Turn any AI agent into an AI Scientist. The #1 Agent Skills library for science, used by 250,000+ scientists worldwide. 177 ready-to-use validated skills plus 100+ scientific databases covering biology, chemistry, medicine, and drug discovery. Compatible with Cursor, Claude Code, Codex, Pi, Antigravity, and the open Agent Skills standard.
Zarr-Python 3.4.0 API reference
These are selected callable forms, not exhaustive copied signatures. Explicitly select format/mode when behavior matters; defaults can also be affected by runtime configuration.
| Operation | Current form / behavior |
|---|---|
| New array | zarr.create_array(store, shape=..., dtype=..., chunks="auto", shards=None, compressors="auto", overwrite=False); store=None gives memory. |
| From NumPy | zarr.create_array(store, data=values, chunks=...); omit both shape and dtype. |
| Copy array | zarr.from_array(store, data=source, chunks="keep", shards="keep"); preserve tested source semantics, validate destination after copying. |
| Convenience | zarr.zeros(shape, **kwargs), ones, empty, full(shape, fill_value, **kwargs), array(data, **kwargs), zeros_like(source, **kwargs). empty does not establish scientifically meaningful values. |
| Open array | zarr.open_array(store, mode="r", path="", zarr_format=None, storage_options=None); mode is forwarded through kwargs. |
| Open unknown node | zarr.open(store, mode="r") detects array/group; default signature mode is None, resolved from store writability (normally "a"). |
| Open group | zarr.open_group(store, mode="r", path=None, use_consolidated=None); None uses consolidation if available, False bypasses, True requires it. |
| Create/require child | group.create_array(name, ...), require_array(name, shape=..., dtype=...), create_group(name), require_group(name). |
| Resize/append | array.resize(new_shape_tuple) returns None; array.append(data, axis=0) returns the new shape. |
| Metadata | array.attrs, group.attrs hold JSON-compatible values; native v3 dimension_names are not coordinate arrays. |
| Sizing | array.nbytes is logical bytes; array.nbytes_stored() computes stored bytes. array.info is metadata; array.info_complete() obtains storage-dependent details. |
| Consolidate | zarr.consolidate_metadata(store, path=None, zarr_format=None) returns Group; supply a configured store for credentials. |
| Require consolidation | zarr.open_consolidated(store, mode="r", ...) returns Group, not Array. |
Modes: r existing read-only; r+ existing read/write; a open or create; w replace;
w- create only. Creation flags are not distributed exclusive-writer locks.
On an existing array, open_array(..., config=...) does not forward that configuration
in 3.4.0. Use a zarr.config.set context before opening when enforcing missing-chunk
policy; see the tested example in performance.
Selection semantics
For a 2-D array, z.vindex[[0, 5], [2, 7]] pairs coordinates and returns two values;
z.oindex[[0, 5], [2, 7]] returns a 2-by-2 Cartesian product. Explicit coordinate
selection takes a tuple of per-axis coordinate arrays:
z.get_coordinate_selection(([0, 5], [2, 7])). z.blocks[i, j] selects chunks.
Negative-step slices are unsupported. Rectilinear grids require an experimental flag;
Array.chunks cannot describe every rectilinear layout with one tuple.
Stores and codecs
LocalStore, MemoryStore, ZipStore, FsspecStore, and ObjectStore live in
zarr.storage. ObjectStore requires obstore; FsspecStore requires fsspec plus the
selected protocol backend. FsspecStore.from_url(url, storage_options=None, read_only=False) resolves the URI. Old DirectoryStore/FSStore names are replaced;
legacy database/N5 stores are removed. S3Map/GCSMap are backend mappings, not newly
supported Zarr store classes.
For format 3, use zarr.codecs.BloscCodec, GzipCodec, ZstdCodec;
pass shuffle="bitshuffle" to BloscCodec (the BloscShuffle enum is deprecated).
compressors=None disables compression, not serialization. Format-2 arrays use
NumCodecs (e.g. numcodecs.Blosc) through compressors= in create_array. Numeric
format-3 defaults are Zstd; use "auto", not the invalid "default" string.
Exceptions and coordination
Missing nodes use NodeNotFoundError / ArrayNotFoundError / GroupNotFoundError;
PathNotFoundError and ReadOnlyError are not exported by this release. Read-only
mutation raises ValueError. UnknownCodecError subclasses ValueError; malformed
metadata may raise MetadataValidationError. Preserve the original error context;
authentication errors must not be treated as missing chunks.
Use one writer per stored object (per shard if sharded), including boundaries.
No working v2 synchronizers are provided. Neither consolidation nor require_array
is a transaction or a lock. Current concurrency configuration keys include
async.concurrency and threading.max_workers.
Sources: API, released sync API, released errors.