|
| 1 | +import numpy as np |
| 2 | + |
| 3 | +from xarray.core.datatree import Variable |
| 4 | + |
| 5 | + |
| 6 | +def align_nd_chunks( |
| 7 | + nd_var_chunks: tuple[tuple[int, ...], ...], |
| 8 | + nd_backend_chunks: tuple[tuple[int, ...], ...], |
| 9 | +) -> tuple[tuple[int, ...], ...]: |
| 10 | + if len(nd_backend_chunks) != len(nd_var_chunks): |
| 11 | + raise ValueError( |
| 12 | + "The number of dimensions on the backend and the variable must be the same." |
| 13 | + ) |
| 14 | + |
| 15 | + nd_aligned_chunks: list[tuple[int, ...]] = [] |
| 16 | + for backend_chunks, var_chunks in zip( |
| 17 | + nd_backend_chunks, nd_var_chunks, strict=True |
| 18 | + ): |
| 19 | + # Validate that they have the same number of elements |
| 20 | + if sum(backend_chunks) != sum(var_chunks): |
| 21 | + raise ValueError( |
| 22 | + "The number of elements in the backend does not " |
| 23 | + "match the number of elements in the variable. " |
| 24 | + "This inconsistency should never occur at this stage." |
| 25 | + ) |
| 26 | + |
| 27 | + # Validate if the backend_chunks satisfy the condition that all the values |
| 28 | + # excluding the borders are equal |
| 29 | + if len(set(backend_chunks[1:-1])) > 1: |
| 30 | + raise ValueError( |
| 31 | + f"This function currently supports aligning chunks " |
| 32 | + f"only when backend chunks are of uniform size, excluding borders. " |
| 33 | + f"If you encounter this error, please report it—this scenario should never occur " |
| 34 | + f"unless there is an internal misuse. " |
| 35 | + f"Backend chunks: {backend_chunks}" |
| 36 | + ) |
| 37 | + |
| 38 | + # The algorithm assumes that there are always two borders on the |
| 39 | + # Backend and the Array if not, the result is going to be the same |
| 40 | + # as the input, and there is nothing to optimize |
| 41 | + if len(backend_chunks) == 1: |
| 42 | + nd_aligned_chunks.append(backend_chunks) |
| 43 | + continue |
| 44 | + |
| 45 | + if len(var_chunks) == 1: |
| 46 | + nd_aligned_chunks.append(var_chunks) |
| 47 | + continue |
| 48 | + |
| 49 | + # Size of the chunk on the backend |
| 50 | + fixed_chunk = max(backend_chunks) |
| 51 | + |
| 52 | + # The ideal size of the chunks is the maximum of the two; this would avoid |
| 53 | + # that we use more memory than expected |
| 54 | + max_chunk = max(fixed_chunk, max(var_chunks)) |
| 55 | + |
| 56 | + # The algorithm assumes that the chunks on this array are aligned except the last one |
| 57 | + # because it can be considered a partial one |
| 58 | + aligned_chunks: list[int] = [] |
| 59 | + |
| 60 | + # For simplicity of the algorithm, let's transform the Array chunks in such a way that |
| 61 | + # we remove the partial chunks. To achieve this, we add artificial data to the borders |
| 62 | + t_var_chunks = list(var_chunks) |
| 63 | + t_var_chunks[0] += fixed_chunk - backend_chunks[0] |
| 64 | + t_var_chunks[-1] += fixed_chunk - backend_chunks[-1] |
| 65 | + |
| 66 | + # The unfilled_size is the amount of space that has not been filled on the last |
| 67 | + # processed chunk; this is equivalent to the amount of data that would need to be |
| 68 | + # added to a partial Zarr chunk to fill it up to the fixed_chunk size |
| 69 | + unfilled_size = 0 |
| 70 | + |
| 71 | + for var_chunk in t_var_chunks: |
| 72 | + # Ideally, we should try to preserve the original Dask chunks, but this is only |
| 73 | + # possible if the last processed chunk was aligned (unfilled_size == 0) |
| 74 | + ideal_chunk = var_chunk |
| 75 | + if unfilled_size: |
| 76 | + # If that scenario is not possible, the best option is to merge the chunks |
| 77 | + ideal_chunk = var_chunk + aligned_chunks[-1] |
| 78 | + |
| 79 | + while ideal_chunk: |
| 80 | + if not unfilled_size: |
| 81 | + # If the previous chunk is filled, let's add a new chunk |
| 82 | + # of size 0 that will be used on the merging step to simplify the algorithm |
| 83 | + aligned_chunks.append(0) |
| 84 | + |
| 85 | + if ideal_chunk > max_chunk: |
| 86 | + # If the ideal_chunk is bigger than the max_chunk, |
| 87 | + # we need to increase the last chunk as much as possible |
| 88 | + # but keeping it aligned, and then add a new chunk |
| 89 | + max_increase = max_chunk - aligned_chunks[-1] |
| 90 | + max_increase = ( |
| 91 | + max_increase - (max_increase - unfilled_size) % fixed_chunk |
| 92 | + ) |
| 93 | + aligned_chunks[-1] += max_increase |
| 94 | + else: |
| 95 | + # Perfect scenario where the chunks can be merged without any split. |
| 96 | + aligned_chunks[-1] = ideal_chunk |
| 97 | + |
| 98 | + ideal_chunk -= aligned_chunks[-1] |
| 99 | + unfilled_size = ( |
| 100 | + fixed_chunk - aligned_chunks[-1] % fixed_chunk |
| 101 | + ) % fixed_chunk |
| 102 | + |
| 103 | + # Now we have to remove the artificial data added to the borders |
| 104 | + for order in [-1, 1]: |
| 105 | + border_size = fixed_chunk - backend_chunks[::order][0] |
| 106 | + aligned_chunks = aligned_chunks[::order] |
| 107 | + aligned_chunks[0] -= border_size |
| 108 | + t_var_chunks = t_var_chunks[::order] |
| 109 | + t_var_chunks[0] -= border_size |
| 110 | + if ( |
| 111 | + len(aligned_chunks) >= 2 |
| 112 | + and aligned_chunks[0] + aligned_chunks[1] <= max_chunk |
| 113 | + and aligned_chunks[0] != t_var_chunks[0] |
| 114 | + ): |
| 115 | + # The artificial data added to the border can introduce inefficient chunks |
| 116 | + # on the borders, for that reason, we will check if we can merge them or not |
| 117 | + # Example: |
| 118 | + # backend_chunks = [6, 6, 1] |
| 119 | + # var_chunks = [6, 7] |
| 120 | + # t_var_chunks = [6, 12] |
| 121 | + # The ideal output should preserve the same var_chunks, but the previous loop |
| 122 | + # is going to produce aligned_chunks = [6, 6, 6] |
| 123 | + # And after removing the artificial data, we will end up with aligned_chunks = [6, 6, 1] |
| 124 | + # which is not ideal and can be merged into a single chunk |
| 125 | + aligned_chunks[1] += aligned_chunks[0] |
| 126 | + aligned_chunks = aligned_chunks[1:] |
| 127 | + |
| 128 | + t_var_chunks = t_var_chunks[::order] |
| 129 | + aligned_chunks = aligned_chunks[::order] |
| 130 | + |
| 131 | + nd_aligned_chunks.append(tuple(aligned_chunks)) |
| 132 | + |
| 133 | + return tuple(nd_aligned_chunks) |
| 134 | + |
| 135 | + |
| 136 | +def build_grid_chunks( |
| 137 | + size: int, |
| 138 | + chunk_size: int, |
| 139 | + region: slice | None = None, |
| 140 | +) -> tuple[int, ...]: |
| 141 | + if region is None: |
| 142 | + region = slice(0, size) |
| 143 | + |
| 144 | + region_start = region.start if region.start else 0 |
| 145 | + # Generate the zarr chunks inside the region of this dim |
| 146 | + chunks_on_region = [chunk_size - (region_start % chunk_size)] |
| 147 | + chunks_on_region.extend([chunk_size] * ((size - chunks_on_region[0]) // chunk_size)) |
| 148 | + if (size - chunks_on_region[0]) % chunk_size != 0: |
| 149 | + chunks_on_region.append((size - chunks_on_region[0]) % chunk_size) |
| 150 | + return tuple(chunks_on_region) |
| 151 | + |
| 152 | + |
| 153 | +def grid_rechunk( |
| 154 | + v: Variable, |
| 155 | + enc_chunks: tuple[int, ...], |
| 156 | + region: tuple[slice, ...], |
| 157 | +) -> Variable: |
| 158 | + nd_var_chunks = v.chunks |
| 159 | + if not nd_var_chunks: |
| 160 | + return v |
| 161 | + |
| 162 | + nd_grid_chunks = tuple( |
| 163 | + build_grid_chunks( |
| 164 | + sum(var_chunks), |
| 165 | + region=interval, |
| 166 | + chunk_size=chunk_size, |
| 167 | + ) |
| 168 | + for var_chunks, chunk_size, interval in zip( |
| 169 | + nd_var_chunks, enc_chunks, region, strict=True |
| 170 | + ) |
| 171 | + ) |
| 172 | + |
| 173 | + nd_aligned_chunks = align_nd_chunks( |
| 174 | + nd_var_chunks=nd_var_chunks, |
| 175 | + nd_backend_chunks=nd_grid_chunks, |
| 176 | + ) |
| 177 | + v = v.chunk(dict(zip(v.dims, nd_aligned_chunks, strict=True))) |
| 178 | + return v |
| 179 | + |
| 180 | + |
| 181 | +def validate_grid_chunks_alignment( |
| 182 | + nd_var_chunks: tuple[tuple[int, ...], ...] | None, |
| 183 | + enc_chunks: tuple[int, ...], |
| 184 | + backend_shape: tuple[int, ...], |
| 185 | + region: tuple[slice, ...], |
| 186 | + allow_partial_chunks: bool, |
| 187 | + name: str, |
| 188 | +): |
| 189 | + if nd_var_chunks is None: |
| 190 | + return |
| 191 | + base_error = ( |
| 192 | + "Specified Zarr chunks encoding['chunks']={enc_chunks!r} for " |
| 193 | + "variable named {name!r} would overlap multiple Dask chunks. " |
| 194 | + "Check the chunk at position {var_chunk_pos}, which has a size of " |
| 195 | + "{var_chunk_size} on dimension {dim_i}. It is unaligned with " |
| 196 | + "backend chunks of size {chunk_size} in region {region}. " |
| 197 | + "Writing this array in parallel with Dask could lead to corrupted data. " |
| 198 | + "To resolve this issue, consider one of the following options: " |
| 199 | + "- Rechunk the array using `chunk()`. " |
| 200 | + "- Modify or delete `encoding['chunks']`. " |
| 201 | + "- Set `safe_chunks=False`. " |
| 202 | + "- Enable automatic chunks alignment with `align_chunks=True`." |
| 203 | + ) |
| 204 | + |
| 205 | + for dim_i, chunk_size, var_chunks, interval, size in zip( |
| 206 | + range(len(enc_chunks)), |
| 207 | + enc_chunks, |
| 208 | + nd_var_chunks, |
| 209 | + region, |
| 210 | + backend_shape, |
| 211 | + strict=True, |
| 212 | + ): |
| 213 | + for i, chunk in enumerate(var_chunks[1:-1]): |
| 214 | + if chunk % chunk_size: |
| 215 | + raise ValueError( |
| 216 | + base_error.format( |
| 217 | + var_chunk_pos=i + 1, |
| 218 | + var_chunk_size=chunk, |
| 219 | + name=name, |
| 220 | + dim_i=dim_i, |
| 221 | + chunk_size=chunk_size, |
| 222 | + region=interval, |
| 223 | + enc_chunks=enc_chunks, |
| 224 | + ) |
| 225 | + ) |
| 226 | + |
| 227 | + interval_start = interval.start if interval.start else 0 |
| 228 | + |
| 229 | + if len(var_chunks) > 1: |
| 230 | + # The first border size is the amount of data that needs to be updated on the |
| 231 | + # first chunk taking into account the region slice. |
| 232 | + first_border_size = chunk_size |
| 233 | + if allow_partial_chunks: |
| 234 | + first_border_size = chunk_size - interval_start % chunk_size |
| 235 | + |
| 236 | + if (var_chunks[0] - first_border_size) % chunk_size: |
| 237 | + raise ValueError( |
| 238 | + base_error.format( |
| 239 | + var_chunk_pos=0, |
| 240 | + var_chunk_size=var_chunks[0], |
| 241 | + name=name, |
| 242 | + dim_i=dim_i, |
| 243 | + chunk_size=chunk_size, |
| 244 | + region=interval, |
| 245 | + enc_chunks=enc_chunks, |
| 246 | + ) |
| 247 | + ) |
| 248 | + |
| 249 | + if not allow_partial_chunks: |
| 250 | + region_stop = interval.stop if interval.stop else size |
| 251 | + |
| 252 | + error_on_last_chunk = base_error.format( |
| 253 | + var_chunk_pos=len(var_chunks) - 1, |
| 254 | + var_chunk_size=var_chunks[-1], |
| 255 | + name=name, |
| 256 | + dim_i=dim_i, |
| 257 | + chunk_size=chunk_size, |
| 258 | + region=interval, |
| 259 | + enc_chunks=enc_chunks, |
| 260 | + ) |
| 261 | + if interval_start % chunk_size: |
| 262 | + # The last chunk which can also be the only one is a partial chunk |
| 263 | + # if it is not aligned at the beginning |
| 264 | + raise ValueError(error_on_last_chunk) |
| 265 | + |
| 266 | + if np.ceil(region_stop / chunk_size) == np.ceil(size / chunk_size): |
| 267 | + # If the region is covering the last chunk then check |
| 268 | + # if the reminder with the default chunk size |
| 269 | + # is equal to the size of the last chunk |
| 270 | + if var_chunks[-1] % chunk_size != size % chunk_size: |
| 271 | + raise ValueError(error_on_last_chunk) |
| 272 | + elif var_chunks[-1] % chunk_size: |
| 273 | + raise ValueError(error_on_last_chunk) |
![pre-commit-ci[bot]](https://avatars.githubusercontent.com/in/68672?v=4&size=40){kind=avatar}
0 commit comments