Reduce memory footprint of P2P shuffling (#8157)

Author: hendrikmakait

continuous_integration/environment-3.9.yaml

@@ -28,7 +28,7 @@ dependencies:
   - pre-commit
   - prometheus_client
   - psutil
-  - pyarrow=7
+  - pyarrow=12
   - pynvml  # Only tested here
   - pytest
   - pytest-cov

distributed/shuffle/_arrow.py

@@ -1,10 +1,12 @@
 from __future__ import annotations
 
-from io import BytesIO
-from typing import TYPE_CHECKING
+from pathlib import Path
+from typing import TYPE_CHECKING, Any
 
 from packaging.version import parse
 
+from dask.utils import parse_bytes
+
 if TYPE_CHECKING:
     import pandas as pd
     import pyarrow as pa
@@ -29,34 +31,27 @@ def check_minimal_arrow_version() -> None:
     """Verify that the the correct version of pyarrow is installed to support
     the P2P extension.
 
-    Raises a RuntimeError in case pyarrow is not installed or installed version
-    is not recent enough.
+    Raises a ModuleNotFoundError if pyarrow is not installed or an
+    ImportError if the installed version is not recent enough.
     """
-    # First version to introduce Table.sort_by
-    minversion = "7.0.0"
+    # First version that supports concatenating extension arrays (apache/arrow#14463)
+    minversion = "12.0.0"
     try:
         import pyarrow as pa
-    except ImportError:
-        raise RuntimeError(f"P2P shuffling requires pyarrow>={minversion}")
-
+    except ModuleNotFoundError:
+        raise ModuleNotFoundError(f"P2P shuffling requires pyarrow>={minversion}")
     if parse(pa.__version__) < parse(minversion):
-        raise RuntimeError(
+        raise ImportError(
             f"P2P shuffling requires pyarrow>={minversion} but only found {pa.__version__}"
         )
 
 
-def convert_partition(data: bytes, meta: pd.DataFrame) -> pd.DataFrame:
+def convert_shards(shards: list[pa.Table], meta: pd.DataFrame) -> pd.DataFrame:
     import pyarrow as pa
 
     from dask.dataframe.dispatch import from_pyarrow_table_dispatch
 
-    file = BytesIO(data)
-    end = len(data)
-    shards = []
-    while file.tell() < end:
-        sr = pa.RecordBatchStreamReader(file)
-        shards.append(sr.read_all())
-    table = pa.concat_tables(shards, promote=True)
+    table = pa.concat_tables(shards)
 
     df = from_pyarrow_table_dispatch(meta, table, self_destruct=True)
     return df.astype(meta.dtypes, copy=False)
@@ -66,9 +61,7 @@ def list_of_buffers_to_table(data: list[bytes]) -> pa.Table:
     """Convert a list of arrow buffers and a schema to an Arrow Table"""
     import pyarrow as pa
 
-    return pa.concat_tables(
-        (deserialize_table(buffer) for buffer in data), promote=True
-    )
+    return pa.concat_tables(deserialize_table(buffer) for buffer in data)
 
 
 def serialize_table(table: pa.Table) -> bytes:
@@ -85,3 +78,42 @@ def deserialize_table(buffer: bytes) -> pa.Table:
 
     with pa.ipc.open_stream(pa.py_buffer(buffer)) as reader:
         return reader.read_all()
+
+
+def read_from_disk(path: Path, meta: pd.DataFrame) -> tuple[Any, int]:
+    import pyarrow as pa
+
+    from dask.dataframe.dispatch import pyarrow_schema_dispatch
+
+    batch_size = parse_bytes("1 MiB")
+    batch = []
+    shards = []
+    schema = pyarrow_schema_dispatch(meta, preserve_index=True)
+
+    with pa.OSFile(str(path), mode="rb") as f:
+        size = f.seek(0, whence=2)
+        f.seek(0)
+        prev = 0
+        offset = f.tell()
+        while offset < size:
+            sr = pa.RecordBatchStreamReader(f)
+            shard = sr.read_all()
+            offset = f.tell()
+            batch.append(shard)
+
+            if offset - prev >= batch_size:
+                table = pa.concat_tables(batch)
+                shards.append(_copy_table(table, schema))
+                batch = []
+                prev = offset
+    if batch:
+        table = pa.concat_tables(batch)
+        shards.append(_copy_table(table, schema))
+    return shards, size
+
+
+def _copy_table(table: pa.Table, schema: pa.Schema) -> pa.Table:
+    import pyarrow as pa
+
+    arrs = [pa.concat_arrays(column.chunks) for column in table.columns]
+    return pa.table(data=arrs, schema=schema)

distributed/shuffle/_core.py

@@ -11,6 +11,7 @@
 from dataclasses import dataclass, field
 from enum import Enum
 from functools import partial
+from pathlib import Path
 from typing import TYPE_CHECKING, Any, Generic, NewType, TypeVar
 
 from distributed.core import PooledRPCCall
@@ -62,6 +63,7 @@ def __init__(
 
         self._disk_buffer = DiskShardsBuffer(
             directory=directory,
+            read=self.read,
             memory_limiter=memory_limiter_disk,
         )
 
@@ -180,10 +182,9 @@ def fail(self, exception: Exception) -> None:
         if not self.closed:
             self._exception = exception
 
-    def _read_from_disk(self, id: NDIndex) -> bytes:
+    def _read_from_disk(self, id: NDIndex) -> list[Any]:  # TODO: Typing
         self.raise_if_closed()
-        data: bytes = self._disk_buffer.read("_".join(str(i) for i in id))
-        return data
+        return self._disk_buffer.read("_".join(str(i) for i in id))
 
     async def receive(self, data: list[tuple[_T_partition_id, bytes]]) -> None:
         await self._receive(data)
@@ -238,6 +239,10 @@ async def _get_output_partition(
     ) -> _T_partition_type:
         """Get an output partition to the shuffle run"""
 
+    @abc.abstractmethod
+    def read(self, path: Path) -> tuple[Any, int]:
+        """Read shards from disk"""
+
 
 def get_worker_plugin() -> ShuffleWorkerPlugin:
     from distributed import get_worker

distributed/shuffle/_disk.py

@@ -3,6 +3,7 @@
 import contextlib
 import pathlib
 import shutil
+from typing import Any, Callable
 
 from distributed.shuffle._buffer import ShardsBuffer
 from distributed.shuffle._limiter import ResourceLimiter
@@ -41,6 +42,7 @@ class DiskShardsBuffer(ShardsBuffer):
     def __init__(
         self,
         directory: str | pathlib.Path,
+        read: Callable[[pathlib.Path], tuple[Any, int]],
         memory_limiter: ResourceLimiter | None = None,
     ):
         super().__init__(
@@ -50,6 +52,7 @@ def __init__(
         )
         self.directory = pathlib.Path(directory)
         self.directory.mkdir(exist_ok=True)
+        self._read = read
 
     async def _process(self, id: str, shards: list[bytes]) -> None:
         """Write one buffer to file
@@ -74,19 +77,15 @@ async def _process(self, id: str, shards: list[bytes]) -> None:
                     for shard in shards:
                         f.write(shard)
 
-    def read(self, id: int | str) -> bytes:
+    def read(self, id: int | str) -> Any:
         """Read a complete file back into memory"""
         self.raise_on_exception()
         if not self._inputs_done:
             raise RuntimeError("Tried to read from file before done.")
 
         try:
             with self.time("read"):
-                with open(
-                    self.directory / str(id), mode="rb", buffering=100_000_000
-                ) as f:
-                    data = f.read()
-                    size = f.tell()
+                data, size = self._read((self.directory / str(id)).resolve())
         except FileNotFoundError:
             raise KeyError(id)
 

distributed/shuffle/_rechunk.py

@@ -102,8 +102,8 @@
 from collections.abc import Callable, Sequence
 from concurrent.futures import ThreadPoolExecutor
 from dataclasses import dataclass
-from io import BytesIO
 from itertools import product
+from pathlib import Path
 from typing import TYPE_CHECKING, Any, NamedTuple
 
 import dask
@@ -258,26 +258,22 @@ def split_axes(old: ChunkedAxes, new: ChunkedAxes) -> SplitAxes:
     return axes
 
 
-def convert_chunk(data: bytes) -> np.ndarray:
+def convert_chunk(shards: list[tuple[NDIndex, np.ndarray]]) -> np.ndarray:
     import numpy as np
 
     from dask.array.core import concatenate3
 
-    file = BytesIO(data)
-    shards: dict[NDIndex, np.ndarray] = {}
+    indexed: dict[NDIndex, np.ndarray] = {}
+    for index, shard in shards:
+        indexed[index] = shard
+    del shards
 
-    while file.tell() < len(data):
-        for index, shard in pickle.load(file):
-            shards[index] = shard
-
-    subshape = [max(dim) + 1 for dim in zip(*shards.keys())]
-    assert len(shards) == np.prod(subshape)
+    subshape = [max(dim) + 1 for dim in zip(*indexed.keys())]
+    assert len(indexed) == np.prod(subshape)
 
     rec_cat_arg = np.empty(subshape, dtype="O")
-    for index, shard in shards.items():
+    for index, shard in indexed.items():
         rec_cat_arg[tuple(index)] = shard
-    del data
-    del file
     arrs = rec_cat_arg.tolist()
     return concatenate3(arrs)
 
@@ -427,8 +423,20 @@ def _() -> dict[str, tuple[NDIndex, bytes]]:
     async def _get_output_partition(
         self, partition_id: NDIndex, key: str, **kwargs: Any
     ) -> np.ndarray:
-        data = self._read_from_disk(partition_id)
-        return await self.offload(convert_chunk, data)
+        def _(partition_id: NDIndex) -> np.ndarray:
+            data = self._read_from_disk(partition_id)
+            return convert_chunk(data)
+
+        return await self.offload(_, partition_id)
+
+    def read(self, path: Path) -> tuple[Any, int]:
+        shards: list[tuple[NDIndex, np.ndarray]] = []
+        with path.open(mode="rb") as f:
+            size = f.seek(0, os.SEEK_END)
+            f.seek(0)
+            while f.tell() < size:
+                shards.extend(pickle.load(f))
+        return shards, size
 
     def _get_assigned_worker(self, id: NDIndex) -> str:
         return self.worker_for[id]

distributed/shuffle/_shuffle.py

@@ -7,6 +7,7 @@
 from concurrent.futures import ThreadPoolExecutor
 from dataclasses import dataclass
 from functools import partial
+from pathlib import Path
 from typing import TYPE_CHECKING, Any
 
 import toolz
@@ -20,8 +21,9 @@
 from distributed.shuffle._arrow import (
     check_dtype_support,
     check_minimal_arrow_version,
-    convert_partition,
+    convert_shards,
     list_of_buffers_to_table,
+    read_from_disk,
     serialize_table,
 )
 from distributed.shuffle._core import (
@@ -321,7 +323,7 @@ def split_by_worker(
     return out
 
 
-def split_by_partition(t: pa.Table, column: str) -> dict[Any, pa.Table]:
+def split_by_partition(t: pa.Table, column: str) -> dict[int, pa.Table]:
     """
     Split data into many arrow batches, partitioned by final partition
     """
@@ -383,6 +385,11 @@ class DataFrameShuffleRun(ShuffleRun[int, "pd.DataFrame"]):
         buffer.
     """
 
+    column: str
+    meta: pd.DataFrame
+    partitions_of: dict[str, list[int]]
+    worker_for: pd.Series
+
     def __init__(
         self,
         worker_for: dict[int, str],
@@ -476,16 +483,22 @@ async def _get_output_partition(
         **kwargs: Any,
     ) -> pd.DataFrame:
         try:
-            data = self._read_from_disk((partition_id,))
 
-            out = await self.offload(convert_partition, data, self.meta)
+            def _(partition_id: int, meta: pd.DataFrame) -> pd.DataFrame:
+                data = self._read_from_disk((partition_id,))
+                return convert_shards(data, meta)
+
+            out = await self.offload(_, partition_id, self.meta)
         except KeyError:
             out = self.meta.copy()
         return out
 
     def _get_assigned_worker(self, id: int) -> str:
         return self.worker_for[id]
 
+    def read(self, path: Path) -> tuple[Any, int]:
+        return read_from_disk(path, self.meta)
+
 
 @dataclass(frozen=True)
 class DataFrameShuffleSpec(ShuffleSpec[int]):