keyten

Python bindings for Keyten, a tiny columnar dataframe engine with compressed blocks and a lazy query optimizer.

Queries stay lazy until collect(), then run entirely in the engine with the GIL released — predicate and projection pushdown, vectorized parallel execution, disk spilling for data bigger than memory, and native-table I/O all happen outside the interpreter.

Install

Keyten currently supports 64-bit Linux and CPython 3.11 or newer:

pip install keyten

A first query

import keyten as kt

trades = kt.DataFrame([
    kt.Series.string("sym", ["a", "b", "a", "b"]),
    kt.Series.int("qty", [1, 2, 3, 4]),
])

result = (
    trades.lazy()
    .filter(kt.col("qty") > 1)
    .with_columns((kt.col("qty") * 10).alias("q10"))
    .group_by("sym")
    .agg([
        kt.col("q10").sum(),
        kt.col("qty").count().alias("n"),
    ])
    .sort("sym")
    .collect()
)

print(result)

What's in the box

• Expressions — arithmetic, comparisons, boolean logic, casts, string operations (substring, regex extract), value recoding, null/NaN shaping, and conditional if_else.
• Aggregations — sum/min/max/mean/count/first/last/std/median/n_unique and pairwise corr, per group or whole-frame.
• Windows — cum_sum, shift, diff, rolling_mean, forward_fill, and over(...) partitioned evaluation.
• Joins — inner, left, semi, anti, and a time-series backward asof join with by-groups.
• Temporal kinds — dates, timestamps, and times are first-class, parse straight from CSV, and ride the integer encodings.
• Sources — scan_csv (strict typed inference), scan_parquet (files or directories, statistics-pruned), and scan_native, the engine's own zero-copy, crash-safe table format.
• Scale — out-of-core execution against a self-derived memory budget; nothing to configure.

The package ships typed stubs (py.typed), so editors and type checkers see the full API with docstrings.

Documentation

Tutorials, the full API reference, and engine internals live at k10.works/docs.

License

Licensed under the MIT License.