etl_utilities
A Python ETL framework for data cleaning, type inference, SQL generation, and database loading that works across three DataFrame backends (Pandas, Polars, Spark) and three database families (MSSQL, MySQL/MariaDB, PostgreSQL).
It exists to remove the boilerplate from the boring-but-critical middle of a data pipeline: taking a messy extract, coercing every column to a sane type, generating a matching CREATE TABLE, and loading it efficiently — without hand-writing per-column casts or dialect-specific SQL.
Table of Contents
- Installation
- Quick Start — an end-to-end pipeline
- Batch-safe cleaning for streaming (Polars)
- Reconciling mismatched Parquet schemas
- Backends & databases
- Class reference
- Package documentation
- Logging
- License
Installation
pip install etl_utilities
The core install pulls in pandas, numpy, SQLAlchemy, python-dateutil, and rich.
Backend-specific dependencies are optional extras:
pip install etl_utilities[polars] # Polars cleaning
pip install etl_utilities[spark] # Spark cleaning
pip install etl_utilities[postgres] # Postgres loading (psycopg2)
pip install etl_utilities[arrow] # Parquet schema reconciliation (pyarrow)
pip install etl_utilities[all] # everything
Requires Python 3.10+.
Quick Start — an end-to-end pipeline
Take a messy CSV export, clean it, generate a Postgres table from the inferred schema, and load it:
import pandas as pd
from etl.dataframe.cleaner import Cleaner
from etl.query.creator import Creator
from etl.database.sql_dialects import postgres
from etl.database.connector import Connector
from etl.database.unified_loader import Loader
# 1. EXTRACT — read a messy upstream export
raw = pd.read_csv("sales_export.csv")
# 2. TRANSFORM — standardize column names (in place) and coerce types
Cleaner.column_names_to_snake_case(raw) # "Order ID" -> "order_id"
clean = Cleaner.clean_df(raw) # drop empty rows/cols + infer bool/int/float/date
# 3. GENERATE DDL — a CREATE TABLE matching the cleaned frame's inferred schema
ddl = Creator.create_table(
clean,
schema_name="analytics",
table_name="sales",
dialect=postgres,
primary_key_column="order_id",
)
# 4. LOAD — open a connection and bulk-insert
connection = Connector(
host="localhost", port=5432, instance="",
database="warehouse", username="etl", password="***",
).to_user_postgres()
connection.execute(ddl)
Loader(connection, clean, schema="analytics", table="sales", dialect=postgres).insert()
Swap postgres for mssql or mariadb (from etl.database.sql_dialects) and the generated DDL and the loader's placeholders/escaping change with it — the rest of the pipeline is identical.
Batch-safe cleaning for streaming (Polars)
Per-frame type inference is correct for one in-memory table but breaks batch-by-batch: batch A might infer Boolean/Int8 while batch B infers String/Int32, and any downstream concat / Parquet-dataset / DB append then rejects the schema mismatch.
PolarsCleaner solves this by separating inference (decide the plan once) from application (apply identically to every batch):
from etl.dataframe.polars.cleaner import PolarsCleaner
# Infer a deterministic plan once, from a representative sample.
# prefer_float keeps numeric columns stable even if a later batch turns out fractional.
plan = PolarsCleaner.infer_cleaning_plan(sample_df, prefer_float=True)
for batch in stream: # every batch gets the SAME schema
batch = PolarsCleaner.apply_cleaning_plan(batch, plan)
sink.write(batch)
For one-shot in-memory cleaning, the original PolarsCleaner.clean_all_types(df) / clean_df(df) still do everything in a single pass.
A few more schema-stability helpers, useful right before a strict sink (Delta/Parquet/warehouse append):
df = PolarsCleaner.sanitize_float_columns(df) # NaN / +-Inf (and "NaN"/"Infinity" strings) -> null
df = PolarsCleaner.localize_naive_datetimes(df) # naive Datetime -> tz-aware (UTC by default)
df = PolarsCleaner.cast_null_columns(df) # all-null (pl.Null) columns -> a concrete dtype
df = PolarsCleaner.clean_df(df, drop_all_null_columns=False) # keep all-null columns for schema parity
# Collision-safe snake_case: when two source columns normalize to the same name,
# coalesce them (default), or pass on_collision="error" / "suffix".
df = PolarsCleaner.column_names_to_snake_case(df)
Reconciling mismatched Parquet schemas
A pile of Parquet files with the same logical columns but incompatible physical types
(Decimal(38,12) vs Decimal(38,13), string-vs-decimal) defeats pl.scan_parquet([...])
and pa.unify_schemas(...) alike. unify_parquet_schemas reads only the footers (concurrently)
and builds one schema every file can be cast to on read:
import pyarrow.dataset as ds
from etl.io.arrow_schema import unify_parquet_schemas
unified = unify_parquet_schemas(parquet_paths) # any decimal -> float64; cross-file conflict -> large_string
table = ds.dataset(parquet_paths, schema=unified).to_table() # PyArrow casts on read
Combine the three: unify schemas → stream batches → apply_cleaning_plan rebuilds a full
schema-stable, memory-safe cleaning pipeline from library primitives alone.
Backends & databases
| DataFrame backend | Module | Notes |
|---|---|---|
| Pandas | etl.dataframe.cleaner / parser / analyzer |
Eager, the most complete implementation |
| Polars | etl.dataframe.polars |
Expression-based; adds the batch-safe streaming API above |
| Spark | etl.dataframe.spark |
Distributed; native Spark SQL (no Python UDFs), sampling-based inference |
All backends follow the same type-narrowing order: Boolean → Integer → Float → Date → String.
Database dialects (etl.database.sql_dialects): mssql, mariadb, postgres — each defines its own
escaping, type mapping, placeholders, and constraint syntax, consumed by the Creator and Loader.
Class reference
- Connector — SQLAlchemy connection factory for MSSQL (trusted/user), PostgreSQL, and MySQL.
- Loader (
unified_loader) — dialect-awareINSERTloader from a DataFrame;MsSqlLoader/MySqlLoaderadd backend-specific optimizations (e.g.fast_executemany). - Parser — static type parsers (boolean, float, date, integer) shared by the cleaners.
- Cleaner / PolarsCleaner — column-name normalization plus value cleaning and type inference.
- Creator — generates
CREATE TABLEDDL from a DataFrame's analyzed schema for any dialect. - Analyzer — finds unique columns/pairs, empty and categorical columns, and column metadata.
- Validator — pre-upload checks: extra columns, type mismatches, truncation detection.
- MsSqlUpdater — builds
MERGE/UPSERT/APPENDSQL between source and target tables.
Package documentation
Each subpackage has its own focused README:
src/etl/dataframe/readme.md— Analyzer, Cleaner, Parser across the Pandas/Polars/Spark backends.src/etl/database/readme.md— Connector, Loaders, Validator.src/etl/query/readme.md— Creator and MsSqlUpdater SQL generation.
Logging
Library modules log through standard logging.getLogger(__name__) loggers and emit no output
unless your application configures logging. The optional etl.logger.Logger singleton is a
convenience for applications: it attaches ANSI-colored dual stdout/stderr handlers.
License
MIT — see LICENSE.
1 maintainer
Metadata
1 maintainer
Release files for etl-utilities 2.0.1
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | |
|---|---|---|---|---|
| etl_utilities-2.0.1.tar.gz | 119.1 kB | With provenance | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| etl_utilities-2.0.1-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 184.8 kB
Release files / etl_utilities-2.0.1.tar.gz
| Download URL | etl_utilities-2.0.1.tar.gz |
|---|---|
| Size | 119.1 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
a0711675bcb33b6f4da4b56804994598a09990d8823d9bb6d315aadf8b989084
|
|
BLAKE2b-256 checksum How to use checksums |
db1884de88fa4dea0f2e54f96a40bfa148b53e289b9bd07fe473b4ad0f5382c5
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 21, 2026.
Transparency logRelease files / etl_utilities-2.0.1-py3-none-any.whl
| Download URL | etl_utilities-2.0.1-py3-none-any.whl |
|---|---|
| Size | 65.8 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
6132182ca062c119f345d9d099cf5ced207e699476d9d95a44cbdf6858344ce3
|
|
BLAKE2b-256 checksum How to use checksums |
03e013f95332eac93489093d7e5e8ac76bb79dac934621f9d760ccc30f833987
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 21, 2026.
Transparency logSecurity for etl-utilities 2.0.1
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
All 2 files in this release have been uploaded with provenance metadata.
Attestation for etl_utilities-2.0.1.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 21, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
