Partial sweep= open renames groups by position, mislabelling elevations

[aladinor]

What happens

Opening a subset of sweeps names the returned groups by position, not by sweep identity. Requesting sweep=[0, 2, 8] returns sweep_0, sweep_1, sweep_2 — so the 25° tilt arrives named sweep_2, whose elevation in the full volume is 1.0°.

_attach_sweep_groups names from the enumeration index:

# xradar/io/backends/common.py:99-101
for i, sw in enumerate(sweeps):
    sw = sw.drop_vars(_STATION_VARS, errors="ignore").drop_attrs(deep=False)
    dtree[f"sweep_{i}"] = xr.DataTree(sw)

The sweep docstring says only "Sweep number(s) to extract" — nothing about renaming. The data is fine; each sweep still carries its true sweep_number, only the group name is wrong.

This matters because it fails silently. Writing a subset into an elevation-indexed schema (a fixed sweep→elevation template) files measurements at the wrong tilt with no error.

Reproducer

Dependencies inline (PEP 723) — save as mre.py, run uv run mre.py. Reads one public IRIS volume from the IDEAM open-data bucket anonymously; no malformed input involved — the same file opens perfectly with a full open.

# /// script
# requires-python = ">=3.11"
# dependencies = [
#     "xradar",
#     "s3fs",
# ]
# ///
import fsspec
import xradar
from xradar.io import open_iris_datatree

KEY = "s3-radaresideam/l2_data/2023/02/07/Tablazo/TAB230207000003.RAW3D38"
data = fsspec.filesystem("s3", anon=True).cat(KEY)

print(f"xradar {xradar.__version__}\n")


def show(tree, title):
    print(title)
    names = sorted(
        (k for k in tree.children if k.startswith("sweep_")),
        key=lambda s: int(s.split("_")[1]),
    )
    for name in names:
        ds = tree[name].to_dataset()
        print(
            f"  {name:9s} sweep_fixed_angle={float(ds['sweep_fixed_angle'].values):5.1f}"
            f"   sweep_number={int(ds['sweep_number'].values)}"
        )
    print()
    return {n: float(tree[n].to_dataset()["sweep_fixed_angle"].values) for n in names}


truth = show(open_iris_datatree(data), "full open (reference):")
subset = show(open_iris_datatree(data, sweep=[0, 2, 8]), "sweep=[0, 2, 8]:")

print("sweep_N name -> elevation, full open vs subset:")
for name in subset:
    ok = "ok" if truth[name] == subset[name] else "MISLABELLED"
    print(f"  {name:9s} {truth[name]:5.1f}  vs {subset[name]:5.1f}   {ok}")

Output on xradar 0.12.0:

xradar 0.12.0

full open (reference):
  sweep_0   sweep_fixed_angle=  0.0   sweep_number=0
  sweep_1   sweep_fixed_angle=  0.5   sweep_number=1
  sweep_2   sweep_fixed_angle=  1.0   sweep_number=2
  sweep_3   sweep_fixed_angle=  2.0   sweep_number=3
  sweep_4   sweep_fixed_angle=  4.5   sweep_number=4
  sweep_5   sweep_fixed_angle=  7.0   sweep_number=5
  sweep_6   sweep_fixed_angle= 10.0   sweep_number=6
  sweep_7   sweep_fixed_angle= 15.0   sweep_number=7
  sweep_8   sweep_fixed_angle= 25.0   sweep_number=8

sweep=[0, 2, 8]:
  sweep_0   sweep_fixed_angle=  0.0   sweep_number=0
  sweep_1   sweep_fixed_angle=  1.0   sweep_number=2
  sweep_2   sweep_fixed_angle= 25.0   sweep_number=8

sweep_N name -> elevation, full open vs subset:
  sweep_0     0.0  vs   0.0   ok
  sweep_1     0.5  vs   1.0   MISLABELLED
  sweep_2     1.0  vs  25.0   MISLABELLED

Expected

Group names preserve sweep identity — sweep=[0, 2, 8] yields sweep_0, sweep_2, sweep_8 — or, if positional naming is intentional, the docstring should say so, since the current behaviour is indistinguishable from correct output until someone checks sweep_number.

Secondary (same code path, lower priority)

A truncated sweep whose times fall back to cftime (dates before the 1582 reform) makes root assembly a mixed-type reduction:

# xradar/io/backends/common.py:139  in _assign_root
time_coverage_start = min(times)
TypeError: '>' not supported between instances of 'int' and 'cftime._cftime.DatetimeGregorian'

The input really is malformed there, so raising is defensible — but the message names neither the sweep nor the cause, and IrisArrayWrapper.__init__ (iris.py:3767) has already warned sweep_N empty or corrupted for that very sweep. Surfacing that instead would be a big diagnostic win. Encountered on 55 of 9,994 IDEAM Tablazo volumes, e.g. s3-radaresideam/l2_data/2023/02/19/Tablazo/TAB230219165503.RAW4WHB, whose sweep_7 holds one ray stamped year 0001 while sweeps 0-6 and 8 are clean.

Working around that one by re-opening with only the good sweeps is what surfaced the naming bug above.

Versions

xradar 0.12.0 (also 0.12.1.dev9+gd723857f2), Python 3.12, Linux.