fix: Read partitioned parquet files from relative paths

ludwig/data/dataset/ray.py

@@ -27,6 +27,7 @@
 import torch
 from packaging import version
 from pyarrow.fs import FSSpecHandler, PyFileSystem
+from pyarrow.lib import ArrowInvalid
 from ray.data import read_parquet
 from ray.data.dataset_pipeline import DatasetPipeline
 
@@ -55,7 +56,20 @@
 @default_retry()
 def read_remote_parquet(path: str):
     fs, path = get_fs_and_path(path)
-    return read_parquet(path, filesystem=PyFileSystem(FSSpecHandler(fs)))
+
+    # Fix for https://github.com/ludwig-ai/ludwig/issues/3440
+    # Parquet file reads will fail with `pyarrow.lib.ArrowInvalid` under the following conditions:
+    #     1) The Parquet data is in multi-file format
+    #     2) A relative filepath is passed to the read function
+    #     3) A filesystem object is passed to the read function
+    # The issue can be resolved by either:
+    #     1) Passing an absolute filepath
+    #     2) Not passing a filesystem object
+    try:
+        df = read_parquet(path, filesystem=PyFileSystem(FSSpecHandler(fs)))
+    except ArrowInvalid:
+        df = read_parquet(path)
+    return df
 
 
 @DeveloperAPI

tests/ludwig/data/test_ray_data.py

@@ -1,11 +1,15 @@
+import os
+import shutil
 from unittest import mock
 
+import pandas as pd
 import pytest
 
 # Skip these tests if Ray is not installed
 ray = pytest.importorskip("ray")  # noqa
+dask = pytest.importorskip("dask")  # noqa
 
-from ludwig.data.dataset.ray import RayDatasetBatcher  # noqa
+from ludwig.data.dataset.ray import RayDatasetBatcher, read_remote_parquet  # noqa
 
 # Mark the entire module as distributed
 pytestmark = pytest.mark.distributed
@@ -31,3 +35,57 @@ def test_async_reader_error():
             samples_per_epoch=100,
             ignore_last=False,
         )
+
+
+@pytest.fixture(scope="module")
+def parquet_file(ray_cluster_2cpu) -> str:
+    """Write a multi-file parquet dataset to the cwd.
+
+    Returns:
+        The path to the parquet dataset.
+    """
+    # The data needs to be written to a multi-file parquet format, otherwise the issue doesn't repro. To do this, we
+    # partitition a test dataframe with dask and then write to file.
+    df = pd.DataFrame({"col1": list(range(1000)), "col2": list(range(1000))})
+    df = dask.dataframe.from_pandas(df, chunksize=100)
+
+    # Typically we would write test data to a temporary directory, but the issue this was set up to test only happens
+    # when using relative filepaths.
+    cwd = os.getcwd()
+    filepath = os.path.join(cwd, "data.training.parquet")
+    df.to_parquet(filepath, engine="pyarrow")
+
+    yield filepath
+
+    # Clean up the data
+    shutil.rmtree(filepath)
+
+
+@pytest.fixture(scope="module", params=["absolute", "relative"])
+def parquet_filepath(parquet_file: str, request: "pytest.FixtureRequest") -> str:
+    """Convert a filepath in the CWD to either an absolute or relative path.
+
+    Args:
+        parquet_file: Absolute path to a parquet file in the CWD
+        request: pytest request fixture with the fixture parameters
+
+    Returns:
+        Either the absolute or relative path of the parquet file.
+    """
+    filepath_type = request.param
+    return parquet_file if filepath_type == "absolute" else os.path.basename(parquet_file)
+
+
+def test_read_remote_parquet(parquet_filepath: str):
+    """Test for the fix to https://github.com/ludwig-ai/ludwig/issues/3440.
+
+    Parquet file reads will fail with `pyarrow.lib.ArrowInvalid` under the following conditions:
+        1) The Parquet data is in multi-file format
+        2) A relative filepath is passed to the read function
+        3) A filesystem object is passed to the read function
+
+    The issue can be resolved by either:
+        1) Passing an absolute filepath
+        2) Not passing a filesystem object
+    """
+    read_remote_parquet(parquet_filepath)