Zarr Python

When to use

Use for chunked scientific arrays, hierarchical stores, codecs, sharding, partial reads, cloud object storage, and NumPy/Dask/Xarray interoperability. This community guide targets Zarr-Python 3.4.0, released 2026-09-15, with Python 3.12+. The package version and on-disk format are separate: this release reads/writes formats 2 and 3; new arrays default to format 3. Keep downstream packages that require zarr<3 in their own environments.

Install

uv pip install "zarr==3.4.0" "numpy==2.5.3"
# Optional remote backends and migration CLI:
uv pip install "zarr[remote,cli]==3.4.0" "fsspec==2026.9.0" "s3fs==2026.9.0" "gcsfs==2026.8.1"

Commit the project's resolved lockfile. Optional integration versions exercised here: Dask 2026.8.0, Xarray 2026.9.0, h5py 3.16.0, NumCodecs 0.17.0, obstore 0.11.1. Local examples below and in the references use tiny synthetic arrays; remote snippets are illustrative and require a real authorized store. This does not establish cloud permissions, production throughput, or compatibility of every downstream reader.

Workflow

1. Inspect shape, dtype, axis names, coordinates, units, missing-value convention, format, codec availability, and intended readers. Preserve sample IDs and axis order.
2. Choose chunks for actual selections and a memory budget. For sharding, choose shard dimensions that are multiples of chunk dimensions. Benchmark representative data.
3. Create a new destination (overwrite=False or mode="w-"). Use mode="r" for inspection; "a" can create a missing store and "w" destroys existing content.
4. Write bounded blocks. Assign one writer per stored chunk, or per shard when sharded; serialize metadata, append, and resize operations.
5. Reopen read-only and compare values, dtype, shape, coordinates, units, masks and metadata. An unwritten or missing chunk normally reads as fill_value; a successful open alone does not prove data completeness.
6. For a completed group hierarchy, optionally consolidate metadata. Format-3 consolidation is experimental; refresh it after metadata changes and verify the actual consumers. Publish a completed store only after validation.

Basic array roundtrip

import numpy as np
import zarr
from zarr.codecs import BloscCodec

expected = np.arange(96, dtype="float32").reshape(12, 8)
z = zarr.create_array(
    "array.zarr", shape=expected.shape, dtype=expected.dtype,
    chunks=(4, 4), zarr_format=3,
    compressors=BloscCodec(cname="zstd", clevel=5, shuffle="bitshuffle"),
    dimension_names=("sample", "feature"),
    attributes={"units": "arbitrary", "source": "synthetic example"},
)
for start in range(0, z.shape[0], 4):
    z[start:start + 4] = expected[start:start + 4]

reopened = zarr.open_array("array.zarr", mode="r")
np.testing.assert_array_equal(reopened[:], expected)
assert reopened.dtype == expected.dtype
assert reopened.metadata.dimension_names == ("sample", "feature")
assert reopened.attrs["units"] == "arbitrary"
subset = reopened[2:6, 1:4]  # Only this selection is materialized.

create_array takes either data= or shape= plus dtype=; do not combine data= with explicit shape/dtype. The format-3 numeric default is a bytes serializer followed by ZstdCodec, not Blosc. Set a codec explicitly for reproducibility.

Creation and indexing

import numpy as np
import zarr

z = zarr.create_array(None, data=np.arange(80).reshape(10, 8), chunks=(2, 4))
zeros = zarr.zeros((10, 8), chunks=(2, 4), dtype="f4")
ones = zarr.ones((10, 8), chunks=(2, 4), dtype="f4")
filled = zarr.full((10, 8), fill_value=42, chunks=(2, 4), dtype="i4")
like = zarr.zeros_like(z)
np.testing.assert_array_equal(filled[:], np.full((10, 8), 42))

# Coordinate indexing pairs corresponding coordinates; orthogonal indexing is a product.
np.testing.assert_array_equal(z.vindex[[0, 5], [2, 7]], [2, 47])
np.testing.assert_array_equal(z.get_coordinate_selection(([0, 5], [2, 7])), [2, 47])
assert z.oindex[[0, 5], [2, 7]].shape == (2, 2)
assert z.blocks[0, 0].shape == (2, 4)
z[0, :] = np.arange(8)

Negative-step slices are unsupported. Array reads return NumPy data in the default CPU configuration. np.asarray(z), np.sum(z), z[:], or a Dask .compute() of a full array can materialize the entire logical dataset; use bounded selections or lazy reductions.

Resize and append

import numpy as np
import zarr


…