tart-msv4

An xarray backend that opens TART (Transient Array Radio Telescope) visibility HDF5 files as Measurement Set v4 (MSv4) xarray.DataTrees. Its output passes xradio.schema.check.check_datatree, and it follows the conventions of xarray-ms.

The only runtime dependencies are numpy, xarray, h5py and pyerfa. It needs no astropy, no network and no dask (dask is supported when installed).

Install

pip install tart-msv4

Usage

import xarray as xr

dt = xr.open_datatree("vis_2026-06-09_08_11_43.476804.hdf", engine="tart")

xr.open_dataset(path, engine="tart") returns just the visibility partition. engine="tart" accepts a single file only. Always pass engine="tart". The backend's guessing claims only HDF5 files that carry the TART signature (vis, baselines, timestamp, and a JSON config with num_antenna), but xarray tries its netCDF engines first, and h5netcdf or netCDF4, when installed, claim every HDF5 file.

Every array on the (time, baseline) grid is lazy: VISIBILITY is read from the file, and FLAG, WEIGHT, UVW, TIME_CENTROID and EFFECTIVE_INTEGRATION_TIME are computed on read from small per-file tables. Opening a file reads only its metadata, and the opened tree pickles small. xarray's CF decoding keywords (decode_times etc.) are accepted and ignored.

Several files

TART writes one file per minute or so. To open the files of one observation as a single partition, lazily concatenated along time:

from pathlib import Path
import tart_msv4

paths = sorted(Path("data").glob("vis_*.hdf"))
dt = tart_msv4.open_tart(paths)  # lazy, no dask
dt = tart_msv4.open_tart(paths, chunks={})  # dask, one chunk per file

• Inputs: open_tart takes one path or a sequence of paths. You expand globs and directories yourself.
• Ordering: files are sorted by timestamp. Files that overlap in time, or repeat, raise ValueError.
• Compatibility: config and baselines must be identical. antenna_positions may differ by up to position_tolerance (default 1 cm), because TART recalibrates them along with its gains. Any other mismatch raises ValueError.

Keyword arguments

The keywords below work with both xr.open_datatree and open_tart:

Schema

/
└── <partition>                     visibility       (time, baseline_id, frequency, polarization, uvw_label)
    ├── antenna_xds                 antenna          (antenna_name, cartesian_pos_label, receptor_label)
    ├── field_and_source_base_xds   field_and_source (field_name, sky_dir_label)
    └── gain_xds                    TART extension   (gain_time, antenna_name, enu_label)

gain_xds has type="extension:gain.tart_msv4" and tart_msv4_extension=True. check_datatree skips it with an ExtensionTypeWarning.

Conventions

• Fields. TART points at the zenith and does not fringe-stop, so every snapshot is its own field. Fields are named zenith_<UTC ISO µs> (e.g. zenith_2026-06-09T08:10:42.993515), which keeps names unique across files. All fields share one partition, as in xradio and xarray-ms. Each file is one scan (scan_<first snapshot ISO>). See ADR 1.
• Visibilities need no rephasing. They are already phased to each snapshot's zenith. They are passed through unconjugated, and are identical to the tart2ms MS DATA at a shared phase centre.
• UVW = pos(ant1) − pos(ant2), the sign tart2ms writes. It is checked against the reference MS (ADR 2). Each ITRS baseline is rotated into the celestial frame at its snapshot (IAU 2006/2000A, erfa.c2t06a) and projected onto the (u, v, w) basis of that snapshot's phase centre.
• Phase centres come from erfa.atoc13, which converts observed (az, zd) = (0, 0) to ICRS with aberration and no refraction. They match astropy to under 0.5″; the difference is polar motion, which is neglected. They match casacore/tart2ms to about 0.25″.
• Time. Timestamps are parsed as ISO 8601. Timestamps without a timezone are UTC, never machine-local. time is the TART timestamp exactly, as tart2ms writes it.
• UT1. No IERS data is used, so by default UT1 = UTC. An error of Δ seconds shifts right ascensions by about 15Δ″ (|Δ| < 0.9 s), which is negligible against TART's ~7° resolution. Pass dut1= for better.
• Antenna positions. ENU offsets are placed on the exact WGS84 tangent plane. tart2ms uses a spherical approximation, so its u and v differ by up to about 2 cm.

Out of scope (TODO)

• Rephasing all snapshots to one common phase centre. That is a processing step and belongs in the consumer (e.g. kremetart). Its result should be a separate field_and_source data group.
• Applying gains.
• Satellite catalogues.
• IERS-based UT1 and polar motion.
• Writing back to HDF5.

Development

uv sync                       # dev + test groups
uv run ruff format . && uv run ruff check .
uv run pytest                 # downloads the real-data bundle once into tests/data

The real-data bundle is the kremetart test bundle (nine TART files and tart2ms reference MSs). Set TART_MSV4_OFFLINE=1 to skip the download, or TART_MSV4_DATA=<dir> to use an existing copy. Tests marked integration skip without it. Domain terms are defined in GLOSSARY.md.

Pre-commit hooks

Install the hooks once per clone:

uv sync
uv run pre-commit install    # installs the pre-commit and commit-msg hooks

If a hook modifies files, the commit is aborted so you can review the changes. Run git add -u && git commit again. Run all hooks on demand with uv run pre-commit run --all-files.

Releasing

Commit messages are the changelog. CHANGELOG.md is generated from them by git-cliff (cliff.toml) at release time, so don't edit it by hand. Releases use tbump:

uv run tbump 0.2.0

tbump updates the version in pyproject.toml and src/tart_msv4/__init__.py, generates CHANGELOG.md, refreshes uv.lock, commits chore(release): bump version to 0.2.0, tags v0.2.0 and pushes. Pushing the tag publishes the package to PyPI.