Major fix to subset_to_blocks (#173)

1. Copy and extend `dask.array.blocks.__getitem__` to support orthogonal indexing. 
    This means each cohort is a single layer in the graph.
2. Significantly extend cohorts.py benchmarks

Author: dcherian

asv_bench/benchmarks/cohorts.py

@@ -29,7 +29,22 @@ def track_num_tasks(self):
         )[0]
         return len(result.dask.to_dict())
 
+    def track_num_tasks_optimized(self):
+        result = flox.groupby_reduce(
+            self.array, self.by, func="sum", axis=self.axis, method="cohorts"
+        )[0]
+        (opt,) = dask.optimize(result)
+        return len(opt.dask.to_dict())
+
+    def track_num_layers(self):
+        result = flox.groupby_reduce(
+            self.array, self.by, func="sum", axis=self.axis, method="cohorts"
+        )[0]
+        return len(result.dask.layers)
+
     track_num_tasks.unit = "tasks"
+    track_num_tasks_optimized.unit = "tasks"
+    track_num_layers.unit = "layers"
 
 
 class NWMMidwest(Cohorts):
@@ -45,16 +60,68 @@ def setup(self, *args, **kwargs):
         self.axis = (-2, -1)
 
 
-class ERA5(Cohorts):
+class ERA5Dataset:
     """ERA5"""
 
+    def __init__(self, *args, **kwargs):
+        self.time = pd.Series(pd.date_range("2016-01-01", "2018-12-31 23:59", freq="H"))
+        self.axis = (-1,)
+        self.array = dask.array.random.random((721, 1440, len(self.time)), chunks=(-1, -1, 48))
+
+    def rechunk(self):
+        self.array = flox.core.rechunk_for_cohorts(
+            self.array, -1, self.by, force_new_chunk_at=[1], chunksize=48, ignore_old_chunks=True
+        )
+
+
+class ERA5DayOfYear(ERA5Dataset, Cohorts):
+    def setup(self, *args, **kwargs):
+        super().__init__()
+        self.by = self.time.dt.dayofyear.values
+
+
+class ERA5DayOfYearRechunked(ERA5DayOfYear, Cohorts):
+    def setup(self, *args, **kwargs):
+        super().setup()
+        super().rechunk()
+
+
+class ERA5MonthHour(ERA5Dataset, Cohorts):
     def setup(self, *args, **kwargs):
-        time = pd.Series(pd.date_range("2016-01-01", "2018-12-31 23:59", freq="H"))
+        super().__init__()
+        by = (self.time.dt.month.values, self.time.dt.hour.values)
+        ret = flox.core._factorize_multiple(
+            by,
+            expected_groups=(pd.Index(np.arange(1, 13)), pd.Index(np.arange(1, 25))),
+            by_is_dask=False,
+            reindex=False,
+        )
+        # Add one so the rechunk code is simpler and makes sense
+        self.by = ret[0][0] + 1
 
-        self.by = time.dt.dayofyear.values
+
+class ERA5MonthHourRechunked(ERA5MonthHour, Cohorts):
+    def setup(self, *args, **kwargs):
+        super().setup()
+        super().rechunk()
+
+
+class PerfectMonthly(Cohorts):
+    """Perfectly chunked for a "cohorts" monthly mean climatology"""
+
+    def setup(self, *args, **kwargs):
+        self.time = pd.Series(pd.date_range("1961-01-01", "2018-12-31 23:59", freq="M"))
         self.axis = (-1,)
+        self.array = dask.array.random.random((721, 1440, len(self.time)), chunks=(-1, -1, 4))
+        self.by = self.time.dt.month.values
 
-        array = dask.array.random.random((721, 1440, len(time)), chunks=(-1, -1, 48))
+    def rechunk(self):
         self.array = flox.core.rechunk_for_cohorts(
-            array, -1, self.by, force_new_chunk_at=[1], chunksize=48, ignore_old_chunks=True
+            self.array, -1, self.by, force_new_chunk_at=[1], chunksize=4, ignore_old_chunks=True
         )
+
+
+class PerfectMonthlyRechunked(PerfectMonthly):
+    def setup(self, *args, **kwargs):
+        super().setup()
+        super().rechunk()

flox/core.py

@@ -6,6 +6,7 @@
 import operator
 from collections import namedtuple
 from functools import partial, reduce
+from numbers import Integral
 from typing import TYPE_CHECKING, Any, Callable, Dict, Literal, Mapping, Sequence, Union
 
 import numpy as np
@@ -288,7 +289,7 @@ def rechunk_for_cohorts(
     divisions = []
     counter = 1
     for idx, lab in enumerate(labels):
-        if lab in force_new_chunk_at:
+        if lab in force_new_chunk_at or idx == 0:
             divisions.append(idx)
             counter = 1
             continue
@@ -305,6 +306,7 @@ def rechunk_for_cohorts(
             divisions.append(idx)
             counter = 1
             continue
+
         counter += 1
 
     divisions.append(len(labels))
@@ -313,6 +315,9 @@ def rechunk_for_cohorts(
         print(labels_at_breaks[:40])
 
     newchunks = tuple(np.diff(divisions))
+    if debug:
+        print(divisions[:10], newchunks[:10])
+        print(divisions[-10:], newchunks[-10:])
     assert sum(newchunks) == len(labels)
 
     if newchunks == array.chunks[axis]:
@@ -1046,26 +1051,18 @@ def _reduce_blockwise(
     return result
 
 
-def subset_to_blocks(
-    array: DaskArray, flatblocks: Sequence[int], blkshape: tuple[int] | None = None
-) -> DaskArray:
+def _normalize_indexes(array, flatblocks, blkshape):
     """
-    Advanced indexing of .blocks such that we always get a regular array back.
+    .blocks accessor can only accept one iterable at a time,
+    but can handle multiple slices.
+    To minimize tasks and layers, we normalize to produce slices
+    along as many axes as possible, and then repeatedly apply
+    any remaining iterables in a loop.
 
-    Parameters
-    ----------
-    array : dask.array
-    flatblocks : flat indices of blocks to extract
-    blkshape : shape of blocks with which to unravel flatblocks
-
-    Returns
-    -------
-    dask.array
+    TODO: move this upstream
     """
-    if blkshape is None:
-        blkshape = array.blocks.shape
-
     unraveled = np.unravel_index(flatblocks, blkshape)
+
     normalized: list[Union[int, np.ndarray, slice]] = []
     for ax, idx in enumerate(unraveled):
         i = _unique(idx).squeeze()
@@ -1077,30 +1074,65 @@ def subset_to_blocks(
             elif np.array_equal(i, np.arange(i[0], i[-1] + 1)):
                 normalized.append(slice(i[0], i[-1] + 1))
             else:
-                normalized.append(i)
+                normalized.append(list(i))
     full_normalized = (slice(None),) * (array.ndim - len(normalized)) + tuple(normalized)
 
     # has no iterables
-    noiter = tuple(i if not hasattr(i, "__len__") else slice(None) for i in full_normalized)
+    noiter = list(i if not hasattr(i, "__len__") else slice(None) for i in full_normalized)
     # has all iterables
-    alliter = {
-        ax: i if hasattr(i, "__len__") else slice(None) for ax, i in enumerate(full_normalized)
-    }
+    alliter = {ax: i for ax, i in enumerate(full_normalized) if hasattr(i, "__len__")}
 
-    # apply everything but the iterables
-    if all(i == slice(None) for i in noiter):
+    mesh = dict(zip(alliter.keys(), np.ix_(*alliter.values())))
+
+    full_tuple = tuple(i if ax not in mesh else mesh[ax] for ax, i in enumerate(noiter))
+
+    return full_tuple
+
+
+def subset_to_blocks(
+    array: DaskArray, flatblocks: Sequence[int], blkshape: tuple[int] | None = None
+) -> DaskArray:
+    """
+    Advanced indexing of .blocks such that we always get a regular array back.
+
+    Parameters
+    ----------
+    array : dask.array
+    flatblocks : flat indices of blocks to extract
+    blkshape : shape of blocks with which to unravel flatblocks
+
+    Returns
+    -------
+    dask.array
+    """
+    import dask.array
+    from dask.array.slicing import normalize_index
+    from dask.base import tokenize
+    from dask.highlevelgraph import HighLevelGraph
+
+    if blkshape is None:
+        blkshape = array.blocks.shape
+
+    index = _normalize_indexes(array, flatblocks, blkshape)
+
+    if all(not isinstance(i, np.ndarray) and i == slice(None) for i in index):
         return array
 
-    subset = array.blocks[noiter]
+    # These rest is copied from dask.array.core.py with slight modifications
+    index = normalize_index(index, array.numblocks)
+    index = tuple(slice(k, k + 1) if isinstance(k, Integral) else k for k in index)
 
-    for ax, inds in alliter.items():
-        if isinstance(inds, slice):
-            continue
-        idxr = [slice(None, None)] * array.ndim
-        idxr[ax] = inds
-        subset = subset.blocks[tuple(idxr)]
+    name = "blocks-" + tokenize(array, index)
+    new_keys = array._key_array[index]
+
+    squeezed = tuple(np.squeeze(i) if isinstance(i, np.ndarray) else i for i in index)
+    chunks = tuple(tuple(np.array(c)[i].tolist()) for c, i in zip(array.chunks, squeezed))
+
+    keys = itertools.product(*(range(len(c)) for c in chunks))
+    layer = {(name,) + key: tuple(new_keys[key].tolist()) for key in keys}
+    graph = HighLevelGraph.from_collections(name, layer, dependencies=[array])
 
-    return subset
+    return dask.array.Array(graph, name, chunks, meta=array)
 
 
 def _extract_unknown_groups(reduced, group_chunks, dtype) -> tuple[DaskArray]:

tests/__init__.py

@@ -115,6 +115,18 @@ def assert_equal(a, b, tolerance=None):
         np.testing.assert_allclose(a, b, equal_nan=True, **tolerance)
 
 
+def assert_equal_tuple(a, b):
+    """assert_equal for .blocks indexing tuples"""
+    assert len(a) == len(b)
+
+    for a_, b_ in zip(a, b):
+        assert type(a_) == type(b_)
+        if isinstance(a_, np.ndarray):
+            np.testing.assert_array_equal(a_, b_)
+        else:
+            assert a_ == b_
+
+
 @pytest.fixture(scope="module", params=["flox", "numpy", "numba"])
 def engine(request):
     if request.param == "numba":

tests/test_core.py

@@ -12,14 +12,23 @@
 from flox.core import (
     _convert_expected_groups_to_index,
     _get_optimal_chunks_for_groups,
+    _normalize_indexes,
     factorize_,
     find_group_cohorts,
     groupby_reduce,
     rechunk_for_cohorts,
     reindex_,
+    subset_to_blocks,
 )
 
-from . import assert_equal, engine, has_dask, raise_if_dask_computes, requires_dask
+from . import (
+    assert_equal,
+    assert_equal_tuple,
+    engine,
+    has_dask,
+    raise_if_dask_computes,
+    requires_dask,
+)
 
 labels = np.array([0, 0, 2, 2, 2, 1, 1, 2, 2, 1, 1, 0])
 nan_labels = labels.astype(float)  # copy
@@ -1035,3 +1044,84 @@ def test_dtype(func, dtype, engine):
     labels = np.array(["a", "a", "c", "c", "c", "b", "b", "c", "c", "b", "b", "f"])
     actual, _ = groupby_reduce(arr, labels, func=func, dtype=np.float64)
     assert actual.dtype == np.dtype("float64")
+
+
+@requires_dask
+def test_subset_blocks():
+    array = dask.array.random.random((120,), chunks=(4,))
+
+    blockid = (0, 3, 6, 9, 12, 15, 18, 21, 24, 27)
+    subset = subset_to_blocks(array, blockid)
+    assert subset.blocks.shape == (len(blockid),)
+
+
+@requires_dask
+@pytest.mark.parametrize(
+    "flatblocks, expected",
+    (
+        ((0, 1, 2, 3, 4), (slice(None),)),
+        ((1, 2, 3), (slice(1, 4),)),
+        ((1, 3), ([1, 3],)),
+        ((0, 1, 3), ([0, 1, 3],)),
+    ),
+)
+def test_normalize_block_indexing_1d(flatblocks, expected):
+    nblocks = 5
+    array = dask.array.ones((nblocks,), chunks=(1,))
+    expected = tuple(np.array(i) if isinstance(i, list) else i for i in expected)
+    actual = _normalize_indexes(array, flatblocks, array.blocks.shape)
+    assert_equal_tuple(expected, actual)
+
+
+@requires_dask
+@pytest.mark.parametrize(
+    "flatblocks, expected",
+    (
+        ((0, 1, 2, 3, 4), (0, slice(None))),
+        ((1, 2, 3), (0, slice(1, 4))),
+        ((1, 3), (0, [1, 3])),
+        ((0, 1, 3), (0, [0, 1, 3])),
+        (tuple(range(10)), (slice(0, 2), slice(None))),
+        ((0, 1, 3, 5, 6, 8), (slice(0, 2), [0, 1, 3])),
+        ((0, 3, 4, 5, 6, 8, 24), np.ix_([0, 1, 4], [0, 1, 3, 4])),
+    ),
+)
+def test_normalize_block_indexing_2d(flatblocks, expected):
+    nblocks = 5
+    ndim = 2
+    array = dask.array.ones((nblocks,) * ndim, chunks=(1,) * ndim)
+    expected = tuple(np.array(i) if isinstance(i, list) else i for i in expected)
+    actual = _normalize_indexes(array, flatblocks, array.blocks.shape)
+    assert_equal_tuple(expected, actual)
+
+
+@requires_dask
+def test_subset_block_passthrough():
+    # full slice pass through
+    array = dask.array.ones((5,), chunks=(1,))
+    subset = subset_to_blocks(array, np.arange(5))
+    assert subset.name == array.name
+
+    array = dask.array.ones((5, 5), chunks=1)
+    subset = subset_to_blocks(array, np.arange(25))
+    assert subset.name == array.name
+
+
+@requires_dask
+@pytest.mark.parametrize(
+    "flatblocks, expectidx",
+    [
+        (np.arange(10), (slice(2), slice(None))),
+        (np.arange(8), (slice(2), slice(None))),
+        ([0, 10], ([0, 2], slice(1))),
+        ([0, 7], (slice(2), [0, 2])),
+        ([0, 7, 9], (slice(2), [0, 2, 4])),
+        ([0, 6, 12, 14], (slice(3), [0, 1, 2, 4])),
+        ([0, 12, 14, 19], np.ix_([0, 2, 3], [0, 2, 4])),
+    ],
+)
+def test_subset_block_2d(flatblocks, expectidx):
+    array = dask.array.from_array(np.arange(25).reshape((5, 5)), chunks=1)
+    subset = subset_to_blocks(array, flatblocks)
+    assert len(subset.dask.layers) == 2
+    assert_equal(subset, array.compute()[expectidx])