Refactor expand_climatology to make it more scalable and to preserve spatial chunks

PiperOrigin-RevId: 575355512

Author: shoyer

scripts/expand_climatology.py

@@ -76,26 +76,33 @@
 
 
 def select_climatology(
-    time_slice: slice, climatology: xarray.Dataset, time_index: pd.DatetimeIndex
-) -> abc.Iterable[tuple[xbeam.Key, xarray.Dataset]]:
+    variable_name_and_time_slice: tuple[str, slice],
+    climatology: xarray.Dataset,
+    time_index: pd.DatetimeIndex,
+    base_chunks: dict[str, int],
+) -> abc.Iterator[tuple[xbeam.Key, xarray.Dataset]]:
   """Select climatology data matching time_index[time_slice]."""
+  variable_name, time_slice = variable_name_and_time_slice
   chunk_times = time_index[time_slice]
   times_array = xarray.DataArray(
       chunk_times, dims=['time'], coords={'time': chunk_times}
   )
   if 'hour' in climatology.coords:
-    chunk = climatology.sel(
+    sliced = climatology[[variable_name]].sel(
         dayofyear=times_array.dt.dayofyear, hour=times_array.dt.hour
     )
-    del chunk.coords['dayofyear']
-    del chunk.coords['hour']
+    del sliced.coords['dayofyear']
+    del sliced.coords['hour']
   else:
-    chunk = climatology.sel(dayofyear=times_array.dt.dayofyear)
-    del chunk.coords['dayofyear']
+    sliced = climatology[[variable_name]].sel(
+        dayofyear=times_array.dt.dayofyear
+    )
+    del sliced.coords['dayofyear']
 
-  for variable_name in chunk:
-    key = xbeam.Key({'time': time_slice.start}, vars={variable_name})  # pytype: disable=wrong-arg-types
-    yield key, chunk[[variable_name]]
+  key = xbeam.Key({'time': time_slice.start}, vars={variable_name})
+  sliced = sliced.compute()
+  target_chunks = {k: v for k, v in base_chunks.items() if k in sliced.dims}
+  yield from xbeam.split_chunks(key, sliced, target_chunks)
 
 
 def main(argv: list[str]) -> None:
@@ -124,8 +131,9 @@ def main(argv: list[str]) -> None:
     time_chunk_size = TIME_CHUNK_SIZE.value
 
   time_chunk_count = math.ceil(times.size / time_chunk_size)
-
-  output_chunks = {dim: -1 for dim in input_chunks if dim not in time_dims}
+  variables = list(climatology.keys())
+  base_chunks = {k: v for k, v in input_chunks.items() if k not in time_dims}
+  output_chunks = dict(base_chunks)
   output_chunks['time'] = time_chunk_size
 
   # Beam type checking is broken with Python 3.10:
@@ -137,7 +145,9 @@ def main(argv: list[str]) -> None:
         root
         | beam.Create([i * time_chunk_size for i in range(time_chunk_count)])
         | beam.Map(lambda start: slice(start, start + time_chunk_size))
-        | beam.FlatMap(select_climatology, climatology, times)
+        | beam.FlatMap(lambda index: [(v, index) for v in variables])
+        | beam.Reshuffle()
+        | beam.FlatMap(select_climatology, climatology, times, base_chunks)
         | xbeam.ChunksToZarr(
             OUTPUT_PATH.value, template=template, zarr_chunks=output_chunks
         )

scripts/expand_climatology_test.py

@@ -52,7 +52,7 @@ def test(self):
     input_path = self.create_tempdir('input_path').full_path
     output_path = self.create_tempdir('output_path').full_path
 
-    climatology.chunk({'dayofyear': 31}).to_zarr(input_path)
+    climatology.chunk({'dayofyear': 31, 'level': 1}).to_zarr(input_path)
 
     with flagsaver.flagsaver(
         input_path=input_path,
@@ -65,6 +65,8 @@ def test(self):
 
     actual = xarray.open_zarr(output_path)
     xarray.testing.assert_allclose(actual, expected)
+    self.assertEqual(actual.chunks['time'][0], 4 * 31)
+    self.assertEqual(actual.chunks['level'][0], 1)
 
 
 if __name__ == '__main__':