zarr-python

name	zarr-python
description	Chunked N-D arrays for cloud storage (Zarr-Python 3). Compressed arrays, parallel I/O, S3/GCS via fsspec, NumPy/Dask/Xarray compatible, for large-scale scientific computing pipelines.
allowed-tools	Read Write Edit Bash
license	MIT license
compatibility	Requires Python 3.12+ and zarr 3.x. Cloud I/O needs zarr[remote] plus s3fs or gcsfs. Legacy Zarr v2 workflows use zarr==2.* on older Python.
metadata	{"version":"1.0","skill-author":"K-Dense Inc."}

Zarr Python

Overview

Zarr is a Python library for storing large N-dimensional arrays with chunking and compression. Apply this skill for efficient parallel I/O, cloud-native workflows, and seamless integration with NumPy, Dask, and Xarray.

Current upstream: zarr 3.2.1 (PyPI, May 2026). Docs: zarr.readthedocs.io. New arrays default to Zarr format 3; set zarr_format=2 for legacy interop. This skill is a community guide maintained by K-Dense Inc., not an official zarr-developers package.

Quick Start

Installation

uv pip install "zarr>=3.2,<4"

Requires Python 3.12+ (per PyPI metadata for zarr 3.2.x). For remote stores (S3, GCS, HTTP):

uv pip install "zarr[remote]"
uv pip install s3fs   # AWS S3
uv pip install gcsfs  # Google Cloud Storage

Pin zarr>=3,<4 in application dependencies. Use uv pip install "zarr==2.*" only when you must stay on Zarr-Python 2 / Python 3.10–3.11.

Basic Array Creation

import zarr
import numpy as np

# Create a 2D array with chunking and compression
z = zarr.create_array(
    store="data/my_array.zarr",
    shape=(10000, 10000),
    chunks=(1000, 1000),
    dtype="f4"
)

# Write data using NumPy-style indexing
z[:, :] = np.random.random((10000, 10000))

# Read data
data = z[0:100, 0:100]  # Returns NumPy array

Core Operations

Creating Arrays

Zarr provides multiple convenience functions for array creation:

# Create empty array
z = zarr.zeros(shape=(10000, 10000), chunks=(1000, 1000), dtype='f4',
               store='data.zarr')

# Create filled arrays
z = zarr.ones((5000, 5000), chunks=(500, 500))
z = zarr.full((1000, 1000), fill_value=42, chunks=(100, 100))

# Create from existing data
data = np.arange(10000).reshape(100, 100)
z = zarr.array(data, chunks=(10, 10), store='data.zarr')

# Create like another array
z2 = zarr.zeros_like(z)  # Matches shape, chunks, dtype of z

Opening Existing Arrays

# Open array (read/write mode by default)
z = zarr.open_array('data.zarr', mode='r+')

# Read-only mode
z = zarr.open_array('data.zarr', mode='r')

# The open() function auto-detects arrays vs groups
z = zarr.open('data.zarr')  # Returns Array or Group

Reading and Writing Data

Zarr arrays support NumPy-like indexing:

# Write entire array
z[:] = 42

# Write slices
z[0, :] = np.arange(100)
z[10:20, 50:60] = np.random.random((10, 10))

# Read data (returns NumPy array)
data = z[0:100, 0:100]
row = z[5, :]

# Advanced indexing
z.vindex[[0, 5, 10], [2, 8, 15]]  # Coordinate indexing
z.oindex[0:10, [5, 10, 15]]       # Orthogonal indexing
z.blocks[0, 0]                     # Block/chunk indexing

Resizing and Appending

# Resize array (v3: pass shape as a tuple)
z.resize((15000, 15000))

# Append data along an axis
z.append(np.random.random((1000, 10000)), axis=0)  # Adds rows

Chunking Strategies

Chunking is critical for performance. Choose chunk sizes and shapes based on access patterns.

Chunk Size Guidelines

• Minimum chunk size: 1 MB recommended for optimal performance
• Balance: Larger chunks = fewer metadata operations; smaller chunks = better parallel access
• Memory consideration: Entire chunks must fit in memory during compression

# Configure chunk size (aim for ~1MB per chunk)
# For float32 data: 1MB = 262,144 elements = 512×512 array
z = zarr.zeros(
    shape=(10000, 10000),
    chunks=(512, 512),  # ~1MB chunks
    dtype='f4'
)

Aligning Chunks with Access Patterns

Critical: Chunk shape dramatically affects performance based on how data is accessed.

# If accessing rows frequently (first dimension)
z = zarr.zeros((10000, 10000), chunks=(10, 10000))  # Chunk spans columns

# If accessing columns frequently (second dimension)
z = zarr.zeros((10000, 10000), chunks=(10000, 10))  # Chunk spans rows

# For mixed access patterns (balanced approach)
z = zarr.zeros((10000, 10000), chunks=(1000, 1000))  # Square chunks

Performance example: For a (200, 200, 200) array, reading along the first dimension:

• Using chunks (1, 200, 200): ~107ms
• Using chunks (200, 200, 1): ~1.65ms (65× faster!)

Sharding for Large-Scale Storage

When arrays have millions of small chunks, use sharding to group chunks into larger storage objects:

from zarr.codecs import BloscCodec, BytesCodec, ShardingCodec

# Create array with sharding
z = zarr.create_array(
    store='data.zarr',
    shape=(100000, 100000),
    chunks=(100, 100),  # Small chunks for access
    shards=(1000, 1000),  # Groups 100 chunks per shard
    dtype='f4'
)

Benefits:

• Reduces file system overhead from millions of small files
• Improves cloud storage performance (fewer object requests)
• Prevents filesystem block size waste

Important: Entire shards must fit in memory before writing.

Compression

Zarr applies compression per chunk to reduce storage while maintaining fast access.

Configuring Compression

from zarr.codecs import BloscCodec, GzipCodec, ZstdCodec, BytesCodec

# Default: Blosc with Zstandard
z = zarr.zeros((1000, 1000), chunks=(100, 100))  # Uses default compression

# Configure Blosc codec
z = zarr.create_array(
    store='data.zarr',
    shape=(1000, 1000),
    chunks=(100, 100),
    dtype='f4',
    codecs=[BloscCodec(cname='zstd', clevel=5, shuffle='shuffle')]
)

# Available Blosc compressors: 'blosclz', 'lz4', 'lz4hc', 'snappy', 'zlib', 'zstd'

# Use Gzip compression
z = zarr.create_array(
    store='data.zarr',
    shape=(1000, 1000),
    chunks=(100, 100),
    dtype='f4',
    codecs=[GzipCodec(level=6)]
)

# Disable compression
z = zarr.create_array(
    store='data.zarr',
    shape=(1000, 1000),
    chunks=(100, 100),
    dtype='f4',
    codecs=[BytesCodec()]  # No compression
)

Compression Performance Tips

• Blosc (default): Fast compression/decompression, good for interactive workloads
• Zstandard: Better compression ratios, slightly slower than LZ4
• Gzip: Maximum compression, slower performance
• LZ4: Fastest compression, lower ratios
• Shuffle: Enable shuffle filter for better compression on numeric data

# Optimal for numeric scientific data
codecs=[BloscCodec(cname='zstd', clevel=5, shuffle='shuffle')]

# Optimal for speed
codecs=[BloscCodec(cname='lz4', clevel=1)]

# Optimal for compression ratio
codecs=[GzipCodec(level=9)]

Storage Backends

Zarr supports multiple storage backends through a flexible storage interface.

Local Filesystem (Default)

from zarr.storage import LocalStore

# Explicit store creation
store = LocalStore('data/my_array.zarr')
z = zarr.open_array(store=store, mode='w', shape=(1000, 1000), chunks=(100, 100))

# Or use string path (creates LocalStore automatically)
z = zarr.open_array('data/my_array.zarr', mode='w', shape=(1000, 1000),
                    chunks=(100, 100))

In-Memory Storage

from zarr.storage import MemoryStore

# Create in-memory store
store = MemoryStore()
z = zarr.open_array(store=store, mode='w', shape=(1000, 1000), chunks=(100, 100))

# Data exists only in memory, not persisted

ZIP File Storage

from zarr.storage import ZipStore

# Write to ZIP file
store = ZipStore('data.zip', mode='w')
z = zarr.open_array(store=store, mode='w', shape=(1000, 1000), chunks=(100, 100))
z[:] = np.random.random((1000, 1000))
store.close()  # IMPORTANT: Must close ZipStore

# Read from ZIP file
store = ZipStore('data.zip', mode='r')
z = zarr.open_array(store=store)
data = z[:]
store.close()

Cloud Storage (S3, GCS)

Zarr 3 uses fsspec backends via URI strings or FsspecStore (preferred over legacy S3Map/GCSMap).

import zarr

# S3 — credentials from standard AWS env vars (scope reads to these keys only)
# AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_DEFAULT_REGION
z = zarr.create_array(
    store="s3://my-bucket/path/to/array.zarr",
    shape=(1000, 1000),
    chunks=(100, 100),
    dtype="f4",
    storage_options={"anon": False},
)
z[:] = data

# GCS — GOOGLE_APPLICATION_CREDENTIALS or gcloud default credentials
z = zarr.open_array(
    "gs://my-bucket/path/to/array.zarr",
    mode="r",
    storage_options={"project": "my-project"},
)

# Explicit store (any fsspec filesystem)
from zarr.storage import FsspecStore
store = FsspecStore.from_url("s3://my-bucket/data.zarr", storage_options={"anon": False})
root = zarr.open_group(store=store, mode="r+")

Cloud backends read credentials from provider environment variables locally via fsspec; they are not sent to third-party endpoints outside your configured bucket/project.

Cloud Storage Best Practices:

• Use consolidated metadata to reduce latency: zarr.consolidate_metadata(store)
• Align chunk sizes with cloud object sizing (typically 5-100 MB optimal)
• Enable parallel writes using Dask for large-scale data
• Consider sharding to reduce number of objects

Groups and Hierarchies

Groups organize multiple arrays hierarchically, similar to directories or HDF5 groups.

Creating and Using Groups

# Create root group
root = zarr.group(store='data/hierarchy.zarr')

# Create sub-groups
temperature = root.create_group('temperature')
precipitation = root.create_group('precipitation')

# Create arrays within groups
temp_array = temperature.create_array(
    name='t2m',
    shape=(365, 720, 1440),
    chunks=(1, 720, 1440),
    dtype='f4'
)

precip_array = precipitation.create_array(
    name='prcp',
    shape=(365, 720, 1440),
    chunks=(1, 720, 1440),
    dtype='f4'
)

# Access using paths
array = root['temperature/t2m']

# Visualize hierarchy
print(root.tree())
# Output:
# /
#  ├── temperature
#  │   └── t2m (365, 720, 1440) f4
#  └── precipitation
#      └── prcp (365, 720, 1440) f4

Group API (v3)

Use create_array / require_array (h5py-style create_dataset / require_dataset were removed in v3):

root = zarr.group('data.zarr')
arr = root.create_array('my_data', shape=(1000, 1000), chunks=(100, 100), dtype='f4')

grp = root.require_group('subgroup')
arr2 = grp.require_array('array', shape=(500, 500), chunks=(50, 50), dtype='i4')

Attributes and Metadata

Attach custom metadata to arrays and groups using attributes:

# Add attributes to array
z = zarr.zeros((1000, 1000), chunks=(100, 100))
z.attrs['description'] = 'Temperature data in Kelvin'
z.attrs['units'] = 'K'
z.attrs['created'] = '2024-01-15'
z.attrs['processing_version'] = 2.1

# Attributes are stored as JSON
print(z.attrs['units'])  # Output: K

# Add attributes to groups
root = zarr.group('data.zarr')
root.attrs['project'] = 'Climate Analysis'
root.attrs['institution'] = 'Research Institute'

# Attributes persist with the array/group
z2 = zarr.open('data.zarr')
print(z2.attrs['description'])

Important: Attributes must be JSON-serializable (strings, numbers, lists, dicts, booleans, null).

Integration with NumPy, Dask, and Xarray

NumPy Integration

Zarr arrays implement the NumPy array interface:

import numpy as np
import zarr

z = zarr.zeros((1000, 1000), chunks=(100, 100))

# Use NumPy functions directly
result = np.sum(z, axis=0)  # NumPy operates on Zarr array
mean = np.mean(z[:100, :100])

# Convert to NumPy array
numpy_array = z[:]  # Loads entire array into memory

Dask Integration

Dask provides lazy, parallel computation on Zarr arrays:

import dask.array as da
import zarr

# Create large Zarr array
z = zarr.open('data.zarr', mode='w', shape=(100000, 100000),
              chunks=(1000, 1000), dtype='f4')

# Load as Dask array (lazy, no data loaded)
dask_array = da.from_zarr('data.zarr')

# Perform computations (parallel, out-of-core)
result = dask_array.mean(axis=0).compute()  # Parallel computation

# Write Dask array to Zarr
large_array = da.random.random((100000, 100000), chunks=(1000, 1000))
da.to_zarr(large_array, 'output.zarr')

Benefits:

• Process datasets larger than memory
• Automatic parallel computation across chunks
• Efficient I/O with chunked storage

Xarray Integration

Xarray provides labeled, multidimensional arrays with Zarr backend:

import xarray as xr
import zarr

# Open Zarr store as Xarray Dataset (lazy loading)
ds = xr.open_zarr('data.zarr')

# Dataset includes coordinates and metadata
print(ds)

# Access variables
temperature = ds['temperature']

# Perform labeled operations
subset = ds.sel(time='2024-01', lat=slice(30, 60))

# Write Xarray Dataset to Zarr
ds.to_zarr('output.zarr')

# Create from scratch with coordinates
ds = xr.Dataset(
    {
        'temperature': (['time', 'lat', 'lon'], data),
        'precipitation': (['time', 'lat', 'lon'], data2)
    },
    coords={
        'time': pd.date_range('2024-01-01', periods=365),
        'lat': np.arange(-90, 91, 1),
        'lon': np.arange(-180, 180, 1)
    }
)
ds.to_zarr('climate_data.zarr')

Benefits:

• Named dimensions and coordinates
• Label-based indexing and selection
• Integration with pandas for time series
• NetCDF-like interface familiar to climate/geospatial scientists

Parallel Computing and Thread Safety

The synchronizer argument (ThreadSynchronizer, ProcessSynchronizer) is not ported to Zarr-Python 3 yet. Use these patterns instead:

• Reads: always safe across threads/processes.
• Writes: safe when each worker writes to non-overlapping chunks; most stores support atomic chunk writes.
• Overlapping writes: coordinate externally (file locks, workflow design) until synchronizers return.

For Dask-heavy workloads, tune Zarr async concurrency — see Optimizing performance.

Consolidated Metadata

For hierarchical stores with many arrays, consolidate metadata into a single file to reduce I/O operations:

import zarr

# After creating arrays/groups
root = zarr.group('data.zarr')
# ... create multiple arrays/groups ...

# Consolidate metadata
zarr.consolidate_metadata('data.zarr')

# Open with consolidated metadata (faster, especially on cloud storage)
root = zarr.open_consolidated('data.zarr')

Benefits:

• Reduces metadata read operations from N (one per array) to 1
• Critical for cloud storage (reduces latency)
• Speeds up tree() operations and group traversal

Cautions:

• Metadata can become stale if arrays update without re-consolidation
• Not suitable for frequently-updated datasets
• Multi-writer scenarios may have inconsistent reads

Performance Optimization

Checklist for Optimal Performance

1. Chunk Size: Aim for 1-10 MB per chunk

   # For float32: 1MB = 262,144 elements
   chunks = (512, 512)  # 512×512×4 bytes = ~1MB
   

2. Chunk Shape: Align with access patterns

   # Row-wise access → chunk spans columns: (small, large)
   # Column-wise access → chunk spans rows: (large, small)
   # Random access → balanced: (medium, medium)
   

3. Compression: Choose based on workload

   # Interactive/fast: BloscCodec(cname='lz4')
   # Balanced: BloscCodec(cname='zstd', clevel=5)
   # Maximum compression: GzipCodec(level=9)
   

4. Storage Backend: Match to environment

   # Local: LocalStore (default)
   # Cloud: fsspec URIs or FsspecStore + consolidated metadata
   # Temporary: MemoryStore
   

5. Sharding: Use for large-scale datasets

   # When you have millions of small chunks
   shards=(10*chunk_size, 10*chunk_size)
   

6. Parallel I/O: Use Dask for large operations

   import dask.array as da
   dask_array = da.from_zarr('data.zarr')
   result = dask_array.compute(scheduler='threads', num_workers=8)
   

Profiling and Debugging

# Print detailed array information
print(z.info)

# Output includes:
# - Type, shape, chunks, dtype
# - Compression codec and level
# - Storage size (compressed vs uncompressed)
# - Storage location

# Check storage size
print(f"Compressed size: {z.nbytes_stored / 1e6:.2f} MB")
print(f"Uncompressed size: {z.nbytes / 1e6:.2f} MB")
print(f"Compression ratio: {z.nbytes / z.nbytes_stored:.2f}x")

Common Patterns and Best Practices

Pattern: Time Series Data

# Store time series with time as first dimension
# This allows efficient appending of new time steps
z = zarr.open('timeseries.zarr', mode='a',
              shape=(0, 720, 1440),  # Start with 0 time steps
              chunks=(1, 720, 1440),  # One time step per chunk
              dtype='f4')

# Append new time steps
new_data = np.random.random((1, 720, 1440))
z.append(new_data, axis=0)

Pattern: Large Matrix Operations

import dask.array as da

# Create large matrix in Zarr
z = zarr.open('matrix.zarr', mode='w',
              shape=(100000, 100000),
              chunks=(1000, 1000),
              dtype='f8')

# Use Dask for parallel computation
dask_z = da.from_zarr('matrix.zarr')
result = (dask_z @ dask_z.T).compute()  # Parallel matrix multiply

Pattern: Cloud-Native Workflow

import zarr

path = "s3://my-bucket/data.zarr"
z = zarr.create_array(
    store=path,
    shape=(10000, 10000),
    chunks=(500, 500),
    dtype="f4",
    storage_options={"anon": False},
)
z[:] = data

zarr.consolidate_metadata(path)
z_read = zarr.open_consolidated(path, storage_options={"anon": False})
subset = z_read[0:100, 0:100]

Pattern: Format Conversion

# HDF5 to Zarr
import h5py
import zarr

with h5py.File('data.h5', 'r') as h5:
    dataset = h5['dataset_name']
    z = zarr.array(dataset[:],
                   chunks=(1000, 1000),
                   store='data.zarr')

# NumPy to Zarr
import numpy as np
data = np.load('data.npy')
z = zarr.array(data, chunks='auto', store='data.zarr')

# Zarr to NetCDF (via Xarray)
import xarray as xr
ds = xr.open_zarr('data.zarr')
ds.to_netcdf('data.nc')

Common Issues and Solutions

Issue: Slow Performance

Diagnosis: Check chunk size and alignment

print(z.chunks)  # Are chunks appropriate size?
print(z.info)    # Check compression ratio

Solutions:

• Increase chunk size to 1-10 MB
• Align chunks with access pattern
• Try different compression codecs
• Use Dask for parallel operations

Issue: High Memory Usage

Cause: Loading entire array or large chunks into memory

Solutions:

# Don't load entire array
# Bad: data = z[:]
# Good: Process in chunks
for i in range(0, z.shape[0], 1000):
    chunk = z[i:i+1000, :]
    process(chunk)

# Or use Dask for automatic chunking
import dask.array as da
dask_z = da.from_zarr('data.zarr')
result = dask_z.mean().compute()  # Processes in chunks

Issue: Cloud Storage Latency

Solutions:

# 1. Consolidate metadata
zarr.consolidate_metadata(store)
z = zarr.open_consolidated(store)

# 2. Use appropriate chunk sizes (5-100 MB for cloud)
chunks = (2000, 2000)  # Larger chunks for cloud

# 3. Enable sharding
shards = (10000, 10000)  # Groups many chunks

Issue: Concurrent Write Conflicts

Solution: Design workflows so each process/thread writes to separate chunks. Zarr-Python 3 does not yet support ThreadSynchronizer / ProcessSynchronizer; see references/v3_migration.md.

Additional Resources

Bundled references

File	Contents
references/api_reference.md	Function signatures, stores, codecs, indexing
references/v3_migration.md	Zarr-Python 2→3 breaking changes and WIP features

Official upstream

• Documentation: https://zarr.readthedocs.io/en/stable/
• 3.0 migration guide: https://zarr.readthedocs.io/en/stable/user-guide/v3_migration/
• Storage backends: https://zarr.readthedocs.io/en/stable/user-guide/storage/
• Zarr specifications: https://zarr-specs.readthedocs.io/
• GitHub: https://github.com/zarr-developers/zarr-python
• Developer chat: https://ossci.zulipchat.com/#narrow/channel/423692-Zarr-Python

Related libraries: Xarray, Dask, NumCodecs