polars-io-tools 0.1.0

Custom parsing extensions for lazy polars

polars io tools

Custom parsing extensions for lazy polars

Overview

polars-io-tools extends Polars lazy execution with custom I/O sources that push filters and column projections all the way down into the systems that hold your data — SQL databases, ClickHouse, Datadog, and Delta Lake — instead of loading everything and filtering in memory. It also adds lazy-friendly operations (joins, multi-source composition, time-series windows, caching, distributed execution) that keep predicate pushdown working where vanilla Polars would otherwise give up and materialize the whole frame.

Everything is exposed through the piot LazyFrame namespace and a handful of top-level scan_* / sink_* functions, so it composes naturally with the Polars API you already use.

Who is this for

Reach for polars-io-tools when you want Polars' lazy API over data that lives in an external store, and you care about not fetching rows or columns you will immediately throw away. It is most valuable for large, partitioned, or remote datasets where a filter on a date or key column should translate into a smaller query against the source. If your data already fits comfortably in memory or lives in local Parquet/CSV, plain Polars is the simpler choice.

Installation

pip install polars-io-tools

polars-io-tools requires Python 3.11 or newer. See the Installation guide for conda and source builds.

Quickstart

Importing the package registers the piot namespace on every Polars LazyFrame:

import polars as pl
import polars_io_tools  # registers the .piot namespace

left = pl.LazyFrame({"x": [1, 2, 3], "y": [4, 5, 6]})
right = pl.LazyFrame({"x": [-1, -2, 3], "z": [7, 8, 9]})

# An inner join where the keys present on the left are pushed down as a
# filter on the right frame *before* the join runs.
result = left.piot.filtered_join(right, on="x").collect()
print(result)
# shape: (1, 3)
# ┌─────┬─────┬─────┐
# │ x   ┆ y   ┆ z   │
# │ --- ┆ --- ┆ --- │
# │ i64 ┆ i64 ┆ i64 │
# ╞═════╪═════╪═════╡
# │ 3   ┆ 6   ┆ 9   │
# └─────┴─────┴─────┘

For a guided walkthrough, start with the Getting Started tutorial.

What's included

• Lazy I/O sources with predicate & projection pushdown — scan_db (any ODBC database), scan_clickhouse, scan_datadog, scan_delta, and from_narwhals. Filters on the resulting LazyFrame are translated into the source's own query language (SQL WHERE, Datadog time ranges, Delta partition pruning) so only the matching rows and columns are fetched.
• Lazy writers — sink_delta and sink_clickhouse write a LazyFrame directly to Delta Lake or ClickHouse, including streaming/chunked writes and transparent handling of types the target store cannot represent natively.
• Pushdown-preserving query building — filtered_join, filtered_join_asof, join_between, multi_source, concat_named, and ts_with_columns express joins, multi-source composition, and rolling/lookback time-series logic without blocking the filter pushdown that those operations normally defeat.
• Caching — cache keeps an in-memory, column- and partition-level cache for iterative work; cache_parquet materializes date-partitioned Parquet on local disk or S3, fetching only the partitions a query needs.
• Distributed execution — execute_on_ray splits a LazyFrame by calendar period and runs the partitions across an existing Ray cluster.
• Ergonomics — iter_rows for memory-efficient row iteration, debug to inspect what Polars pushes into a source, and disable_optimizations to compare against plain Polars.

Documentation

Full documentation lives in the project wiki:

• Getting Started — a guided tutorial.
• Reading and Writing Data — recipes for each I/O source and sink.
• Query Optimization — joins, composition, time-series, and caching.
• Distributed Execution — running on Ray.
• API Reference — the public functions and piot namespace.
• Concepts — how predicate pushdown into custom sources works.

Contributing

Contributions are welcome. See the Contributing guide and Local Development Setup to get started.

License

polars-io-tools is licensed under the Apache 2.0 license.