arnio 1.6.1

C++ accelerated data preparation for pandas and the Python data stack

Fast data preparation for the Python data stack.

Arnio is a compiled C++ data preparation engine for messy CSV and pandas workflows.
It parses, infers types, strips whitespace, deduplicates, validates, and profiles data —
then hands clean results back to the tools you already use.
Use Arnio before and alongside pandas, NumPy, scikit-learn, DuckDB, and Arrow.

           

pip install arnio

Colab install smoke test: COLAB_SMOKE_TEST.md

Quickstart · Integrations · Why Arnio · Architecture · Benchmarks · Community · Contribute

---

⚡ Quickstart

Three lines. That's the entire workflow.

import arnio as ar

# Load CSV directly through C++ — no Python parsing overhead
frame = ar.read_csv("messy_sales_data.csv")

# Declare what clean data looks like — arnio handles the rest
clean = ar.pipeline(frame, [
    ("strip_whitespace",),
    ("normalize_case", {"case_type": "lower"}),
    ("fill_nulls", {"value": 0.0, "subset": ["revenue"]}),
    ("drop_nulls",),
    ("drop_duplicates",),
])

# Out comes a standard pandas DataFrame — use it like you always have
df = ar.to_pandas(clean)

# Use copy=True when you need defensive pandas-owned buffers
safe_df = ar.to_pandas(clean, copy=True)

Already have a pandas DataFrame? Use Arnio in-place in your existing pandas workflow:

import pandas as pd
import arnio as ar

df = pd.read_csv("messy_sales_data.csv")

clean_df = df.arnio.clean([
    ("strip_whitespace",),
    ("normalize_case", {"case_type": "lower"}),
    ("drop_duplicates",),
])

report = clean_df.arnio.profile()

Select specific columns

Use select_columns() to create a new ArFrame with only the required columns before converting to pandas.

selected = frame.select_columns(["name", "revenue"])

print(selected.columns)
# ['name', 'revenue']

Every step above executes in C++. Your Python code is a configuration — not the execution engine.

📸 Peek at a 100 GB file without loading it

scan_csv reads only the header + a sample to infer the schema. Zero data loaded.

schema = ar.scan_csv("100GB_file.csv")
# {'id': 'int64', 'name': 'string', 'is_active': 'bool', 'revenue': 'float64'}

Useful for exploring datasets before committing memory.

👀 Preview rows without pandas conversion or full-column Python list materialization

preview() reads only the first n rows directly from the C++ frame — no pandas conversion triggered.

frame = ar.read_csv("huge_file.csv")

print(frame.preview())      # first 5 rows (default)
print(frame.preview(n=10))  # first 10 rows

Raises ValueError for invalid n (zero, negative, or non-integer).

🧩 Add custom steps without touching C++

Register any Python function as a pipeline step. It receives a DataFrame, returns a DataFrame.

def remove_outliers(df, column="revenue", threshold=100_000):
    return df[df[column] <= threshold]

ar.register_step("remove_outliers", remove_outliers)

# Now use it in any pipeline alongside native C++ steps
clean = ar.pipeline(frame, [
    ("strip_whitespace",),
    ("remove_outliers", {"column": "revenue", "threshold": 50000}),
    ("drop_duplicates",),
])

Custom steps run through a pandas↔ArFrame conversion bridge. Prototype in Python, then optionally migrate hot paths to C++ for full speed.

---

🔗 Integrations

Arnio is designed to make the rest of the Python data stack more productive, not to replace it.

Workflow	How Arnio helps
pandas	Clean, validate, and profile messy DataFrames through df.arnio.
NumPy	Prepare typed numeric data before array/modeling workflows.
scikit-learn	Use Arnio cleaning as a preprocessing layer before model training.
DuckDB / Arrow	Validate and prepare data before analytics and columnar exchange.
notebooks	Inspect quality issues and cleaning suggestions before analysis.

Pandas accessor

df = pd.read_csv("raw_customers.csv")

clean_df = df.arnio.clean(drop_duplicates=True)
quality = clean_df.arnio.profile()
validation = clean_df.arnio.validate({
    "email": ar.Email(nullable=False),
    "age": ar.Int64(nullable=True, min=0),
})

This keeps pandas as the analysis tool while Arnio handles the preparation, quality, and validation layer.

Product direction: PROJECT_DIRECTION.md

---

🔍 Why Arnio exists

Every data project starts the same way:

df = pd.read_csv("data.csv")              # 💥 RAM spike — entire file as raw strings
df.columns = df.columns.str.strip()        # Why is this not automatic?
df["name"] = df["name"].str.strip()        # Python loop over every cell
df["name"] = df["name"].str.lower()        # Another Python loop
df = df.dropna()                           # Another pass
df = df.drop_duplicates()                  # Another pass

Six lines. Four full-data passes. All in interpreted Python. This is fine for a Jupyter demo — but it doesn't scale, it doesn't compose, and it definitely doesn't belong in production.

Arnio intercepts this entire pattern. It moves the preparation layer into a predictable pipeline, accelerates supported operations in C++, and gives you clean data for pandas, NumPy, scikit-learn, DuckDB, or notebooks.

Without Arnio df = pd.read_csv(path) df.columns = df.columns.str.strip() for col in str_cols: df[col] = df[col].str.strip() df[col] = df[col].str.lower() df = df.dropna(subset=["revenue"]) df = df.drop_duplicates() # 6+ lines, multiple passes, pure Python

With Arnio frame = ar.read_csv(path) df = ar.to_pandas(ar.pipeline(frame, [ ("strip_whitespace",), ("normalize_case", {"case_type": "lower"}), ("drop_nulls", {"subset": ["revenue"]}), ("drop_duplicates",), ])) # Declarative. Single pipeline. C++ execution.

---

🏗️ Architecture

Arnio is not a pandas wrapper. It's a separate runtime with its own data model.

┌──────────────────────────────────────────────────────────────┐
│  Your Python Code                                            │
│  frame = ar.read_csv("data.csv")                             │
│  clean = ar.pipeline(frame, [...])                           │
│  df = ar.to_pandas(clean)                                    │
└────────────────────────┬─────────────────────────────────────┘
                         │  pybind11 boundary
┌────────────────────────▼─────────────────────────────────────┐
│  C++ Runtime  (_arnio_cpp)                                   │
│                                                              │
│  ┌─────────────┐  ┌─────────────────┐  ┌──────────────────┐ │
│  │  CsvReader   │  │  Frame/Column   │  │  Cleaning Engine │ │
│  │  • RFC 4180  │  │  • Columnar     │  │  • drop_nulls    │ │
│  │  • BOM strip │  │  • std::variant │  │  • fill_nulls    │ │
│  │  • Type      │  │  • Bool null    │  │  • drop_dupes    │ │
│  │    inference │  │    masks        │  │  • strip_ws      │ │
│  │  • Quoted    │  │  • O(1) column  │  │  • normalize     │ │
│  │    fields    │  │    lookup       │  │  • rename/cast   │ │
│  └─────────────┘  └─────────────────┘  └──────────────────┘ │
│                                                              │
│  to_pandas() ──→ zero-copy NumPy buffer (numerics/bools)     │
└──────────────────────────────────────────────────────────────┘

Design decisions that matter

Decision	What it means
Columnar storage	Data lives in typed std::vectors — vector<int64_t>, vector<double>, vector<string> — not rows of variants. Cache-friendly and SIMD-ready.
Boolean null masks	Nulls are tracked in a separate vector<bool>, keeping data vectors dense. No sentinel values, no NaN tricks.
Two-pass CSV read	Pass 1 infers types across all rows. Pass 2 parses values directly into the correct typed column. No string→object→cast overhead.
Zero-copy bridge	to_pandas() exposes C++ memory directly via NumPy's buffer protocol where supported. Numeric columns preserve the fast zero-copy path by default, while copy=True requests defensive pandas-owned buffers.
Step registry	Pipeline steps map to C++ function pointers. Adding a new cleaning primitive is a single function + one registry entry.

Full architecture documentation: ARCHITECTURE.md

---

🏎️ Benchmarks

Reference environment: Ubuntu, Python 3.12, synthetic messy CSV inputs.
Reproduce: make benchmark — generates deterministic tall and wide datasets and runs both engines.

To reproduce the published numbers from a fresh checkout:

python -m venv .venv
source .venv/bin/activate
python -m pip install -U pip
python -m pip install -e .
python benchmarks/generate_data.py
python benchmarks/benchmark_vs_pandas.py

benchmarks/generate_data.py uses deterministic NumPy seeds, so every run creates the same benchmarks/benchmark_1m.csv tall input and benchmarks/benchmark_wide.csv wide input. The benchmark then executes three pandas runs and three arnio runs for each case, printing average wall-clock time from time.perf_counter() and peak Python allocation from tracemalloc. For cleaner comparisons, close other memory-heavy processes and run the script from the repository root after installing the same Python, pandas, NumPy, compiler, and arnio commit you want to compare.

Expected output format:

Tall CSV (1,000,000 rows x 12 columns)
Metric                     pandas        arnio
────────────────────────────────────────────
Exec Time (avg)       4.73s         5.75s
Peak RAM               211MB         212MB
Speed: 0.8x | RAM: -1% reduction

Wide CSV (5,000 rows x 256 columns)
Metric                     pandas        arnio
────────────────────────────────────────────
Exec Time (avg)       ...s          ...s
Peak RAM              ...MB         ...MB
Speed: ...x | RAM: ...% reduction

Small differences are expected across CPUs, operating systems, compilers, Python builds, and pandas/NumPy versions. If you share benchmark results in an issue or PR, include your OS, Python version, CPU model, pandas/NumPy versions, arnio commit, and the full command output so maintainers can compare like for like.

Arnio is near memory parity in the reference benchmark while replacing ad-hoc Python string loops with a compiled, declarative pipeline. Validate memory and speed on your own workload. The execution time gap is a known, active optimization target — the current drop_duplicates and strip_whitespace implementations use unoptimized row-key serialization.

✅ What's already won	🎯 What's being optimized
Native C++ parsing eliminates Python memory spikes Columnar storage matches pandas' internal efficiency Declarative API eliminates .apply() spaghetti Zero-copy bridge for numeric conversions	drop_duplicates — replace string serialization with hash-based comparisons strip_whitespace — in-place mutation instead of copy-on-write Parallel column processing via std::thread Help close the gap →

---

🧰 Cleaning primitives

Most operations below run natively in C++. The current filter_rows step uses the Python pipeline backend and may be optimized in C++ later.

Primitive	What it does	Example
drop_nulls	Remove rows with null/empty values	ar.drop_nulls(frame, subset=["age"])
validate_columns_exist	Fail early when required columns are missing	ar.validate_columns_exist(frame, ["age"])
filter_rows	Filter rows using comparison operators	ar.filter_rows(frame, column="age", op=">", value=18)
fill_nulls	Replace nulls with a scalar	ar.fill_nulls(frame, 0, subset=["revenue"])
drop_duplicates	Deduplicate rows (first/last/none)	ar.drop_duplicates(frame, keep="first")
drop_constant_columns	Remove columns with only one unique value	ar.drop_constant_columns(frame)
clip_numeric	Clip numeric values to lower and/or upper bounds	ar.clip_numeric(frame, lower=0, upper=100)
strip_whitespace	Trim leading/trailing spaces from strings	ar.strip_whitespace(frame)
normalize_case	Force lower/upper/title case	ar.normalize_case(frame, case_type="title")
rename_columns	Rename columns via mapping	ar.rename_columns(frame, {"old": "new"})
cast_types	Cast column types	ar.cast_types(frame, {"age": "int64"})
round_numeric_columns	Round numeric columns (non-numeric columns in subset ignored safely)	ar.round_numeric_columns(frame, decimals=2)
clean	Convenience shorthand	ar.clean(frame, drop_nulls=True)
safe_divide_columns	Divide one column by another, handling zero/null denominators	ar.safe_divide_columns(frame, numerator="revenue", denominator="cost", output_column="ratio")

Or compose them all into a pipeline:

clean = ar.pipeline(frame, [
    ("validate_columns_exist", {"columns": ["name", "city", "revenue"]}),
    ("strip_whitespace",),
    ("normalize_case", {"case_type": "lower"}),
    ("fill_nulls", {"value": "unknown", "subset": ["city"]}),
    ("drop_duplicates", {"keep": "first"}),
])

🔎 Filter rows inside pipelines

Use filter_rows to keep only rows matching a condition.

clean = ar.pipeline(frame, [
    ("filter_rows", {
        "column": "revenue",
        "op": ">=",
        "value": 1000
    }),
])

Supported operators:

• >
• <
• >=
• <=
• ==
• !=

Works with:

• integers
• floats
• strings
• booleans

### 🔢 Safe column division

Divide one column by another while handling division by zero and null denominators explicitly:

result = ar.safe_divide_columns(
    frame,
    numerator="revenue",
    denominator="cost",
    output_column="ratio",
    fill_value=0.0,  # used when denominator is zero or null
)

When the denominator is zero or null, the result is replaced with fill_value (default 0.0) instead of raising an error or producing NaN/Inf.

---

📊 Pandas Dtype Support Matrix

This table helps users understand which pandas dtypes and workflows are fully supported, partially supported, unsupported, or planned.

If a dtype is partially supported, users may need conversion before processing. Unsupported dtypes should raise clear errors where applicable.

Pandas Dtype	Support Status	Notes
int64	✅ Supported	Fully supported with native C++ columnar storage
float64	✅ Supported	Fully supported with zero-copy conversion where possible
bool	✅ Supported	Native supported boolean type
string	✅ Supported	Recommended over object dtype for text workflows
datetime64[ns]	❌ Unsupported	No native datetime parsing or conversion support yet
category	⚠️ Limited	Converted to string/object during processing
object (mixed columns)	⚠️ Limited	Mixed object columns may coerce to string and reduce type inference reliability
nullable pandas dtypes (Int64, boolean)	⚠️ Limited	Supported through pandas extension dtypes with null-mask handling
timedelta64[ns]	❌ Unsupported	Not currently supported

Notes

• Numeric columns are optimized for zero-copy conversion between C++ and pandas where supported.
• Pass copy=True to to_pandas() when downstream pandas code needs defensive pandas-owned column buffers.
• Boolean conversion is already copied by the binding because std::vector<bool> cannot be exposed as a zero-copy NumPy buffer in the current implementation.
• Columns with null masks may require copies so pandas can apply nullable values safely.
• String columns require Python string object creation during to_pandas() conversion.
• Mixed object columns may reduce type inference accuracy and may require preprocessing.
• Unsupported dtypes should raise clear user-facing errors instead of silent failures.

---

🧠 Data quality engine

Arnio now includes built-in dataset understanding before you analyze in pandas.

report = ar.profile(frame)
print(report.summary())

suggestions = ar.suggest_cleaning(frame)
clean = ar.pipeline(frame, suggestions)

For production data contracts:

schema = ar.Schema({
    "id": ar.Int64(nullable=False, unique=True),
    "email": ar.Email(nullable=False),
    "username": ar.String(min_length=3, max_length=20),
    "revenue": ar.Float64(nullable=True, min=0),
})

result = ar.validate(frame, schema)
if not result.passed:
    summary = result.summary()
    print(summary["issues_by_rule"])
    print(summary["issues_by_column"])
    print(summary["issues_by_column_and_rule"])
    print(result.to_pandas())
    print(result.to_markdown(max_issues=10))

ValidationResult.to_markdown() is useful in CI logs, GitHub comments, or data quality reports because it renders a compact validation summary plus a GitHub-friendly issue table. Severity counts are not included in summary() yet because ValidationIssue does not currently carry severity information.

For low-risk automatic cleanup:

clean, report = ar.auto_clean(frame, mode="strict", return_report=True)

This is the layer pandas does not try to own: profiling, data contracts, row-level validation issues, and safe cleaning suggestions for messy incoming datasets.

Beginner-friendly auto-clean tutorial

Use this workflow when you receive a small messy dataset and want to inspect what Arnio will change before applying it.

import arnio as ar
import pandas as pd

raw = pd.DataFrame(
    {
        "order_id": [1001, 1002, 1002, 1003, 1004],
        "customer": [" Ishan ", " Prasoon ", " Prasoon ", " Pranay ", " Dhruv "],
        "city": [" Paris ", "London", "London", " New York ", " Tokyo "],
    }
)

frame = ar.from_pandas(raw)

report = ar.profile(frame)
summary = report.summary()
print(summary)

suggestions = ar.suggest_cleaning(frame)
print(suggestions)
# [('strip_whitespace', {'subset': ['customer', 'city']}), ('drop_duplicates', {'keep': 'first'})]

safe = ar.auto_clean(frame)
strict = ar.auto_clean(frame, mode="strict")

Messy input:

order_id	customer	city
1001	Ishan	Paris
1002	Prasoon	London
1002	Prasoon	London
1003	Pranay	New York
1004	Dhruv	Tokyo

Expected cleaned output with mode="strict":

order_id	customer	city
1001	Ishan	Paris
1002	Prasoon	London
1003	Pranay	New York
1004	Dhruv	Tokyo

mode="safe" only trims whitespace. Use mode="strict" when you also want deterministic built-in cleanup such as exact duplicate removal.

See examples/auto_clean_tutorial.py for a runnable version of this walkthrough.

Data Quality Reports

Arnio provides detailed profiling for datasets via ar.profile(). To generate the report shown in these examples, the following code was used:

import arnio as ar
import pandas as pd

# Sample dataset used for these examples
data = {
    "user_id": [101, 102, 103, 104],
    "email": ["[e-mail]", "invalid-email", None, "[e-mail]"],
    "score": [85.5, 90.0, None, 88.2]
}
df = ar.from_pandas(pd.DataFrame(data))
# Bounded profiling for large datasets (controls how many sample values are kept)
report = ar.profile(df, sample_size=5)

1. Terminal Representation (Simplified Example)

A simplified view of the standard string representation of the report object:

DataQualityReport(
    row_count=4,
    column_count=3,
    memory_usage=733,
    duplicate_rows=0,
    columns={
        'user_id': ColumnProfile(dtype='int64', semantic_type='identifier', unique_count=4),
        'email': ColumnProfile(dtype='string', semantic_type='categorical', null_count=1, unique_ratio=0.666667),
        'score': ColumnProfile(dtype='float64', semantic_type='numeric', mean=87.9, min=85.5, max=90.0)
    }
)

2. JSON Format (Excerpts from .to_dict())

Key fields from the structured JSON export for integration with APIs or dashboards:

{
  "row_count": 4,
  "column_count": 3,
  "memory_usage": 733,
  "duplicate_rows": 0,
  "duplicate_ratio": 0.0,
  "columns": {
    "user_id": {
      "dtype": "int64",
      "semantic_type": "identifier",
      "null_count": 0,
      "unique_ratio": 1.0
    },
    "email": {
      "dtype": "string",
      "semantic_type": "categorical",
      "null_count": 1,
      "unique_ratio": 0.666667,
      "warnings": ["contains_nulls"]
    },
    "score": {
      "dtype": "float64",
      "semantic_type": "numeric",
      "null_count": 1,
      "mean": 87.9,
      "min": 85.5,
      "max": 90.0,
      "warnings": ["contains_nulls"]
    }
  }
}

3. Example Summary Table

A manually formatted Markdown table representing the core metrics:

Metric	Value
Row Count	4
Column Count	3
Memory Usage	733 bytes
Duplicates	0 (0.0%)

🗺️ Roadmap

Version	Focus	Status
v1.0	Stable release · cross-platform wheels · CI/CD · PyPI publishing · Google Colab support	✅ Shipped
v1.1	Production readiness · release hardening · docs/tooling	✅ Shipped
v1.2	C++ pipeline optimization · speed parity with pandas · hash-based deduplication	🔨 Active
v1.3	Chunked / streaming processing · Parquet & JSON readers	📋 Planned
v1.4	Parallel column processing · SIMD string operations	💭 Exploring

---

💬 Community

Join the Arnio Discord Community for quick setup help, contributor onboarding, GSSoC 2026 coordination, feature discussion, and community updates.

Discord is for fast conversation and support. GitHub remains the source of truth for issue assignment, PR reviews, bugs, roadmap decisions, and releases.

---

🤝 Contribute

Arnio is a GSSoC 2026 project with a structured contributor backlog across beginner, intermediate, and advanced tracks.

You don't need C++ to contribute

Most new features are pure Python pipeline steps:

# 1. Write a function that takes a DataFrame and returns a DataFrame
def remove_special_chars(df, columns=None):
    cols = columns or df.select_dtypes("object").columns
    for col in cols:
        df[col] = df[col].str.replace(r"[^a-zA-Z0-9\s]", "", regex=True)
    return df

# 2. Register it
ar.register_step("remove_special_chars", remove_special_chars)

# 3. Write tests, open a PR. That's it.

If you do know C++

The biggest performance wins are in:

• drop_duplicates — replacing std::ostringstream row serialization with proper hash-based comparisons
• strip_whitespace — converting from copy-on-write to in-place mutation
• Parallel column processing — std::thread across independent columns

Getting started

# macOS / Linux
git clone https://github.com/im-anishraj/arnio.git && cd arnio
make install   # pip install -e ".[dev]" + pre-commit
make test      # pytest with coverage
make lint      # ruff + black

# Windows
pip install -e ".[dev]"
pre-commit install
pytest tests/ -v

PR titles must follow Conventional Commits — feat:, fix:, docs:, chore:. Our release pipeline auto-generates changelogs from these.

For GSSoC contributors, please read GSSOC_GUIDE.md before asking to be assigned. It explains issue claiming, contribution levels, review expectations, and what maintainers look for in a strong PR. If you want a quick onboarding refresher, see the GSSoC FAQ. If you are new to Arnio terms, see the contributor glossary.

📖 Full Contributing Guide ·  GSSoC Guide ·  🐛 Open Issues ·  💬 Discussions ·  Discord

---

🚢 Release process

Arnio releases are automated through Release Please and GitHub Actions.

1. Merge user-facing changes with Conventional Commit PR titles (feat:, fix:, docs:, or chore:) so Release Please can choose the version bump and changelog entries.
2. Review and merge the Release Please PR on main; this updates release metadata and creates the GitHub release and tag.
3. Confirm the Build & Publish Wheels workflow succeeds for the release tag. It builds the sdist and wheels, then publishes to PyPI through Trusted Publishing.
4. Smoke test the published package in a clean environment:

python -m venv /tmp/arnio-smoke
source /tmp/arnio-smoke/bin/activate
python -m pip install -U pip
python -m pip install arnio
printf 'name,revenue\n Ada,10\n' > /tmp/arnio-smoke.csv
python - <<'PY'
import arnio as ar
print(ar.__version__)
print(ar.scan_csv("/tmp/arnio-smoke.csv"))
PY

5. Verify the GitHub release, PyPI project page, and install command all show the expected version before announcing the release.

If any publish or smoke-test step fails, leave the failed tag and GitHub release in place until maintainers agree on the recovery plan.

---

📐 Project structure

arnio/
├── cpp/
│   ├── include/arnio/      # C++ headers — types, column, frame, csv_reader, cleaning
│   └── src/                 # C++ implementations (~30 KB of compiled logic)
├── bindings/
│   └── bind_arnio.cpp       # pybind11 module — the Python↔C++ bridge
├── arnio/
│   ├── __init__.py          # Public API surface
│   ├── io.py                # read_csv, scan_csv
│   ├── cleaning.py          # Python wrappers for C++ cleaning functions
│   ├── pipeline.py          # Step registry + pipeline executor
│   ├── convert.py           # to_pandas (zero-copy), from_pandas
│   ├── frame.py             # ArFrame — lightweight C++ Frame wrapper
│   └── exceptions.py        # ArnioError, UnknownStepError, CsvReadError, TypeCastError
├── tests/                   # pytest suite — CSV, cleaning, pipeline, conversions
├── benchmarks/              # Reproducible arnio vs pandas benchmark
├── examples/                # basic_usage.py, auto_clean_tutorial.py, custom_step.py
└── website/                 # Project website — arnio.vercel.app

---

Stop writing cleaning scripts. Declare clean data.

       

_{Built with C++ and pybind11 · Licensed under MIT · Maintained by @im-anishraj}