polars-checkpoint

Materialise Polars LazyFrames to parquet files and scan them back lazily. Defaults to the streaming engine for sink/scan. Useful for managing & reducing memory pressure due to expensive intermediate results in complex multi-step transforms.

Installation

Requires polars and Python 3.10+.

Quick Start

import polars as pl
from polars_checkpoint import checkpoint

lf = pl.LazyFrame({"x": range(1_000_000)}).with_columns(y=pl.col("x") * 2)

# Materialises to a temp parquet file; returns a lazy re-scan
lf = checkpoint(lf)

lf.filter(pl.col("y") > 100).collect()

A process-wide default session manages the temp directory and cleans it up at exit.

Session API

For explicit control over storage location and lifecycle, use CheckpointSession:

from polars_checkpoint import CheckpointSession

# As a context manager — cleans up on exit from the block
with CheckpointSession(root_dir="./my_checkpoints") as sess:
    lf = pl.scan_csv("big.csv")
    lf = sess.checkpoint(lf, name="after-parse")
    lf = lf.filter(pl.col("status") == "active")
    lf = sess.checkpoint(lf, name="filtered")

# Without a context manager — cleans up at GC or interpreter shutdown,
# or when you call close() explicitly
sess = CheckpointSession(root_dir="./my_checkpoints")
lf = sess.checkpoint(pl.scan_csv("big.csv"), name="raw")
reloaded = sess["raw"]
print(sess.summary())

sess.close()  # optional; triggers early cleanup

CheckpointSession constructor

Key methods & features

• checkpoint(lf, *, name=None, streaming=True, ...) — Materialise a LazyFrame to parquet. Auto-generates a name if none given. Falls back to collect().write_parquet() when streaming=False.
• session[name] — Retrieve a checkpoint as a LazyFrame.
• name in session — Check existence.
• len(session) / iter(session) — Count / list checkpoints.
• summary() — Returns a Polars DataFrame with name, size (MB), and path of each checkpoint.
• close(timeout=None) — Waits for in-flight writes, then cleans up. Also usable as a context manager.

Thread Safety

Sessions are internally locked. Concurrent checkpoint() calls from multiple threads are safe; close() waits for all in-flight materialisations before removing files.

Cleanup Behaviour

When root_dir is auto-generated, the entire temp directory is removed. When user-supplied, only the individual checkpoint subdirectories created by the session are removed.