Use RangeSource in ReadDatasets

PiperOrigin-RevId: 827046303

Author: shoyer

xarray_beam/_src/core.py

@@ -27,14 +27,15 @@
 import immutabledict
 import numpy as np
 import xarray
+from xarray_beam._src import range_source
 from xarray_beam._src import threadmap
 
 
-T = TypeVar('T')
+T = TypeVar("T")
 
 
 def export(obj: T) -> T:
-  obj.__module__ = 'xarray_beam'
+  obj.__module__ = "xarray_beam"
   return obj
 
 
@@ -122,7 +123,6 @@ class Key:
     Key(indices={'x': 4}, vars={'bar'})
     >>> key.with_indices(x=5)
     Key(indices={'x': 5}, vars={'bar'})
-
   """
 
   # pylint: disable=redefined-builtin
@@ -184,8 +184,8 @@ def with_indices(self, **indices: int | None) -> Key:
     """Replace some indices with new values.
 
     Args:
-      **indices: indices to override (for integer values) or remove, with
-        values of ``None``.
+      **indices: indices to override (for integer values) or remove, with values
+        of ``None``.
 
     Returns:
       New Key with the specified indices.
@@ -421,49 +421,19 @@ def normalize_expanded_chunks(
 )
 
 
-@export
-class DatasetToChunks(beam.PTransform, Generic[DatasetOrDatasets]):
-  """Split one or more xarray.Datasets into keyed chunks."""
+class _DatasetToChunksBase(beam.PTransform, Generic[DatasetOrDatasets]):
+  """Base class for PTransforms that split Datasets into chunks."""
 
   def __init__(
       self,
       dataset: DatasetOrDatasets,
       chunks: Mapping[str, int | tuple[int, ...]] | None = None,
       split_vars: bool = False,
-      num_threads: int | None = None,
-      shard_keys_threshold: int = 200_000,
-      tasks_per_shard: int = 10_000,
   ):
-    """Initialize DatasetToChunks.
-
-    Args:
-      dataset: dataset or datasets to split into (Key, xarray.Dataset) or (Key,
-        [xarray.Dataset, ...]) pairs.
-      chunks: optional chunking scheme. Required if the dataset is *not* already
-        chunked. If the dataset *is* already chunked with Dask, `chunks` takes
-        precedence over the existing chunks.
-      split_vars: whether to split the dataset into separate records for each
-        data variable or to keep all data variables together. This is
-        recommended if you don't need to perform joint operations on different
-        dataset variables and individual variable chunks are sufficiently large.
-      num_threads: optional number of Dataset chunks to load in parallel per
-        worker. More threads can increase throughput, but also increases memory
-        usage and makes it harder for Beam runners to shard work. Note that each
-        variable in a Dataset is already loaded in parallel, so this is most
-        useful for Datasets with a small number of variables or when using
-        split_vars=True.
-      shard_keys_threshold: threshold at which to compute keys on Beam workers,
-        rather than only on the host process. This is important for scaling
-        pipelines to millions of tasks.
-      tasks_per_shard: number of tasks to emit per shard. Only used if the
-        number of tasks exceeds shard_keys_threshold.
-    """
+    """Initialize _DatasetToChunksBase."""
     self.dataset = dataset
     self._validate(dataset, split_vars)
     self.split_vars = split_vars
-    self.num_threads = num_threads
-    self.shard_keys_threshold = shard_keys_threshold
-    self.tasks_per_shard = tasks_per_shard
 
     if chunks is None:
       dask_chunks = self._first.chunks
@@ -542,6 +512,77 @@ def _task_count(self) -> int:
       total += int(np.prod(count_list))
     return total
 
+  def _key_to_chunks(self, key: Key) -> tuple[Key, DatasetOrDatasets]:
+    """Convert a Key into an in-memory (Key, xarray.Dataset) pair."""
+    with inc_timer_msec(self.__class__, "read-msec"):
+      sizes = {
+          dim: self.expanded_chunks[dim][self.offset_index[dim][offset]]
+          for dim, offset in key.offsets.items()
+      }
+      slices = offsets_to_slices(key.offsets, sizes)
+      results = []
+      for ds in self._datasets:
+        dataset = ds if key.vars is None else ds[list(key.vars)]
+        valid_slices = {k: v for k, v in slices.items() if k in dataset.dims}
+        chunk = dataset.isel(valid_slices)
+        # Load the data, using a separate thread for each variable
+        num_threads = len(dataset)
+        result = chunk.chunk().compute(num_workers=num_threads)
+        results.append(result)
+
+    inc_counter(self.__class__, "read-chunks")
+    inc_counter(
+        self.__class__, "read-bytes", sum(result.nbytes for result in results)
+    )
+
+    if isinstance(self.dataset, xarray.Dataset):
+      return key, results[0]
+    else:
+      return key, results
+
+
+@export
+class DatasetToChunks(_DatasetToChunksBase):
+  """Split one or more xarray.Datasets into keyed chunks."""
+
+  def __init__(
+      self,
+      dataset: DatasetOrDatasets,
+      chunks: Mapping[str, int | tuple[int, ...]] | None = None,
+      split_vars: bool = False,
+      num_threads: int | None = None,
+      shard_keys_threshold: int = 200_000,
+      tasks_per_shard: int = 10_000,
+  ):
+    """Initialize DatasetToChunks.
+
+    Args:
+      dataset: dataset or datasets to split into (Key, xarray.Dataset) or (Key,
+        [xarray.Dataset, ...]) pairs.
+      chunks: optional chunking scheme. Required if the dataset is *not* already
+        chunked. If the dataset *is* already chunked with Dask, `chunks` takes
+        precedence over the existing chunks.
+      split_vars: whether to split the dataset into separate records for each
+        data variable or to keep all data variables together. This is
+        recommended if you don't need to perform joint operations on different
+        dataset variables and individual variable chunks are sufficiently large.
+      num_threads: optional number of Dataset chunks to load in parallel per
+        worker. More threads can increase throughput, but also increases memory
+        usage and makes it harder for Beam runners to shard work. Note that each
+        variable in a Dataset is already loaded in parallel, so this is most
+        useful for Datasets with a small number of variables or when using
+        split_vars=True.
+      shard_keys_threshold: threshold at which to compute keys on Beam workers,
+        rather than only on the host process. This is important for scaling
+        pipelines to millions of tasks.
+      tasks_per_shard: number of tasks to emit per shard. Only used if the
+        number of tasks exceeds shard_keys_threshold.
+    """
+    super().__init__(dataset, chunks, split_vars)
+    self.num_threads = num_threads
+    self.shard_keys_threshold = shard_keys_threshold
+    self.tasks_per_shard = tasks_per_shard
+
   @cached_property
   def sharded_dim(self) -> str | None:
     # We use the simple heuristic of only sharding inputs along the dimension
@@ -610,34 +651,6 @@ def _shard_inputs(self) -> list[tuple[int | None, str | None]]:
         inputs.append((None, name))
     return inputs  # pytype: disable=bad-return-type  # always-use-property-annotation
 
-  def _key_to_chunks(self, key: Key) -> Iterator[tuple[Key, DatasetOrDatasets]]:
-    """Convert a Key into an in-memory (Key, xarray.Dataset) pair."""
-    with inc_timer_msec(self.__class__, "read-msec"):
-      sizes = {
-          dim: self.expanded_chunks[dim][self.offset_index[dim][offset]]
-          for dim, offset in key.offsets.items()
-      }
-      slices = offsets_to_slices(key.offsets, sizes)
-      results = []
-      for ds in self._datasets:
-        dataset = ds if key.vars is None else ds[list(key.vars)]
-        valid_slices = {k: v for k, v in slices.items() if k in dataset.dims}
-        chunk = dataset.isel(valid_slices)
-        # Load the data, using a separate thread for each variable
-        num_threads = len(dataset)
-        result = chunk.chunk().compute(num_workers=num_threads)
-        results.append(result)
-
-    inc_counter(self.__class__, "read-chunks")
-    inc_counter(
-        self.__class__, "read-bytes", sum(result.nbytes for result in results)
-    )
-
-    if isinstance(self.dataset, xarray.Dataset):
-      yield key, results[0]
-    else:
-      yield key, results
-
   def expand(self, pcoll):
     if self.shard_count is None:
       # Create all keys on the machine launching the Beam pipeline. This is
@@ -652,11 +665,95 @@ def expand(self, pcoll):
           | beam.Reshuffle()
       )
 
-    return key_pcoll | "KeyToChunks" >> threadmap.FlatThreadMap(
+    return key_pcoll | "KeyToChunks" >> threadmap.ThreadMap(
         self._key_to_chunks, num_threads=self.num_threads
     )
 
 
+# TODO(shoyer): expose this function as a public API, after switching it to
+# generate Key objects using `indices` instead of `offsets`.
+class ReadDataset(_DatasetToChunksBase):
+  """Read chunks from an xarray.Dataset into a Beam pipeline.
+
+  This PTransform is a Beam "splittable DoFn", which means that it may be
+  dynamically split by Beam runners into smaller chunks for efficient parallel
+  execution.
+  """
+
+  def __init__(
+      self,
+      dataset: xarray.Dataset,
+      chunks: Mapping[str, int | tuple[int, ...]] | None = None,
+      split_vars: bool = False,
+  ):
+    """Initialize ReadDatasets.
+
+    Args:
+      dataset: dataset to split into (Key, xarray.Dataset) chunks.
+      chunks: optional chunking scheme. Required if the dataset is *not* already
+        chunked. If the dataset *is* already chunked with Dask, `chunks` takes
+        precedence over the existing chunks.
+      split_vars: whether to split the dataset into separate records for each
+        data variable or to keep all data variables together. This is
+        recommended if you don't need to perform joint operations on different
+        dataset variables and individual variable chunks are sufficiently large.
+    """
+    super().__init__(dataset, chunks, split_vars)
+
+  @cached_property
+  def _var_chunk_counts(
+      self,
+  ) -> list[tuple[str | None, list[str], tuple[int, ...]]]:
+    out = []
+    if not self.split_vars:
+      dims = sorted(self.expanded_chunks)
+      shape = tuple(len(self.expanded_chunks[dim]) for dim in dims)
+      out.append((None, dims, shape))
+    else:
+      for name, variable in self._first.items():
+        dims = sorted([d for d in variable.dims if d in self.expanded_chunks])
+        shape = tuple(len(self.expanded_chunks[dim]) for dim in dims)
+        out.append((name, dims, shape))
+    return out  # pytype: disable=bad-return-type
+
+  @cached_property
+  def _var_sizes(self) -> list[int]:
+    return [int(np.prod(shape)) for _, _, shape in self._var_chunk_counts]
+
+  @cached_property
+  def _cumulative_sizes(self) -> np.ndarray:
+    return np.cumsum([0] + self._var_sizes)
+
+  def _index_to_key(self, position: int) -> Key:
+    assert 0 <= position < self._cumulative_sizes[-1]
+    var_index = (
+        np.searchsorted(self._cumulative_sizes, position, side="right") - 1
+    )
+    offset = position - self._cumulative_sizes[var_index]
+    name, dims, shape = self._var_chunk_counts[var_index]
+    indices = np.unravel_index(offset, shape)
+    offsets = {dim: self.offsets[dim][idx] for dim, idx in zip(dims, indices)}
+    return Key(offsets, vars=None if name is None else {name})
+
+  def _get_element(self, position: int) -> tuple[Key, xarray.Dataset]:
+    return self._key_to_chunks(self._index_to_key(position))  # pytype: disable=bad-return-type
+
+  def expand(
+      self, pbegin: beam.PBegin
+  ) -> beam.PCollection[tuple[Key, xarray.Dataset]]:
+    element_count = self._task_count()
+    assert element_count > 0
+    # For simplicity, assume that all chunks are approximately the same size,
+    # even if variables are being split and some variables have different
+    # variables. This assumption could be relaxed in the future, with an
+    # improved version of RangeSource.
+    avg_chunk_bytes = math.ceil(self._first.nbytes / element_count)
+    source = range_source.RangeSource(
+        element_count, avg_chunk_bytes, self._get_element
+    )
+    return pbegin | beam.io.Read(source)
+
+
 def _ensure_chunk_is_computed(key: Key, dataset: xarray.Dataset) -> None:
   """Ensure that a dataset contains no chunked variables."""
   for var_name, variable in dataset.variables.items():