Use Ray iter_batches instead of to_arrow_refs

python/src/space/ray/ops/join.py

@@ -29,7 +29,7 @@
                                      row_id_field_name)
 import space.core.transform.utils as transform_utils
 from space.core.utils import errors
-from space.ray.ops.utils import singleton_storage
+from space.ray.ops.utils import iter_batches, singleton_storage
 from space.ray.options import RayOptions
 
 if TYPE_CHECKING:
@@ -138,14 +138,7 @@ def _join(left: _JoinInputInternal, right: _JoinInputInternal, join_key: str,
 
 
 def _read_all(ds: ray.data.Dataset) -> Optional[pa.Table]:
-  results = []
-  for ref in ds.to_arrow_refs():
-    data = ray.get(ref)
-    if data is None or data.num_rows == 0:
-      continue
-
-    results.append(data)
-
+  results = list(iter_batches(ds))
   if not results:
     return None
 

python/src/space/ray/ops/utils.py

@@ -15,7 +15,10 @@
 """Utilities for Ray operations."""
 
 from __future__ import annotations
-from typing import TYPE_CHECKING
+from typing import Iterator, TYPE_CHECKING
+
+import ray
+import pyarrow as pa
 
 from space.core.storage import Storage
 from space.core.utils import errors
@@ -35,3 +38,13 @@ def singleton_storage(view: View) -> Storage:
     raise errors.UserInputError("Joining results of joins is not supported")
 
   return list(view.sources.values())[0].storage
+
+
+def iter_batches(ds: ray.data.Dataset) -> Iterator[pa.Table]:
+  """Return an iterator of PyArrow tables from a Ray dataset."""
+  # batch_size is None to use entire Ray blocks.
+  for data in ds.iter_batches(batch_size=None,
+                              batch_format="pyarrow",
+                              drop_last=False):
+    if data.num_rows > 0:
+      yield data

python/src/space/ray/runners.py

@@ -21,7 +21,6 @@
 
 import pyarrow as pa
 import pyarrow.compute as pc
-import ray
 
 from space.core.jobs import JobResult
 from space.core.loaders.array_record import ArrayRecordIndexFn
@@ -41,7 +40,7 @@
 from space.ray.ops.change_data import read_change_data
 from space.ray.ops.delete import RayDeleteOp
 from space.ray.ops.insert import RayInsertOp
-from space.ray.ops.utils import singleton_storage
+from space.ray.ops.utils import iter_batches, singleton_storage
 from space.ray.options import RayOptions
 
 if TYPE_CHECKING:
@@ -83,9 +82,8 @@ def read(
     read_options = ReadOptions(filter_, fields, snapshot_id, reference_read,
                                batch_size)
 
-    for ref in self._view.ray_dataset(self._ray_options, read_options,
-                                      join_options).to_arrow_refs():
-      yield ray.get(ref)
+    return iter_batches(
+        self._view.ray_dataset(self._ray_options, read_options, join_options))
 
   def diff(self,
            start_version: Union[Version],
@@ -102,11 +100,8 @@ def diff(self,
       # TODO: skip processing the data for deletions; the caller is usually
       # only interested at deleted primary keys.
       # TODO: to split change data into chunks for parallel processing.
-      processed_remote_data = self._view.process_source(change.data)
-      for ref in processed_remote_data.to_arrow_refs():
-        data = ray.get(ref)
-        if data.num_rows > 0:
-          yield ChangeData(change.snapshot_id, change.type_, data)
+      for data in iter_batches(self._view.process_source(change.data)):
+        yield ChangeData(change.snapshot_id, change.type_, data)
 
   @property
   def _source_storage(self) -> Storage:
@@ -150,10 +145,8 @@ def read(
                    self._storage.version_to_snapshot_id(version))
     read_options = ReadOptions(filter_, fields, snapshot_id, reference_read,
                                batch_size)
-
-    for ref in self._storage.ray_dataset(self._ray_options,
-                                         read_options).to_arrow_refs():
-      yield ray.get(ref)
+    return iter_batches(
+        self._storage.ray_dataset(self._ray_options, read_options))
 
   def refresh(self,
               target_version: Optional[Version] = None,