datapill 0.2.2

A modular data preparation CLI for ingestion, profiling, classification, preprocessing, and export pipelines

🗂️ Data Pill

A modular, open-source data preparation CLI for developers, data analysts, data scientists, and data engineers.

datapill gives you a single command-line tool to ingest, profile, classify, preprocess, and export data - from any source, to any destination - with a clean pipeline model and full artifact tracking. A hosted SaaS is coming soon.

---

💡 Why Data Pill?

Most data prep work happens in ad-hoc scripts, notebooks, or half-finished internal tools. datapill replaces that chaos with a reproducible, composable pipeline that works the same way whether you're exploring a CSV on your laptop or processing millions of rows from Kafka in production.

	If you are…	datapill helps you…
🧑‍💻	Developer	Build data pipelines from config files, run them in CI, generate standalone Python scripts
📊	Data Analyst	Profile any dataset in seconds, detect nulls, distributions, and anomalies without writing code
🤖	Data Scientist	Classify columns by semantic type, preprocess features (scaling, imputation, encoding) with one command
⚙️	Data Engineer	Stream-ingest from Postgres, MySQL, S3, Kafka, REST APIs; write back with upsert support

---

📦 Installation

pip install datapill

Requires Python 3.11+.

This installs the core CLI (~200 MB) with rule-based classification, profiling, preprocessing, and export. No ML model is required.

With embedding support

To use --mode embedding or --mode hybrid in dp classify, install the ML extras (~3.5 GB, includes PyTorch and sentence-transformers):

pip install "datapill[ml]"

Note: Without [ml], hybrid mode still works — it runs rule-based classification for all columns and skips the embedding fallback. Only columns that would otherwise be sent to the embedding model are affected.

Verify the install:

dp --help

---

🚀 Quick Start

1. Ingest a local CSV

dp ingest --source local_file --path data/sales.csv

2. Profile the dataset

dp profile --input <run_id>

3. Classify columns by semantic type

dp classify --input <run_id> --mode hybrid

4. Preprocess with a pipeline config

dp preprocess --input <run_id> --pipeline pipeline.json

5. Export to Parquet

dp export --input <run_id> --format parquet --out-path output/result.parquet

---

🛠️ Commands

dp ingest

Stream data from any supported source into the artifact store.

dp ingest --source postgresql --config pg.json --table orders
dp ingest --source s3 --config s3.json --url s3://my-bucket/data.parquet
dp ingest --source kafka --config kafka.json --topic events --max-records 10000
dp ingest --source rest --config api.json --endpoint /users
dp ingest --source local_file --path data/sales.csv --limit 50000

Supported sources: local_file · postgresql · mysql · s3 · rest · kafka

Supported formats (local / S3): csv · parquet · json · jsonl · excel

---

dp profile

Compute a full statistical profile of any ingested dataset.

dp profile --input <run_id>
dp profile --input <run_id> --mode summary
dp profile --input <run_id> --sample-strategy random --sample-size 100000
dp profile --input <run_id> --correlation spearman

What you get:

• Per-column: null rate, distinct count, min/max/mean/median, std, skewness, kurtosis, percentiles, histogram
• Top value frequencies with percentages
• Pattern detection: email, URL, phone, UUID, ISO date
• Correlation matrix (Pearson or Spearman) for all numeric columns
• Warnings: HIGH_NULL_RATE, CONSTANT_COLUMN, SKEWED_DISTRIBUTION, HIGH_CARDINALITY, POTENTIAL_IDENTIFIER

---

dp classify

Classify every column in a dataset by its semantic type - automatically.

dp classify --input <run_id> --mode hybrid
dp classify --input <run_id> --mode rule_based --threshold 0.65
dp classify --input <run_id> --overrides '{"age": "numerical_continuous", "y": "target_label"}'

Modes:

Mode	How it works	Requires
rule_based	Regex patterns on column names + dtype heuristics. Fast, zero ML dependencies.	core
embedding	Semantic similarity via sentence-transformers (all-MiniLM-L6-v2) against anchor texts per type.	datapill[ml]
hybrid	Rule-based first; embedding kicks in only for ambiguous or unknown columns.	datapill[ml] for full accuracy

Without [ml]: hybrid mode runs entirely on rule-based logic. Columns that cannot be resolved by rules are returned as unknown instead of being sent to the embedding model.

Semantic types detected: identifier · numerical_continuous · numerical_discrete · categorical_nominal · categorical_ordinal · text_freeform · text_structured · datetime · boolean · geospatial · embedding · target_label

---

dp preprocess

Apply a preprocessing pipeline defined in a JSON config file.

dp preprocess --input <run_id> --pipeline pipeline.json
dp preprocess --input <run_id> --pipeline pipeline.json --dry-run
dp preprocess --input <run_id> --pipeline pipeline.json --checkpoint

Pipeline config format:

{
  "steps": [
    { "type": "impute_mean",      "scope": { "columns": ["age", "income"] } },
    { "type": "clip_iqr",         "scope": { "columns": ["income"] } },
    { "type": "standard_scaler",  "scope": { "columns": ["age", "income"] } },
    { "type": "onehot",           "scope": { "columns": ["category"] } },
    { "type": "drop_missing",     "scope": { "columns": [] } }
  ]
}

Available steps:

Category	Steps
Missing values	impute_mean · impute_median · impute_mode · drop_missing
Outliers	clip_iqr · clip_zscore
Scaling	standard_scaler · minmax_scaler · robust_scaler
Encoding	onehot · ordinal
Structure	select_columns · drop_columns · rename_columns · cast_dtype · deduplicate
Custom	custom_python (sandboxed via RestrictedPython)

--dry-run runs on the first 1,000 rows and prints a preview without saving any artifact.
--checkpoint saves the DataFrame after each step so you can resume from any point.

---

dp export

Export a processed dataset to a file or write it back to a database or S3.

# Export to file
dp export --input <run_id> --format parquet --out-path output/result.parquet

# Write back to PostgreSQL (upsert)
dp export --input <run_id> --format parquet \
  --connector pg.json --write-mode upsert --primary-keys id

# Write to S3
dp export --input <run_id> --format csv --connector s3.json

Write modes: replace · append · upsert
Output formats: csv · parquet · json · jsonl · excel · arrow

---

dp pipeline export

Generate a standalone Python script from a preprocess pipeline artifact — no datapill dependency required at runtime.

dp pipeline export -i <run_id> -s local_file --path data.csv
dp pipeline export -i <run_id> -s postgresql -c pg.json
dp pipeline export -i <run_id> -s local_file --path data.csv --with-tests
dp pipeline export -i <run_id> --out-dir ./generated

What you get:

• Preprocessing steps reconstructed from the saved config artifact
• Ingest configuration merged into a single self-contained script
• A run_<name>.py entry point with a --dry-run flag
• An optional test_<name>.py scaffold (with --with-tests)

Generated files:

File	Description
run_<name>.py	Main pipeline script — runs without datapill
test_<name>.py	pytest scaffold (only with --with-tests)

Options:

Option	Description
--input, -i	run_id or preprocess artifact ID
--source, -s	Connector type: local_file · postgresql · mysql · s3
--ingest-config, -c	Connector JSON config (same as dp ingest --config)
--path	File path (local_file)
--table	Table name (postgresql | mysql)
--url	S3 URL
--format, -f	Output format: csv · parquet · json · jsonl · excel
--out-path	Output path hard-coded into the generated script
--name, -n	Base name for generated files, e.g. orders → run_orders.py
--compression	snappy · zstd · gzip (parquet only)
--with-tests	Also generate test_<name>.py
--out-dir, -o	Directory to write generated files (default: generated/)

Run the generated pipeline:

python generated/run_<name>.py --dry-run
python generated/run_<name>.py

Run the generated tests:

python -m pytest generated/test_<name>.py -v

Note: dp pipeline export requires a preprocess artifact saved without --dry-run. If only a dry-run artifact exists, re-run dp preprocess without that flag first.

---

dp connector

Inspect and interact with any connector directly - without running a full pipeline.

dp connector test     --source postgresql --config pg.json
dp connector schema   --source postgresql --config pg.json --table orders
dp connector upload   --source s3 --config s3.json --src-path data.csv --dest-url s3://bucket/data.csv
dp connector download --source s3 --config s3.json --url s3://bucket/data.csv --out-path ./data.csv
dp connector list     --source s3 --config s3.json --prefix input/
dp connector exec     --source postgresql --config pg.json --sql "DELETE FROM orders WHERE status='cancelled'"
dp connector truncate --source postgresql --config pg.json --table orders
dp connector produce  --source kafka --config kafka.json --topic events --file records.json

---

dp list

List all artifacts in the store.

dp list
dp list --feature ingest
dp list --limit 50

---

dp run

Run a full ingest + profile pipeline from a single config file.

dp run pipeline.json

Config format:

{
  "source": "postgresql",
  "connector": { "host": "localhost", "database": "mydb", "user": "u", "password": "p" },
  "query": { "table": "orders" },
  "ingest": { "batch_size": 10000 },
  "profile": { "mode": "full", "correlation": "pearson" }
}

---

🔌 Connector Configuration

All connectors are configured via JSON files passed with --config.

PostgreSQL / MySQL

{
  "host": "localhost",
  "port": 5432,
  "database": "mydb",
  "user": "myuser",
  "password": "mypassword"
}

S3

{
  "aws_access_key_id": "AKIA...",
  "aws_secret_access_key": "...",
  "region": "us-east-1",
  "bucket": "my-bucket"
}

Kafka

{
  "bootstrap_servers": ["localhost:9092"],
  "group_id": "datapill",
  "value_format": "json",
  "security_protocol": "PLAINTEXT"
}

SASL/SSL is supported - add sasl_mechanism, sasl_username, sasl_password, and ssl_cafile as needed.

REST API

{
  "base_url": "https://api.example.com",
  "headers": { "Authorization": "Bearer <token>" },
  "response_path": "data",
  "pagination": {
    "type": "offset",
    "limit": 100,
    "limit_param": "limit",
    "offset_param": "offset"
  }
}

Pagination modes: offset · cursor · link_header

---

🗃️ Artifact Store

Every pipeline run produces artifacts - Parquet files and JSON metadata - stored locally and tracked in a registry.

src/datapill/artifacts/
├── registry.json
├── a1b2c3d4_ingest_output.parquet
├── a1b2c3d4_ingest_schema.json
├── e5f6g7h8_profile_detail.json
└── e5f6g7h8_profile_summary.json

You can reference any artifact by its run_id (short 8-char hex) or full artifact_id. datapill resolves ambiguity automatically using feature-aware priority rules.

Change the artifact directory with --out:

dp ingest --source local_file --path data.csv --out /my/artifacts

---

🏗️ Architecture

datapill/
├── cli/            # Typer CLI - entry point for all commands
├── connectors/     # Source adapters (local, PG, MySQL, S3, REST, Kafka)
├── core/           # PipelineContext, ProgressEvent, FeaturePipeline interface
├── executor/       # Sandboxed code execution (RestrictedPython + Docker)
├── features/
│   ├── ingest/     # Stream ingestion → Parquet artifacts
│   ├── profile/    # Statistical profiling + correlation
│   ├── classify/   # Semantic type classification (rule-based + embedding)
│   ├── preprocess/ # Step-based transformation pipeline
│   └── export/     # File export + DB/S3 write-back + code generation
└── storage/        # ArtifactStore - registry, save/load, resolve

Every feature implements the same FeaturePipeline interface: validate → plan → execute. Pipelines emit async ProgressEvent streams so the CLI can render live progress bars.

---

🐍 Custom Python Steps (Sandboxed)

You can write arbitrary Python transformation logic and run it safely inside the preprocess pipeline.

# my_transform.py
def transform(df):
    return df.with_columns(
        (pl.col("revenue") / pl.col("units")).alias("avg_price")
    )

{
  "steps": [
    {
      "type": "custom_python",
      "scope": { "columns": [] },
      "params": { "code": "<contents of my_transform.py>", "func": "transform" }
    }
  ]
}

Custom code is validated by an AST analyzer (banned imports, banned builtins, dunder access) before execution. Two sandbox backends are available:

• RestrictedPython - in-process, low overhead, suitable for most use cases
• Docker - full container isolation (--network none, read-only FS, memory + CPU limits), for untrusted code

---

🧑‍💻 Development Setup

git clone https://github.com/your-org/datapill.git
cd datapill
python -m venv .venv
source .venv/bin/activate       # Windows: .venv\Scripts\activate
pip install -e ".[dev]"

For full embedding support in development:

pip install -e ".[ml,dev]"

Run the tests:

pytest
pytest -m "not integration"     # skip tests that require Docker services
pytest --cov=datapill

Lint:

ruff check src/
ruff format src/

---

🗺️ Roadmap

• Web UI / dashboard for artifact browsing and profile visualization
• Custom step registry - register and share reusable step plugins
• datapill SaaS - hosted pipelines, scheduling, collaboration, and monitoring
• dbt integration - use datapill as a pre-processing layer before dbt models
• Great Expectations integration - attach data quality assertions to any pipeline step

---

🤝 Contributing

Contributions are welcome. Please open an issue before submitting a large pull request so we can discuss the approach.

1. Fork the repository
2. Create a feature branch: git checkout -b feature/my-feature
3. Make your changes with tests
4. Run ruff check and pytest before pushing
5. Open a pull request against main

---

📄 License

MIT License. See LICENSE for details.

---

datapill SaaS - hosted pipelines, scheduling, and collaboration - coming soon.