"source" encoding for datasets opened from fsspec objects (#8923)

* draft for setting `source` from pre-opened `fsspec` file objects

* refactor to only import `fsspec` if we're actually going to check

Could use `getattr(filename_or_obj, "path", filename_or_obj)` to avoid
`isinstance` checks.

* replace with a simple `getattr` on `"path"`

* add a test

* whats-new entry

* open the file as a context manager

Author: keewis

doc/whats-new.rst

@@ -24,6 +24,8 @@ New Features
 ~~~~~~~~~~~~
 - Allow chunking for arrays with duplicated dimension names (:issue:`8759`, :pull:`9099`).
   By `Martin Raspaud <https://github.com/mraspaud>`_.
+- Extract the source url from fsspec objects (:issue:`9142`, :pull:`8923`).
+  By `Justus Magin <https://github.com/keewis>`_.
 
 Breaking changes
 ~~~~~~~~~~~~~~~~

xarray/backends/api.py

@@ -382,8 +382,11 @@ def _dataset_from_backend_dataset(
     ds.set_close(backend_ds._close)
 
     # Ensure source filename always stored in dataset object
-    if "source" not in ds.encoding and isinstance(filename_or_obj, (str, os.PathLike)):
-        ds.encoding["source"] = _normalize_path(filename_or_obj)
+    if "source" not in ds.encoding:
+        path = getattr(filename_or_obj, "path", filename_or_obj)
+
+        if isinstance(path, (str, os.PathLike)):
+            ds.encoding["source"] = _normalize_path(path)
 
     return ds
 

xarray/tests/test_backends.py

@@ -5151,6 +5151,21 @@ def test_source_encoding_always_present_with_pathlib() -> None:
             assert ds.encoding["source"] == tmp
 
 
+@requires_h5netcdf
+@requires_fsspec
+def test_source_encoding_always_present_with_fsspec() -> None:
+    import fsspec
+
+    rnddata = np.random.randn(10)
+    original = Dataset({"foo": ("x", rnddata)})
+    with create_tmp_file() as tmp:
+        original.to_netcdf(tmp)
+
+        fs = fsspec.filesystem("file")
+        with fs.open(tmp) as f, open_dataset(f) as ds:
+            assert ds.encoding["source"] == tmp
+
+
 def _assert_no_dates_out_of_range_warning(record):
     undesired_message = "dates out of range"
     for warning in record: