dyna-zarr

A lightweight, dask-free Python library for lazy, memory-bounded operations on large Zarr (and TIFF) arrays, with an optional GPU path.

Overview

dyna-zarr is a thin, pull-based array layer over Zarr. Instead of building a task graph, every operation is a lazy transform whose read(key) maps an output slice back to a bounded input read, ending at a direct zarr/TensorStore read. Slicing a result pulls only that region through the whole operation chain, so no intermediates are materialized.

The practical consequence is memory-boundedness. When you stream a result to disk with io.write, the array is processed region by region, so peak RAM is a function of the region and worker budget rather than the array size. This makes it possible to read, transform, and write arrays far larger than memory.

Segmenting a volume that does not fit in RAM, on a laptop:

from dyna_zarr import io, operations as ops

vol = io.read("volume.zarr")                             # nothing read yet

smoothed = ops.gaussian_filter(vol, sigma=(1, 2, 2))     # halo filter, still lazy
mask     = (smoothed > 0.5).astype("uint8")              # lazy threshold

io.write(mask, "mask.zarr", region_size_mb=64, max_workers=4)

Only the last line touches the disk. The halo filter and the threshold are fused into one pull-based chain, and io.write streams it region by region, so peak RAM is set by the region budget and not by the size of the volume: a 2 GB array costs the same as a 200 MB one. The result is identical to the equivalent eager NumPy/SciPy code, including at region boundaries.

A threshold computed from the data itself, such as smoothed.mean() + 3 * smoothed.std(), works the same way; see Reductions and statistics.

Memory-boundedness

There are two ways to run a lazy result, with different memory behavior:

• io.write(result, path) streams the result to disk region by region. Peak RAM is set by region_size_mb, max_workers and the dtypes in the chain, and is independent of the array size. This is the memory-bounded path; see What peak RAM actually is.
• result.compute() returns a single in-memory NumPy array. It materializes the whole result by design (mirroring dask.array.compute).

Every operation is memory-bounded on the io.write path except median, argmin, argmax, and unique, which are flagged in the operations catalog below.

Features

• Pull-based and lazy. Operations defer until .compute() (materialize) or io.write (stream to disk).
• Memory-bounded streaming. Region-wise io.write with per-worker memory and worker-count knobs. Even reshape and flatten stay bounded, by staging through local disk.
• NumPy-like. Operator overloads, array methods (.astype, .clip, .round), and the NumPy ufunc protocol (np.sqrt(a), np.add(a, 2)) all work on a DynamicArray.
• Rich op set. About 140 operations: pointwise ufuncs, streaming reductions, neighborhood (halo) filters, structural reshaping, differences, prefix scans, and array creation.
• Multi-format I/O. Read TIFF, Zarr v2, and Zarr v3 (local, S3/GCS, HTTP); write Zarr v2/v3 (local, S3/GCS) with optional sharding. Optionally, zarrista can serve reads and writes in place of the default.
• Optional GPU. Run an op chain on CUDA via CuPy, with a single host-to-device transfer per region.

Installation

pip install dyna-zarr

Optional GPU support (pick the extra matching your CUDA toolkit from nvidia-smi):

pip install "dyna-zarr[gpu-cu12]"   # CUDA 12.x  ([gpu] is an alias for this)
pip install "dyna-zarr[gpu-cu11]"   # CUDA 11.x
pip install "dyna-zarr[gpu-cu13]"   # CUDA 13.x (e.g. Blackwell)

Optional faster storage backend, zarrista (see zarrista backend):

pip install "dyna-zarr[zarrista]"

Quick start

Read

from dyna_zarr import io

arr = io.read("image.tiff")      # TIFF via tifffile's zarr bridge
arr = io.read("array_v2.zarr")   # Zarr v2
arr = io.read("array_v3.zarr")   # Zarr v3  (also s3://, gs://, http://)

print(arr.shape, arr.dtype, arr.chunks)

data   = arr.compute()                           # materialize the whole array
region = arr[10:20, 50:150, 100:200].compute()   # pull just this region

With the optional [zarrista] extra, a Zarr source can be read through zarrista instead, a Rust-backed Zarr implementation built on zarrs. Everything downstream is unchanged:

arr = io.read("array.zarr", backend="zarrista")   # default is "tensorstore"

Note the backend parameter is specific for a Zarr storage backend. It should not be supplied with any value when the input is a TIFF. Supplying a backend with TIFF input will result in an error.

Write (memory-bounded streaming)

from dyna_zarr import io, Codecs

io.write(arr, "out_v3.zarr", zarr_format=3)
io.write(arr, "out_chunked.zarr", chunks=(64, 64, 64), zarr_format=3)
io.write(arr, "out_float32.zarr", dtype="float32", zarr_format=3)     # cast on write
io.write(arr, "out_zstd.zarr", compressor=Codecs(compressor="zstd", clevel=5), zarr_format=3)

# memory and parallelism controls (see "What peak RAM actually is" below)
io.write(arr, "out_bounded.zarr", region_size_mb=64, max_workers=4)

io.write does not replace an existing array: writing to a path that already holds a Zarr array or group raises OutputExistsError. Choose a new path, or pass overwrite=True to delete what is there and write it anew. This is the same for both backends, locally and on S3. Only a path that already looks like a Zarr store is ever removed, so a mistyped path cannot delete an unrelated directory: if it holds other files, dyna refuses to write there at all.

Unless told otherwise, the output also keeps the input's Zarr format, compressor and sharding, so reading an array and writing it back gives the same kind of store. A source that has no Zarr format of its own (a NumPy array, a TIFF, a generated array such as ops.zeros) is written as Zarr v3 with Blosc/LZ4. Passing zarr_format=, compressor= or shard_coefficients= overrides each of them.

Output chunks

By default, the output keeps the input's chunks. A read-modify-write round trip preserves the chunk grid, and the output layout is what you already know about the input. Inheriting the grid also inherits its problems, so dyna prints a note when the inherited chunks are very large (over 8 MiB, which sets the memory floor of the write, since a region always holds at least one whole chunk) or very small (under 64 KiB, which means a great many small files).

There are two ways to choose a different chunk, and passing both is an error:

• chunks= is an exact shape, used as given.
• chunk_size_mb= is a size budget. The chunk is then solved for, together with the read region, from sizes that divide or are multiples of the input chunks, so reads stay aligned to the source grid. Because only those sizes are allowed, the result is approximate, and dyna prints a note when it misses the budget.

What chunk_size_mb=1 gives for a few sources:

When the source has no chunk grid to align to (a reshaped or padded array, for example), the budget alone decides. The chunk gets power-of-two sides, as close to a cube over the last three axes as the budget allows (both axes for a 2D array), with every leading axis at 1, on the assumption that the trailing axes are spatial and the leading ones are t/c. A cube rather than a slab matters for what comes next: a (1, 256, 1024) chunk of the same size would force a full XY plane to be read for every step along Z. With no chunk_size_mb= in that case, the budget is 1 MiB: (64, 64, 64) for float32, (64, 128, 128) for uint8, (32, 64, 64) for float64. Generated arrays (ops.zeros, ops.random, ...) get the same grid unless you pass chunks= to them.

io.write(arr, "out_1mb.zarr", chunk_size_mb=1)          # budget, aligned to the source
io.write(arr, "out_exact.zarr", chunks=(32, 128, 128))  # exact

Zarr v2 is written the same way, with zarr_format=2:

io.write(arr, "out_v2.zarr", zarr_format=2, chunks=(8, 64, 64))
io.write(arr, "out_v2_zstd.zarr", zarr_format=2,
         compressor=Codecs(compressor="blosc", cname="zstd", clevel=5))

Sharding is v3 only. shard_coefficients is given in units of chunks, not in voxels, so the shard shape is the product of the two:

# chunks (8, 64, 64) x coefficients (2, 2, 2) -> shards of (16, 128, 128)
io.write(arr, "out_sharded.zarr", zarr_format=3,
         chunks=(8, 64, 64), shard_coefficients=(2, 2, 2))

Sharding keeps the chunk as the unit of compression and random access while storing many chunks per file, which suits object stores and filesystems that dislike very large numbers of small files. Supported only with zarr_format=3.

Regions vs chunks

An important concept in dyna-zarr is the concept of regions. Regions are subsets of an array, similar to chunks. The two, however, are completely different:

• chunks is the storage grid: what lands on disk.
• region_shape (or region_size_mb) is the read grid: how much is pulled through the operation chain at a time, and therefore what bounds peak RAM. Typically a region contains multiple chunks. Regions are purely a performance parameter and different region sizes supplied to the write function will lead to byte-identical output.

By default, region_size_mb controls the region size (8 MiB if not given). The region shape is then chosen automatically from that budget, the output chunks, the input chunks and the widest dtype in the chain. The other parameter, region_shape, sets the dimensions of the region directly. The two are alternatives: passing both is an error.

When regions must be aligned to another particular grid, region_shape can be useful. It must be a per-axis multiple of chunks, and is clamped to the array shape.

# e.g. a tile-wise producer whose processing grid is (8, 256, 256)
io.write(result, "out_tiled.zarr", chunks=(1, 64, 64), region_shape=(8, 256, 256))

The zarrista backend operates through the same function call, with the streaming pipeline, chunking, sharding, and compression options:

io.write(arr, "out_zarrista.zarr", zarr_format=3, backend="zarrista")

An important point is that, with backend="zarrista" writing a sharded output, the region is additionally grown to whole output shards, since the shard is the contention unit there. This applies to a region_shape you passed as well, and dyna prints a note when it happens.

Read and write backends are independent, so they can differ:

arr = io.read("s3://bucket/image.zarr", backend="zarrista")   # zarrista
io.write(arr, "out_from_s3.zarr")                             # tensorstore

In our hands, backend="zarrista" has been faster on both the read and the write paths (see zarrista backend below for the measurements and the caveats).

What peak RAM actually is

The guarantee is that peak RAM does not grow with the array. Measured on one machine with a fixed 64 MB region and a single worker, while the array grew 32x:

It flattens out: past the first few regions, doubling the array changes nothing.

What peak RAM is set by, in decreasing order of size:

• A fixed cost of roughly 200 MB per write, plus about 150 MB for importing the stack. This dominates when regions are small and no setting reduces it.
• Roughly the input plus the output region, since a region in flight is held as source data and as the result at the same time, with copies on top. Budget for about 2.5x region_size_mb per live region rather than 1x.
• max_workers, which multiplies the second term but not the first.

Two consequences worth knowing. A chain that narrows the dtype (a uint8 mask from a float32 volume) is the cheapest case, because region_size_mb is measured in the widest dtype the chain carries; a chain that widens is the most expensive. And at small region sizes the fixed cost dominates, so lowering region_size_mb below about 16 MB buys much less than it appears to.

Lazy operation chains

from dyna_zarr import io, operations as ops

arr = io.read("input.zarr")

result = ops.sqrt(ops.clip(ops.abs(arr), 0, 1))   # nothing computed yet
io.write(result, "output.zarr", zarr_format=3)     # streamed, region by region
# ...or result.compute() to materialize

NumPy-like interface

A DynamicArray behaves like a NumPy or dask array. Operators, methods, and ufuncs are all lazy:

import numpy as np

masked = (arr > 3) & (arr < 100)     # elementwise operators build a lazy mask
scaled = (arr.astype("float32") / 255).clip(0, 1)
out    = np.sqrt(np.abs(arr))        # NumPy ufunc protocol dispatches to lazy ops

Neighborhood filters

Neighborhood (halo) filters wrap scipy.ndimage. Each read pulls its own halo, so results are chunk-invariant and exact, and stay memory-bounded when streamed.

import numpy as np
from dyna_zarr import io, operations as ops

img = io.read("volume.zarr")   # e.g. (z, y, x)

# LoG filtering
log    = ops.gaussian_laplace(img, sigma=2)
io.write(log, "log.zarr", zarr_format=3)   # halo handled per region

# median denoise
denoised = ops.median_filter(img, size=3)
io.write(denoised, "denoised.zarr")

# a custom per-plane kernel
kernel   = np.ones((1, 3, 3), dtype="float32") / 9   # 3x3 mean within each z-plane
blurred  = ops.convolve(img, kernel)
io.write(blurred, "blurred.zarr")

Reductions and statistics

Reductions are lazy, like every other operation. vol.mean() returns a DynamicArray and nothing is read until it is needed. A one-element result converts on demand with float(), int() or bool().

vol      = io.read("volume.zarr")
smoothed = ops.gaussian_filter(vol, sigma=(1, 2, 2))
level    = smoothed.mean() + 3 * smoothed.std()          # lazy, 0-d
mask     = (smoothed > level).astype("uint8")            # nd compared with 0-d

io.write(mask, "mask_3sd.zarr", region_size_mb=64, max_workers=4)
print(float(level))                                      # already computed, no extra read

The threshold depends on every voxel, so the volume has to be read twice: once for the statistics and once for the write. That is what dyna does. Before the first region is written, io.write computes the small reductions in the chain, in one pass for those that read the same input (mean and std above share it). Results of up to 1 MiB, such as a 0-d statistic or per-channel values, are then kept in memory, so every region reuses them and a later float(level) costs nothing.

These results are kept with the array they were computed from, so a later smoothed.max() does not read anything either. When dyna computes any of min, max, sum or mean, it computes the other three in the same pass, since they cost very little next to reading the data. var and std need a float64 sum of squares, which costs roughly as much as the read itself, so they are only computed when you ask for them. If you do, the rest comes along. This works for reductions along any axis, as long as the whole set still fits in 1 MiB.

The cache assumes the data does not change underneath. If you rewrite a source in place, call io.clear_cache() (or arr.clear_cache() for one array) and the next read recomputes.

Shapes broadcast like in NumPy, so a reduction with keepdims=True can be combined with the full array:

centred = vol - vol.mean(axis=0, keepdims=True)

An important point is that a larger result like this one, broadcast back over the axis it reduced, is needed in full by every region along that axis. When it fits the region budget, io.write shapes the regions to span that axis, so it is still read once. When it does not, dyna recomputes it for each region and warns with the number of repeats (RepeatedReductionWarning). The result is correct either way. To compute it only once, persist it first:

m = vol.mean(axis=0, keepdims=True).persist()    # computed once, into a temporary store
io.write(vol - m, "centred.zarr")

persist() writes to a temporary Zarr store that is removed when nothing uses it anymore, or to a path you give it (persist("mean.zarr")). A result of up to 1 MiB is simply kept in memory.

histogram is lazy as well and returns (counts, edges) like numpy.histogram. Without range=, finding the range takes one extra pass. unique is lazy too, with one catch: the length of its result depends on the data. So it reads nothing until something needs its shape or values (.shape, .size, compute(), a write, or building another operation on it), and then computes both in one pass and keeps them, like the small reductions above.

Operations catalog

About 140 operations, plus the three primitives they are built from. All are available flat on dyna_zarr.operations (ops.gaussian_filter) and also grouped by category submodule (ops.neighborhood.gaussian_filter). Every operation is lazy. Every operation is memory-bounded on the io.write path except median, argmin, and argmax (see Memory-boundedness), and unique, whose memory grows with the number of distinct values.

• Pointwise, unary.
  • Arithmetic and rounding: abs, fabs, negative, positive, sign, sqrt, cbrt, square, reciprocal, floor, ceil, trunc, rint, round, clip, astype, conjugate.
  • Exponents and logarithms: exp, exp2, expm1, log, log2, log10, log1p.
  • Trigonometric and hyperbolic: sin, cos, tan, arcsin, arccos, arctan, sinh, cosh, tanh, arcsinh, arccosh, arctanh, deg2rad, rad2deg, degrees, radians.
  • Floating-point tests: isfinite, isinf, isnan, signbit, spacing.
• Pointwise, binary.
  • Arithmetic: add, subtract, multiply, divide, floor_divide, mod, remainder, fmod, power, float_power, maximum, minimum, fmax, fmin, hypot, arctan2, copysign, nextafter, logaddexp, logaddexp2, heaviside, ldexp.
  • Integer and bitwise: gcd, lcm, bitwise_and, bitwise_or, bitwise_xor, invert, left_shift, right_shift.
  • Comparisons: greater, greater_equal, less, less_equal, equal, not_equal.
  • Logical: logical_and, logical_or, logical_xor, logical_not.
  • Selection and binning: where, isin, digitize.
• Reductions. Streaming and memory-bounded, with axis= and keepdims=: min, max, sum, prod, mean, any, all, var, std. Not fully bounded (hold the full reduced axis): median, argmin, argmax. Whole-array: histogram (lazy, returns (counts, bin_edges) like numpy.histogram, memory-bounded) and unique (lazy until its length is needed, a sorted array of the distinct values, bounded by how many there are). See Reductions and statistics.
• Neighborhood (halo/overlap). gaussian_filter, uniform_filter, median_filter, minimum_filter, maximum_filter, grey_erosion, grey_dilation, convolve, correlate, laplace, gaussian_laplace, gaussian_gradient_magnitude.
• Structural. concatenate, stack, transpose, swap_axes, reshape, flatten, squeeze, expand_dims, pad, tile, roll, flip, rot90, slice_array. Slicing follows NumPy, including negative steps (arr[::-1]). pad reads only what it needs at the borders, except for the statistic modes (mean, median, maximum, minimum, linear_ramp), which need the whole axis.
• Differences. diff, gradient.
• Scan (prefix, along one axis). cumsum, cumprod, cummax, cummin. Streamed with a bounded carry on the io.write path, so memory-bounded despite the sequential dependency.
• Creation. zeros, ones, full, empty, random, and zeros_like, ones_like, full_like, empty_like. random is position-deterministic, so the result is independent of chunking.
• Primitives. map_blocks (pointwise), map_overlap (neighborhood with a halo), reduce (streaming). Use these to build your own ops.

Memory-bounded reshape and flatten

C-order reshape and flatten conflict with n-dimensional chunk layout, so a naive implementation blows up. dyna-zarr stages these through disk (a Rechunker-style two-phase, read-once/write-once copy), so peak RAM stays a function of the per-worker budget rather than the array size. When you io.write an outermost reshape or flatten, this path is used automatically:

from dyna_zarr import io, operations as ops

arr = io.read("big_4d.zarr")                 # e.g. 5 GB, awkward chunks
io.write(ops.flatten(arr), "flat.zarr", region_size_mb=128, max_workers=2)
io.write(ops.reshape(arr, (a, b)), "reshaped.zarr")   # (a, b) is any target shape of the same size

The staging happens in the system temp directory (set TMPDIR, or TEMP on Windows, to move it), never next to the output, so the output can also be remote. It needs free local disk of up to about 2x the uncompressed array for flatten and 3x for reshape. dyna checks this before anything is written and stops with an error that says how much is needed and where. An outermost scan (cumsum and friends) is streamed with a bounded carry and needs no staging. Everything else about the output (format, chunks, compression, sharding, backend, overwrite) works as for any other write. Note that region_shape, memory_budget_mb and num_readers belong to the region pipeline and are refused for these writes.

Changing only the chunk grid needs no special call: io.write(arr, path, chunks=...) streams it like any other write.

GPU (optional)

With a CuPy install, run a chain on the GPU. Setting device='cuda' on a terminal call (compute or io.write) makes device-inheriting ops run on the GPU. A single host-to-device transfer happens at the first CUDA op and the data stays resident up the chain. Results are returned or written from the host.

result = ops.gaussian_filter(arr, sigma=3)
out = result.compute(device="cuda")          # whole chain on the GPU
io.write(result, "out_gpu.zarr", device="cuda")  # per-region GPU compute, streamed write

Remote storage

Both backends read and write s3:// and gs:// arrays, and read http(s):// ones, with the same storage_options. Credentials come from the usual environment variables (and ~/.aws for S3). For any non-AWS S3 (MinIO, Ceph, ...), pass the endpoint:

opts = {"endpoint": "https://s3.example.org",
        "region": "us-east-1",
        "virtual_hosted_style_request": False}

arr = io.read("s3://bucket/image.zarr", storage_options=opts)
io.write(arr, "s3://bucket/out.zarr", storage_options=opts)
io.write(arr, "s3://bucket/out2.zarr", backend="zarrista", storage_options=opts)

Add "skip_signature": True for a public bucket without credentials. Note that TensorStore's AWS client may log that ~/.aws/config or ~/.aws/credentials cannot be found. That is harmless when the credentials come from the environment.

The default TensorStore backend understands endpoint, region, skip_signature, virtual_hosted_style_request and client_options={"allow_http": ...}, and refuses anything else with an error rather than ignoring it. It takes no credentials in storage_options, and skip_signature needs tensorstore 0.1.72 or newer. It cannot write over plain http(s):// (its HTTP store is read-only) and has no store for az://. backend="zarrista" covers both; see below.

zarrista backend (optional)

zarrista is a Rust-backed Zarr implementation (built on zarrs). Installed via the [zarrista] extra, it can serve reads and writes in place of the default TensorStore path.

Measured against that default on the same machine, with matched codecs, matched concurrency, and byte-identical output verified in both directions:

Single machine, local filesystem, float32, so treat these as an indication, not a guarantee, and measure on your own data and hardware. The scripts are in benchmarks/.

Note that the default backend is tensorstore. Zarrista backend needs to be explicitly supplied. Simply pass the backend="zarrista" to the reader:

arr = io.read("image.zarr", backend = "zarrista")

With zarrista, remote stores (s3://, gs://, az://, http://) are read and written through zarrista's async API over obstore, which the [zarrista] extra installs. Credentials come from obstore's usual environment variables. It is possible to pass custom options with storage_options, which then goes directly to obstore.store.from_url:

opts = {"endpoint": "https://s3.example.org",       # any non-AWS S3: MinIO, Ceph, ...
        "region": "us-east-1",
        "virtual_hosted_style_request": False,
        "skip_signature": True}                     # public bucket, no credentials

arr = io.read("s3://bucket/image.zarr", backend="zarrista", storage_options=opts)
io.write(arr, "s3://bucket/out.zarr", backend="zarrista", storage_options=opts)

A public bucket also works addressed directly by its https:// URL, with no options at all:

arr = io.read("https://s3.example.org/bucket/image.zarr/0", backend="zarrista")

obstore picks the store from the scheme: https:// gives an HTTP store, which serves ordinary key GETs and PUTs, enough to read and write an array, while s3:// with endpoint= gives a real S3 client. Prefer the s3:// form when you need credentials, signing, or anything that lists keys, since an HTTP store cannot list an S3 bucket (it would need WebDAV PROPFIND).

Relationship to dask

dyna-zarr is not a general replacement for dask.array. It targets one job: memory-bounded read, transform, and write of large Zarr/TIFF arrays.

The core idea is to drop the task graph. Because every operation is a pull-based chain, where each output slice maps back to a bounded input read, there is no graph to build and no scheduler to run it. That keeps the engine small, keeps peak RAM bounded by the region budget rather than the array size on the io.write path, and avoids scheduling overhead, which makes the read-transform-write pipeline efficient.

The tradeoff is that only operations that fit this slice-pushdown model belong in the chain: pointwise math, neighborhood/halo filters, streaming reductions, and structural reshaping. These are operations that are commonly used in image processing, which is what dyna-zarr is mainly built for. Operations that would need a global, data-dependent graph do not fit directly, and a few that do (such as non-associative reductions) trade extra reads or memory to stay correct.

Two more differences worth knowing:

• Single machine, for now. Parallelism today is threaded I/O within one process, plus the optional GPU path. There is no cluster or distributed execution yet; better and process-based parallelism is a possible future direction.
• Narrower surface. About 140 operations today, extended where the slice-pushdown model permits.

Core components

• io.read(source, backend=..., storage_options=...) reads TIFF, Zarr v2, or Zarr v3 (local or remote) into a DynamicArray; backend is "tensorstore" (default) or the optional "zarrista".
• io.write(array, path, ...) streams a DynamicArray to Zarr v2/v3, locally or remotely. Output: zarr_format, chunks or chunk_size_mb, shard_coefficients, compressor, dtype, overwrite. Pipeline: region_size_mb or region_shape, max_workers, num_readers, memory_budget_mb, device. Storage: backend, storage_options.
• io.clear_cache() forgets the cached small reductions (see Reductions and statistics).
• operations is the lazy op set above.
• DynamicArray is the pull-based lazy array (slicing, .compute(), .persist(), operators with NumPy broadcasting, lazy reductions such as .mean(), .astype/.clip/.round, ufunc protocol, .shape/.dtype/.chunks/.size/.nbytes).
• Codecs is the compression configuration for Zarr v2 and v3.
• backends.zarrista_backend implements the optional backend="zarrista" path used by io.read and io.write.

Running the tests

[dev] installs the tooling only, including moto, an in-process S3 emulator for the remote-storage tests, so they never touch a real cloud store. Tests for the optional backends are marked, so they are selected explicitly rather than skipped silently:

pip install -e ".[dev]"
pytest -m "not zarrista and not gpu"    # core suite

pip install -e ".[dev,zarrista]"
pytest -m zarrista                      # the optional storage backend

pip install -e ".[dev,gpu-cu12]"
pytest -m gpu                           # needs a CUDA device

CI runs the core suite on Linux, Windows, and macOS across Python 3.11-3.13, the zarrista suite on all three OSes as a separate job, and a wheel install-check on each OS. A separate job installs every dependency at its minimum version on Python 3.11 and runs the suite there, so the version floors below are tested, not guessed. The GPU tests are not run by CI, since GitHub's standard runners have no CUDA device, so run -m gpu locally on each platform you support before releasing a change to the device path.

Requirements

• Python 3.11 or newer
• zarr 3.1.5+, numpy 1.26+, scipy 1.10+, tensorstore 0.1.62+, tifffile 2025.5.21+
• tensorstore 0.1.77 to 0.1.84 are excluded: after any S3 access they crash when the Python process exits (seen on Windows), which turns a finished job into a failure.
• Optional: CuPy (via the gpu-cuXX extras) for the GPU path
• Optional: zarrista + obstore (via the zarrista extra) for the faster storage backend; this extra also needs numpy 2.1+