Multi-file and Parquet-aware prefetching from remote storage

python/cudf/cudf/io/parquet.py

@@ -577,11 +577,51 @@ def read_parquet(
         )
     filepath_or_buffer = paths if paths else filepath_or_buffer
 
+    # Prepare remote-IO options
+    prefetch_options = kwargs.pop("prefetch_options", {})
+    if not ioutils._is_local_filesystem(fs):
+        # The default prefetch method depends on the
+        # `row_groups` argument. In most cases we will use
+        # method="all" by default, because it is fastest
+        # when we need to read most of the file(s).
+        # If a (simple) `row_groups` selection is made, we
+        # use method="parquet" to avoid transferring the
+        # entire file over the network
+        method = prefetch_options.get("method")
+        _row_groups = None
+        if method in (None, "parquet"):
+            if row_groups is None:
+                # Don't use 'parquet' prefetcher for column
+                # projection alone.
+                method = method or "all"
+            elif all(r == row_groups[0] for r in row_groups):
+                # Row group selection means we are probably
+                # reading half the file or less. We should
+                # avoid a full file transfer by default.
+                method = "parquet"
+                _row_groups = row_groups[0]
+            elif (method := method or "all") == "parquet":
+                raise ValueError(
+                    "The 'parquet' prefetcher requires a uniform "
+                    "row-group selection for all paths within the "
+                    "same `read_parquet` call. "
+                    "Got: {row_groups}"
+                )
+        if method == "parquet":
+            prefetch_options = prefetch_options.update(
+                {
+                    "method": method,
+                    "columns": columns,
+                    "row_groups": _row_groups,
+                }
+            )
+
     filepaths_or_buffers = ioutils.get_reader_filepath_or_buffer(
         path_or_data=filepath_or_buffer,
         fs=fs,
         storage_options=storage_options,
         bytes_per_thread=bytes_per_thread,
+        prefetch_options=prefetch_options,
     )
 
     # Warn user if they are not using cudf for IO

python/cudf/cudf/tests/test_s3.py

@@ -229,6 +229,53 @@ def test_read_parquet(
     assert_eq(expect, got2)
 
 
+@pytest.mark.parametrize("method", ["all", "parquet"])
+@pytest.mark.parametrize("blocksize", [1024 * 1024, 1024])
+def test_read_parquet_prefetch_options(
+    s3_base,
+    s3so,
+    pdf,
+    method,
+    blocksize,
+):
+    bucket = "parquet"
+    fname_1 = "test_parquet_reader_prefetch_options_1.parquet"
+    buffer_1 = BytesIO()
+    pdf.to_parquet(path=buffer_1)
+    buffer_1.seek(0)
+
+    fname_2 = "test_parquet_reader_prefetch_options_2.parquet"
+    buffer_2 = BytesIO()
+    pdf_2 = pdf.copy()
+    pdf_2["Integer"] += 1
+    pdf_2.to_parquet(path=buffer_2)
+    buffer_2.seek(0)
+
+    with s3_context(
+        s3_base=s3_base,
+        bucket=bucket,
+        files={
+            fname_1: buffer_1,
+            fname_2: buffer_2,
+        },
+    ):
+        got = cudf.read_parquet(
+            [
+                f"s3://{bucket}/{fname_1}",
+                f"s3://{bucket}/{fname_2}",
+            ],
+            storage_options=s3so,
+            prefetch_options={
+                "method": method,
+                "blocksize": blocksize,
+            },
+            columns=["String", "Integer"],
+        )
+
+    expect = pd.concat([pdf, pdf_2], ignore_index=True)[["String", "Integer"]]
+    assert_eq(expect, got)
+
+
 @pytest.mark.parametrize("bytes_per_thread", [32, 1024])
 @pytest.mark.parametrize("columns", [None, ["List", "Struct"]])
 @pytest.mark.parametrize("index", [None, "Integer"])

python/cudf/cudf/utils/ioutils.py

@@ -1,6 +1,8 @@
 # Copyright (c) 2019-2024, NVIDIA CORPORATION.
 
 import datetime
+import functools
+import operator
 import os
 import urllib
 import warnings
@@ -18,6 +20,12 @@
 from cudf.core._compat import PANDAS_LT_300
 from cudf.utils.docutils import docfmt_partial
 
+try:
+    import fsspec.parquet as fsspec_parquet
+
+except ImportError:
+    fsspec_parquet = None
+
 _BYTES_PER_THREAD_DEFAULT = 256 * 1024 * 1024
 _ROW_GROUP_SIZE_BYTES_DEFAULT = 128 * 1024 * 1024
 
@@ -187,6 +195,10 @@
 allow_mismatched_pq_schemas : boolean, default False
     If True, enables reading (matching) columns specified in `columns` and `filters`
     options from the input files with otherwise mismatched schemas.
+prefetch_options : dict, default None
+    WARNING: This is an experimental feature (added in 24.10).
+    Dictionary of options to use to prefetch bytes from remote storage.
+    These options are passed through to `get_reader_filepath_or_buffer`.
 
 Returns
 -------
@@ -1439,6 +1451,13 @@
     Glob pattern to use when expanding directories into file paths
     (e.g. "*.json"). If this parameter is not specified, directories
     will not be expanded.
+prefetch_options : dict, default None
+    WARNING: This is an experimental feature (added in 24.10).
+    Dictionary of options to use to prefetch bytes from remote storage.
+    These options are only used when `path_or_data` is a list of remote
+    paths. If 'method' is set to 'all' (the default), the only supported
+    option is 'blocksize' (default 256 MB). If method is set to 'parquet',
+    'columns' and 'row_groups' are also supported (default None).
 
 Returns
 -------
@@ -1620,6 +1639,7 @@ def get_reader_filepath_or_buffer(
     warn_on_raw_text_input=None,
     warn_meta=None,
     expand_dir_pattern=None,
+    prefetch_options=None,
 ):
     """{docstring}"""
 
@@ -1690,26 +1710,15 @@ def get_reader_filepath_or_buffer(
                 raw_text_input = True
 
         elif fs is not None:
-            # TODO: We can use cat_ranges and/or parquet-aware logic
-            # to copy all remote data into host memory at once here.
-            # The current solution iterates over files, and copies
-            # ALL data from each file (even when we are performing
-            # partial IO, and don't need the entire file)
             if len(paths) == 0:
                 raise FileNotFoundError(
                     f"{input_sources} could not be resolved to any files"
                 )
-            filepaths_or_buffers = [
-                BytesIO(
-                    _fsspec_data_transfer(
-                        fpath,
-                        fs=fs,
-                        mode=mode,
-                        bytes_per_thread=bytes_per_thread,
-                    )
-                )
-                for fpath in paths
-            ]
+            filepaths_or_buffers = _prefetch_remote_buffers(
+                paths,
+                fs,
+                **(prefetch_options or {}),
+            )
         else:
             raw_text_input = True
 
@@ -2099,3 +2108,100 @@ def _read_byte_ranges(
 
     for worker in workers:
         worker.join()
+
+
+def _get_remote_bytes_all(
+    remote_paths, fs, *, blocksize=_BYTES_PER_THREAD_DEFAULT
+):
+    # TODO: Experiment with a heuristic to avoid the fs.sizes
+    # call when we are reading many files at once (the latency
+    # of collecting the file sizes is unnecessary in this case)
+    if max(sizes := fs.sizes(remote_paths)) <= blocksize:
+        # Don't bother breaking up individual files
+        return fs.cat_ranges(remote_paths, None, None)
+    else:
+        # Construct list of paths, starts, and ends
+        paths, starts, ends = map(
+            list,
+            zip(
+                *(
+                    (r, j, min(j + blocksize, s))
+                    for r, s in zip(remote_paths, sizes)
+                    for j in range(0, s, blocksize)
+                )
+            ),
+        )
+
+        # Collect the byte ranges
+        chunks = fs.cat_ranges(paths, starts, ends)
+
+        # Construct local byte buffers
+        # (Need to make sure path offsets are ordered correctly)
+        unique_count = dict(zip(*np.unique(paths, return_counts=True)))
+        offset = np.cumsum([0] + [unique_count[p] for p in remote_paths])
+        buffers = [
+            functools.reduce(operator.add, chunks[offset[i] : offset[i + 1]])
+            for i in range(len(remote_paths))
+        ]
+        return buffers
+
+
+def _get_remote_bytes_parquet(
+    remote_paths,
+    fs,
+    *,
+    columns=None,
+    row_groups=None,
+    blocksize=_BYTES_PER_THREAD_DEFAULT,
+):
+    if fsspec_parquet is None or (columns is None and row_groups is None):
+        return _get_remote_bytes_all(remote_paths, fs, blocksize=blocksize)
+
+    sizes = fs.sizes(remote_paths)
+    data = fsspec_parquet._get_parquet_byte_ranges(
+        remote_paths,
+        fs,
+        columns=columns,
+        row_groups=row_groups,
+        max_block=blocksize,
+    )
+
+    buffers = []
+    for size, path in zip(sizes, remote_paths):
+        path_data = data[path]
+        buf = np.empty(size, dtype="b")
+        for range_offset in path_data.keys():
+            chunk = path_data[range_offset]
+            buf[range_offset[0] : range_offset[1]] = np.frombuffer(
+                chunk, dtype="b"
+            )
+        buffers.append(buf.tobytes())
+    return buffers
+
+
+def _prefetch_remote_buffers(
+    paths,
+    fs,
+    *,
+    method="all",
+    **prefetch_options,
+):
+    # Gather bytes ahead of time for remote filesystems
+    if fs and paths and not _is_local_filesystem(fs):
+        try:
+            prefetcher = {
+                "parquet": _get_remote_bytes_parquet,
+                "all": _get_remote_bytes_all,
+            }[method]
+        except KeyError:
+            raise NotImplementedError(
+                f"{method} is not a supported remote-data prefetcher"
+            )
+        return prefetcher(
+            paths,
+            fs,
+            **prefetch_options,
+        )
+
+    else:
+        return paths