Zero-Effort Data Discovery & Analysis

C++17-powered EDA and data cleaning engine for Python.

Guide · API reference · How it works · Contributing · Security

---

What is Zedda

Zedda profiles, cleans, and validates datasets from a single Python call. Its core is written in C++17 and streams data in constant memory, so it scales from a 900-row CSV to a terabyte-scale Parquet file without changing how you use it.

Installation

pip install --upgrade pip                  # Ensure pip >= 22.3 for abi3 wheel recognition
pip install zedda                          # CSV only — zero native deps
pip install "zedda[parquet]"               # adds Parquet/Arrow/Feather support
pip install "zedda[clean]"                 # adds fuzzy typo detection
pip install "zedda[ai]"                    # adds AI Q&A (zd.ask with Groq/OpenAI)
pip install "zedda[parquet,clean,ai]"      # everything together

Platform support:

Platform	Base install	[parquet] extra
Linux x86_64 (glibc >= 2.17 / musl)	✅ prebuilt wheel	✅ prebuilt pyarrow wheel
Linux ARM64 (aarch64)	✅ prebuilt wheel	✅ prebuilt pyarrow wheel
macOS Intel (x86_64)	✅ prebuilt wheel	✅ prebuilt pyarrow wheel
macOS Apple Silicon (ARM64)	✅ prebuilt wheel	✅ prebuilt pyarrow wheel
Windows x86_64	✅ prebuilt wheel	✅ prebuilt pyarrow wheel
Windows ARM64	✅ prebuilt wheel	⚠️ pyarrow has no win_arm64 wheel — Parquet requires manual build
Python 3.13 free-threaded (cp313t)	⚠️ Not yet supported — use standard Python 3.13

conda users: Zedda is not yet on conda-forge. Install via pip inside your conda environment:

conda activate myenv
pip install zedda

How to use

import zedda as zd

zd.profile("data.csv")  # full EDA report in terminal
zd.scan("data.csv")  # silent scan for CI/CD pipelines
zd.compare("train.csv", "test.csv")  # train/test drift detection
zd.fix("data.csv", apply=True)  # generate or apply pandas fix code
zd.ask("data.csv", "any nulls here?")  # plain-English dataset Q&A
zd.ml_ready("data.csv")  # readiness score for ML training
zd.warnings("data.csv")  # list all issues ranked by severity
zd.clean("data.csv", output="clean.csv")  # safe, backed-up auto-clean
zd.merge(["jan.csv", "feb.csv"], "out")  # safely combine multiple files
zd.report("data.csv", output="rep.html")  # export full report to offline HTML

Every function also accepts a pandas DataFrame directly — a file path is never required.

import pandas as pd

df = pd.read_csv("data.csv")
zd.profile(df)

See the full list of available functions in the API reference, and how the underlying engine works in How it works.

Installation from source

git clone https://github.com/Zedda-Labs/Zedda.git --recursive
cd Zedda
# C++17 build tools (cmake, ninja) are required
pip install cmake ninja
pip install -e ".[dev]"
pytest tests/

Enterprise Data Quality FAQ

1. How does ZEDDA prevent false-positive cleaning recommendations?

ZEDDA evaluates statistical distributions across single-pass scans before flagging issues:

• ID Columns: Flagged only when cardinality is 100.0% unique across all rows.
• Zero-Variance Columns: Flagged only when a single distinct value spans 100.0% of non-null cells.
• Outliers: Identified using IQR bounds (1.5 * IQR) combined with skewness thresholds.
• Multicollinearity: Evaluated via single-pass Pearson correlation matrices (|r| >= 0.95).

2. Is auto-cleaning (`zd.clean()`) safe for production data ingestion?

Yes. zd.clean() is fully deterministic, atomic, and safe:

• Automatic Backup: Creates {filename}.zedda-backup prior to mutating any dataset.
• Audit Trail: Outputs {stem}.audit.json containing exact before/after row counts, quality scores, applied transformations, and timing logs.
• Rollback Support: Run zd.clean.undo("filename.csv") at any time to instantly restore the original dataset from backup.

3. How does ZEDDA integrate into CI/CD pipelines and orchestrators?

• Programmatic Python: Use zd.scan("data.csv") inside Airflow / Prefect tasks to access .to_dict() and .to_json() metadata objects.
• CLI Ingestion: Run zedda scan data.csv --json inside GitHub Actions or GitLab CI to pipe structured JSON metadata directly into downstream data warehouse checks.
• Air-Gapped Compliance: zd.report() generates completely self-contained HTML files (zero CDN links, zero tracking, zero external network calls) suitable for banking and healthcare environments.

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md.

License

MIT — see LICENSE.

PyPI · GitHub · Issues · Guide · Architecture · API Docs · Changelog · Releases