to_dataset: without ORDER BY, coordinates come back in the engine's nondeterministic row order #276

to_dataset builds each dimension's coordinate in the order its values first appear in the result. That keeps an ORDER BY level DESC in the Dataset, which is the intent (see the comment in xarray_sql/ds.py, where pd.unique builds coord_arrays). But without an ORDER BY, SQL leaves the row order unspecified. DuckDB's parallel scans return rows in a different order on each run, so the same query gives Datasets whose coordinates are in different orders, and rarely in the template's order.

Repro

import numpy as np, pandas as pd, xarray as xr, duckdb
import xarray_sql as xql

ds = xr.Dataset(
    {"temp": (("time", "lat", "lon"), np.arange(12.0).reshape(3, 2, 2) + 10)},
    coords={"time": pd.date_range("2024-01-01", periods=3),
            "lat": [10.0, 20.0], "lon": [100.0, 110.0]},
)
con = duckdb.connect()
xql.register(con, "weather", ds)
for _ in range(3):
    rel = con.sql("SELECT lat, lon, MAX(temp) AS temp FROM weather GROUP BY lat, lon")
    out = xql.to_dataset(rel, template=ds)
    print(out.lat.values, out.lon.values)

Output (xarray-sql main, DuckDB 1.5.5):

[20. 10.] [100. 110.]
[10. 20.] [110. 100.]
[20. 10.] [110. 100.]

Labels are right (out.temp.sel(lat=20.0, lon=110.0) is 21.0 on every run), so nothing is wrong cell by cell. But anything positional is affected:

• .values;
• .isel;
• plots, which draw the axis in that order;
• xr.testing.assert_equal against the source;
• concatenating results from two runs.

That's surprising for an Xarray user, since Xarray coordinates from a store are almost always monotonic.

Expected

When the result's order says nothing about a dimension, its coordinate should come back in a deterministic order. Two options:

• Follow the template's order for that dimension. The template is usually ascending, and this makes assert_equal against the source work.
• Sort ascending when there's no template.

A result whose rows are monotonic along a dimension (an ORDER BY, including DESC) would keep that order, as now.

One way to decide: if each dimension's first-appearance coordinate is monotonic, keep it; otherwise sort it, or reindex it to the template. The scatter path already argsorts, so the cost is in choosing coord_arrays, not in placing values.

Workaround

Add ORDER BY on the dimensions. The Performance guide already recommends it for speed ("Round-trip faster with ORDER BY"), but not for determinism.