Polars expression registration (mask_pii, contains_pii)
│   └── patterns/
│       ├── mod.rs           # mask_all() and contains_any_pii() aggregators
│       ├── iban.rs          # IBAN regex + masking
│       ├── vat.rs           # EU VAT regex + masking
│       ├── email.rs         # Email regex + masking (local part)
│       ├── phone.rs         # E.164 phone regex + masking
│       ├── ip.rs            # IPv4/IPv6 regex + masking
│       ├── latam_id.rs      # RUT (Chile), CPF (Brazil), CURP (Mexico)
│       ├── european_id.rs   # DNI/NIE (Spain), NIN (UK), Personalausweis (Germany)
│       ├── credit_card.rs   # Visa, Mastercard, Amex, Discover, Maestro + Luhn
│       └── country_codes.rs # Country prefix lookup table
├── maskops/
│   └── __init__.py          # Python API via register_plugin_function
├── benchmarks/
│   └── benchmark.py         # Per-family throughput benchmarks (1M rows)
└── tests/
    ├── test_masking.py      # pytest suite (66 tests)
    ├── generate_fixtures.py # Faker-based EU test data generator
    └── fixtures/            # Generated CSVs (gitignored)

The Rust layer operates directly on Arrow buffers — zero Python object overhead per row. Each PII type is its own module: adding a new pattern = new file + one line in mod.rs.

Install

pip install maskops

Usage

import polars as pl
import maskops

df = pl.read_csv("payments.csv")

# Mask all PII in a column
df.with_columns(maskops.mask_pii("notes"))

# Filter rows that contain PII
df.filter(maskops.contains_pii("free_text"))

Supported patterns (v0.1.4)

Pattern	Example input	Masked output
IBAN	DE89370400440532013000	DE89******************
EU VAT	DE123456789	DE*********
Email	[e-mail]	********@example.com
Phone	[phone]	+1**********
IP Address	192.168.1.100	192.168.*.*
RUT (Chile)	76.354.771-K	**********-K
CPF (Brazil)	529.982.247-25	*********-25
CURP (Mexico)	BADD110313HCMLNS09	******************
DNI (Spain)	12345678Z	********Z
NIE (Spain)	X1234567L	********L
NIN (UK)	AB 12 34 56 C	*********** C
Personalausweis (Germany)	T220001293	**********
Credit Card (Visa/MC/Amex/Discover/Maestro)	4111111111111111	411111******1111

Tested against 8 EU locales: DE, FR, ES, IT, NL, PL, PT, SE. Email and phone follow RFC 5322 and E.164 respectively. RUT and CPF include Módulo 11 check digit validation. DNI and NIE include modulo 23 check letter validation. Credit cards include Luhn validation — format-only matches are rejected. Personalausweis and NIN: format-only matching; check digit validation pending (v0.2.0+).

Roadmap

• Email, phone patterns
• IP address patterns
• Latin American IDs (RUT, CPF, CURP)
• European IDs (DNI/NIE Spain, NIN UK, Personalausweis Germany)
• Credit cards (Visa, Mastercard, Amex, Discover, Maestro) with Luhn validation
• PyPI publish via GitHub Actions
• Check digit validation for Personalausweis (Germany) and NIN (UK)
• Format-Preserving Encryption (FPE/FF3-1) for reversible masking
• Benchmark improvement vs pure Python
• Benchmark vs Presidio
• Parquet streaming support

Build from source

Windows (PowerShell)

python -m venv .venv
.venv\Scripts\activate
pip install maturin faker polars pytest
maturin develop --release
python tests/generate_fixtures.py
pytest tests/ -v

Linux / macOS

python -m venv .venv
source .venv/bin/activate
pip install maturin faker polars pytest
maturin develop --release
python tests/generate_fixtures.py
pytest tests/ -v

Key dependency versions

Package	Version
pyo3	0.21
pyo3-polars	0.18
polars	0.46
maturin	>=1.7,<2.0

Note: pyo3 must be 0.21 to match pyo3-polars 0.18. Do not bump pyo3 independently.

License

MIT

Benchmarks

Tested on 1,000,000 rows, Intel i-series CPU, Python 3.14, Windows.

Median of 3 runs per benchmark. Baseline uses equivalent regex coverage to maskops per family.

Note on per-family benchmarks: maskops always runs the full pattern set — there is no per-family dispatch. A "Credit Card only" benchmark still pays for IBAN, VAT, email, phone, LatAm ID, and EU ID checks. The Python baseline only runs one regex. This is why maskops underperforms on isolated families with dense PII. The advantage emerges when all patterns are active simultaneously, which is the realistic production case.

EU patterns (IBAN, VAT, Email, Phone)

Profile	Expression	Time	Rows/s	Python re	Speedup
clean	mask_pii	2.455s	407,300	4.268s	1.7×
clean	contains_pii	1.184s	844,846	—	—
dense	mask_pii	3.184s	314,093	1.784s	0.6×
dense	contains_pii	0.133s	7,497,325	—	—
mixed	mask_pii	2.943s	339,774	1.993s	0.7×
mixed	contains_pii	0.282s	3,551,833	—	—

LatAm patterns (RUT, CPF, CURP)

Profile	Expression	Time	Rows/s	Python re	Speedup
clean	mask_pii	2.276s	439,367	2.319s	1.0×
clean	contains_pii	0.795s	1,258,169	—	—
dense	mask_pii	3.048s	328,080	1.690s	0.6×
dense	contains_pii	0.640s	1,562,313	—	—
mixed	mask_pii	2.880s	347,173	1.854s	0.6×
mixed	contains_pii	0.705s	1,418,784	—	—

RUT and CPF include Módulo 11 check digit validation per row — this is the cost of zero false positives.

Network patterns (IP)

Profile	Expression	Time	Rows/s	Python re	Speedup
clean	mask_pii	2.301s	434,502	2.093s	0.9×
clean	contains_pii	0.799s	1,251,735	—	—
dense	mask_pii	2.509s	398,628	1.553s	0.6×
dense	contains_pii	0.215s	4,655,272	—	—
mixed	mask_pii	2.504s	399,408	1.684s	0.7×
mixed	contains_pii	0.374s	2,671,550	—	—

Credit card patterns (Visa, Mastercard, Amex, Discover, Maestro)

Profile	Expression	Time	Rows/s	Python re	Speedup
clean	mask_pii	2.243s	445,762	0.954s	0.4×
clean	contains_pii	0.792s	1,261,873	—	—
dense	mask_pii	2.797s	357,473	1.005s	0.4×
dense	contains_pii	0.628s	1,591,805	—	—
mixed	mask_pii	2.687s	372,166	1.014s	0.4×
mixed	contains_pii	0.674s	1,484,572	—	—

Luhn validation runs per candidate match — this eliminates false positives at the cost of single-family throughput.

European ID patterns (DNI/NIE, NIN, Personalausweis)

Profile	Expression	Time	Rows/s	Python re	Speedup
clean	mask_pii	2.282s	438,149	1.410s	0.6×
clean	contains_pii	0.801s	1,248,547	—	—
dense	mask_pii	2.609s	383,334	1.107s	0.4×
dense	contains_pii	0.604s	1,654,937	—	—
mixed	mask_pii	2.590s	386,037	1.179s	0.5×
mixed	contains_pii	0.665s	1,504,806	—	—

All patterns active

This is the realistic production workload — all 15 pattern types running simultaneously. maskops is up to 5.7× faster than an equivalent pure Python approach. contains_pii reaches 1.9M rows/s on mixed data — use it to pre-filter before masking in hot pipelines.

Profile	Expression	maskops	Python re	Speedup
clean	mask_pii	2.344s	13.445s	5.7×
clean	contains_pii	0.822s	—	—
dense	mask_pii	3.269s	6.625s	2.0×
dense	contains_pii	0.520s	—	—
mixed	mask_pii	3.285s	6.581s	2.0×
mixed	contains_pii	0.545s	—	—

maskops throughput stays roughly flat as pattern count grows — Python regex degrades with each additional pattern. The clean profile gap (5.7×) reflects Python's overhead of compiling and scanning a large combined regex on short-circuit misses.

vs Microsoft Presidio (estimated)

Presidio processes structured DataFrames via presidio-structured, which runs a spaCy NLP pipeline per row. Based on community reports and the architecture:

Tool	Throughput (structured data)	Requires NLP model
maskops	~305K–7.5M rows/s (measured)	No
Presidio (regex-only recognizers)	~10–50K rows/s*	No
Presidio (spaCy NER)	~1–5K rows/s*	Yes (250MB+)

* Estimated from community benchmarks and Presidio's own documentation noting it is "not optimized for bulk structured data." Microsoft confirmed no official throughput benchmarks exist.

maskops is purpose-built for structured data pipelines where Presidio's NLP overhead is unnecessary.

---

This project was developed with AI assistance from Claude (Anthropic). All architecture decisions, security properties, and code were reviewed and validated by the author.