tet-py

Python bindings for Tetration — mmap-friendly .tet tensors and the JSON/TOML query engine.

Quick start

pip install tet-py
python -c "import tet; print(tet.__version__, tet.core_version())"

import tet

with tet.open("data.tet") as f:
    print(f.mean("temperature"), f.quantile("temperature", 0.5))
    arr = f.read_numpy("temperature")                 # ram (memmap for full-width raw slabs)
    z = f.transform.to_numpy.zscore("temperature")  # transform → ndarray

More examples: Python quick start · operations · NumPy

Convert

tet.convert writes one .tet from a foreign file. .npy, multi-member .npz, and .csv / .tsv are in the core wheel. Other formats are optional extras: tet-py[h5], [netcdf], [zarr], [parquet], or [convert] for all of them.

src may be s3://bucket/object.ext (full download, then the same importer) or an s3:// Zarr v2/v3 prefix (each chunk is fetched as the zarr library reads it). Credentials stay in TETRATION_S3_*.

Compared with Rust tet convert: Python does not apply CF scale/offset, chunks the whole array unless you pass chunk_shape, and can read codecs the Rust importer skips when the extra library can decode them. Columns that are not layout-v1 numeric (or category text) are omitted.

Export

tet.export (and TetFile.export) writes a .tet to HDF5, NetCDF, or Zarr. Same optional extras as convert ([h5], [netcdf], [zarr], or [export] for those three). Dense numeric datasets only; carries dim_names and basic string attrs when present. Values are densified via read_numpy — prefer Rust tet export when you need native HDF5/NetCDF or chunk-preserving Zarr.

Packages

tet.open("pkg.tet.d") reads the generation named by CURRENT (packed.tet). tet.ingest copies a sealed .tet into staging, and tet.publish seals those parts and updates CURRENT. tet.pack(package, "out.tet") copies that sealed image to a standalone file; pass remove_package=True to delete the .tet.d after a verified copy.

For an S3-compatible backend:

from tet import S3Store

store = S3Store.connect(endpoint="http://127.0.0.1:9000", bucket="…",
                        access_key_id="…", secret_access_key="…", prefix="demo")
store.ingest("part.tet", session_id="s0")
store.publish(["part"])
f = store.open("current")  # pin; later publish does not retarget f
# or, with TETRATION_S3_* set: tet.open("s3://bucket/demo")

Endpoint and credentials are arguments or TETRATION_S3_* (not in the URI). tet.open("s3://bucket/prefix") range-reads the pinned package. S3Store.open still downloads packed.tet. See docs/operations.md.

Documentation

Develop

uv sync --extra dev
mise run develop          # or: uv run maturin develop
uv run pytest -q
uv run mypy python/tet

Requires Python 3.11+ and Rust 1.95+ (.mise.toml). Fixtures: tests/fixtures/.

License

MIT OR Apache-2.0, same as tetration.