Skip to content

Indexing sharded array with oindex is slow #3829

@slevang

Description

@slevang

Indexing a sharded array with oindex[[0], ...] is much slower than the equivalent slice 0:1.

import time
from tempfile import TemporaryDirectory

import numpy as np
import zarr

with TemporaryDirectory() as tmp:
    shape = (2, 1000, 100, 100)
    chunks = (1, 1000, 10, 10)
    shards = (1, 1000, 100, 100)

    z = zarr.create_array(
        tmp,
        shape=shape,
        chunks=chunks,
        shards=shards,
        dtype="f4",
    )
    z[:] = np.random.randn(*shape).astype(np.float32)

    t0 = time.perf_counter()
    a = z[0:1, :, :, :]
    t_slice = time.perf_counter() - t0

    t0 = time.perf_counter()
    b = z.oindex[[0], :, :, :]
    t_oindex = time.perf_counter() - t0

    print(f"slice   : {t_slice:.3f} s")
    print(f"oindex  : {t_oindex:.3f} s")

Gives:

slice   : 0.098 s
oindex  : 1.178 s

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions