Fix reads from local dir that changes directory

dask_expr/_collection.py

@@ -3,6 +3,7 @@
 import datetime
 import functools
 import inspect
+import os
 import warnings
 from collections.abc import Callable, Hashable, Mapping
 from numbers import Integral, Number
@@ -4526,6 +4527,7 @@ def read_csv(
             storage_options=storage_options,
             kwargs=kwargs,
             header=header,
+            _cwd=_get_cwd(path, kwargs),
         )
     )
 
@@ -4551,6 +4553,7 @@ def read_table(
             storage_options=storage_options,
             kwargs=kwargs,
             header=header,
+            _cwd=_get_cwd(path, kwargs),
         )
     )
 
@@ -4576,10 +4579,25 @@ def read_fwf(
             storage_options=storage_options,
             kwargs=kwargs,
             header=header,
+            _cwd=_get_cwd(path, kwargs),
         )
     )
 
 
+def _get_protocol(urlpath):
+    if "://" in urlpath:
+        protocol, _ = urlpath.split("://", 1)
+        if len(protocol) > 1:
+            # excludes Windows paths
+            return protocol
+    return None
+
+
+def _get_cwd(path, kwargs):
+    protocol = kwargs.pop("protocol", None) or _get_protocol(path) or "file"
+    return os.getcwd() if protocol == "file" else None
+
+
 def read_parquet(
     path=None,
     columns=None,
@@ -4630,6 +4648,7 @@ def read_parquet(
             filesystem=filesystem,
             engine=_set_parquet_engine(engine),
             kwargs=kwargs,
+            _cwd=_get_cwd(path, kwargs),
             _series=isinstance(columns, str),
         )
     )

dask_expr/io/csv.py

@@ -14,6 +14,7 @@ class ReadCSV(PartitionsFiltered, BlockwiseIO):
         "_partitions",
         "storage_options",
         "kwargs",
+        "_cwd",  # needed for tokenization
         "_series",
     ]
     _defaults = {
@@ -24,6 +25,7 @@ class ReadCSV(PartitionsFiltered, BlockwiseIO):
         "_partitions": None,
         "storage_options": None,
         "_series": False,
+        "_cwd": None,
     }
     _absorb_projections = True
 

dask_expr/io/parquet.py

@@ -426,6 +426,7 @@ class ReadParquet(PartitionsFiltered, BlockwiseIO):
         "filesystem",
         "engine",
         "kwargs",
+        "_cwd",  # needed for tokenization
         "_partitions",
         "_series",
         "_dataset_info_cache",
@@ -449,6 +450,7 @@ class ReadParquet(PartitionsFiltered, BlockwiseIO):
         "_partitions": None,
         "_series": False,
         "_dataset_info_cache": None,
+        "_cwd": None,
     }
     _pq_length_stats = None
     _absorb_projections = True

dask_expr/io/tests/test_io.py

@@ -1,5 +1,6 @@
 import glob
 import os
+from pathlib import Path
 
 import dask.array as da
 import dask.dataframe as dd
@@ -30,14 +31,14 @@
 pd = _backend_library()
 
 
-def _make_file(dir, format="parquet", df=None):
+def _make_file(dir, format="parquet", df=None, **kwargs):
     fn = os.path.join(str(dir), f"myfile.{format}")
     if df is None:
         df = pd.DataFrame({c: range(10) for c in "abcde"})
     if format == "csv":
-        df.to_csv(fn)
+        df.to_csv(fn, **kwargs)
     elif format == "parquet":
-        df.to_parquet(fn)
+        df.to_parquet(fn, **kwargs)
     else:
         ValueError(f"{format} not a supported format")
     return fn
@@ -413,6 +414,33 @@ def test_combine_similar_no_projection_on_one_branch(tmpdir):
     assert_eq(df, pdf)
 
 
+@pytest.mark.parametrize(
+    "fmt, func, kwargs",
+    [
+        ("parquet", read_parquet, {}),
+        ("csv", read_csv, {"index": False}),
+    ],
+)
+def test_chdir_different_files(tmpdir, fmt, func, kwargs):
+    cwd = os.getcwd()
+
+    try:
+        pdf = pd.DataFrame({"x": [0, 1, 2, 3] * 4, "y": range(16)})
+        os.chdir(tmpdir)
+        _make_file(tmpdir, format=fmt, df=pdf, **kwargs)
+        df = func(f"myfile.{fmt}")
+
+        new_dir = Path(tmpdir).joinpath("new_dir")
+        new_dir.mkdir()
+        os.chdir(new_dir)
+        pdf2 = pd.DataFrame({"x": [0, 100, 200, 300] * 4, "y": range(16)})
+        _make_file(new_dir, format=fmt, df=pdf2, **kwargs)
+        df2 = func(f"myfile.{fmt}")
+        assert_eq(df.sum() + df2.sum(), pd.Series([2424, 240], index=["x", "y"]))
+    finally:
+        os.chdir(cwd)
+
+
 @pytest.mark.parametrize("meta", [True, False])
 @pytest.mark.parametrize("label", [None, "foo"])
 @pytest.mark.parametrize("allow_projection", [True, False])