atlas-python 0.13.1

Python bindings for the ATLAS array store

atlas-python

Python bindings for ATLAS (Aggregated Tensor Large Array Store) — a directory-based store for many similarly-shaped N-dimensional arrays, backed by local files or any object store (S3 / GCS / Azure / HTTP). Built on a Rust core with a synchronous, NumPy-native API and first-class xarray integration.

pip install atlas-python

import atlas

Extra	Install	Adds
cloud	pip install "atlas-python[cloud]"	S3 / GCS / Azure / HTTP backends via obstore

numpy, xarray, and dask are installed automatically.

Quick start

import numpy as np
import atlas

# The `with` block flushes (== close) on exit. Nothing is persisted before that.
with atlas.Atlas.create("/tmp/my_store", codec="zstd") as store:   # "zstd" | "lz4" | "none"
    ds = store.create_dataset("jan_2024")
    ds.define_array(
        "temperature",
        dtype="float32",
        dims=["lat", "lon"],
        shape=[8, 16],
        chunk_shape=[4, 8],
        fill_value=float("nan"),   # unwritten cells read back as NaN; NaN cells count as nulls in stats
    )
    ds.write_array("temperature", start=[0, 0], data=np.full((8, 16), 20.0, dtype=np.float32))
    ds.set_attribute("month", 1)
    ds.set_attribute("station", "KNMI")

# Reopen and read
store = atlas.Atlas.open("/tmp/my_store")
ds = store.open_dataset("jan_2024")
arr = ds.read_array("temperature")                    # full read -> np.ndarray
chunk = ds.read_array("temperature", [0, 0], [4, 8])  # partial read
stats = ds.array_stats("temperature")                 # {"row_count", "null_count", "min", "max"}
month = ds.get_attribute("month")                     # 1

Durability model

This is the one concept to internalise: writes are buffered in memory and only hit disk on flush().

The store's metadata is loaded once on open/create. Every subsequent mutation — creating datasets, defining arrays, write_array, set_attribute — updates in-memory state only. Nothing reaches disk until store.flush() (equivalently store.close(), or the with store: block exiting). Dropping an Atlas without flushing abandons every pending write.

The payoff: N consecutive writes amortise to a single flush — one delta file per touched array name and one metadata rewrite, no matter how many datasets you touched.

store = atlas.Atlas.create("/tmp/my_store")
# ... many create_dataset / write_array calls ...
store.flush()   # the single durability boundary

xarray integration

Importing atlas registers an accessor at xr.Dataset.atlas, so the integration is always available. The store must exist first; you then append xarray datasets to it.

import numpy as np, xarray as xr, atlas

ds = xr.Dataset(
    data_vars={
        "temperature": (["lat", "lon"], np.arange(8 * 16, dtype=np.float32).reshape(8, 16),
                        {"units": "C", "long_name": "surface temperature"}),
    },
    coords={"lat": np.arange(8, dtype=np.float32), "lon": np.arange(16, dtype=np.float32)},
    attrs={"month": 1, "station": "KNMI"},
)

with atlas.Atlas.create("/tmp/my_store") as store:
    store.add_xarray_dataset(ds, "jan_2024")     # store-side method
    ds.atlas.write(store, "jan_2025")        # xarray accessor (same effect)

# Read back as an xr.Dataset
store = atlas.Atlas.open("/tmp/my_store")
ds_back = store.open_as_xarray_dataset("jan_2024")
xr.testing.assert_identical(ds, ds_back)

Bulk ingestion

add_xarray_dataset never flushes by itself — N consecutive calls accumulate in memory and a single flush() (or the with exit) persists everything.

import glob, os, atlas, xarray as xr

with atlas.Atlas.create("/tmp/store") as store:
    for nc_path in sorted(glob.glob("*.nc")):
        name = os.path.splitext(os.path.basename(nc_path))[0]
        store.add_xarray_dataset(xr.open_dataset(nc_path), name)
# One delta file per array name across the whole batch (not one per file).

Streaming dask-backed writes

If a variable's .data is a dask.array.Array (e.g. from xr.open_dataset(path, chunks=...) or ds.chunk({...})), add_xarray_dataset / ds.atlas.write stream one dask block at a time into the store rather than materialising the whole array. The dask chunk shape becomes the on-disk chunk_shape, so the layout maps 1:1. Peak memory ≈ one chunk per variable.

ds = xr.open_dataset("big.nc", chunks={"time": 100, "lat": -1, "lon": -1})
with atlas.Atlas.create("/tmp/store") as store:
    store.add_xarray_dataset(ds, "big")     # streams chunk-by-chunk

Pass chunks={var: [...]} to add_xarray_dataset / ds.atlas.write to override the on-disk chunk shape independently of dask's chunking.

Lazy dask-backed reads

store.open_as_xarray_dataset(name) returns each variable dask-backed whenever it was stored with non-trivial chunking (chunk_shape != shape); the dask chunks tuple mirrors the on-disk chunk grid and each on-disk chunk is one dask task. Full-shape arrays (and 0-D scalars) come back eager as numpy. Call .compute() to materialise, or slice / map_blocks to operate lazily.

ds_back = atlas.Atlas.open("/tmp/store").open_as_xarray_dataset("big")
ds_back["temperature"].data              # -> dask.array.Array
ds_back["temperature"][0:100].compute()  # reads exactly one chunk

Reads run under dask's threaded scheduler only — the DatasetView captured in the graph isn't picklable, so call .compute() before handing off to distributed/multiprocessing schedulers.

How xarray maps onto the store

Item	How it's stored
Each coord / data variable	A separate array, with dims mapped 1:1.
Dataset attrs	Dataset-level (global) attributes, plain keys.
Per-variable attrs	Real per-variable attributes on each variable's array.
Per-variable _FillValue	Consumed by define_array as a typed fill value (source Dataset.attrs is not mutated). See Missing data.
Coord vs data_var distinction	JSON list in the internal _pyatlas_coords attr.
Non-scalar attr values (list, ndarray)	JSON-encoded string with a json: prefix marker.

Each add_xarray_dataset / ds.atlas.write creates a new dataset — there is no append-into-existing mode.

Missing data

When a dataset is opened with mask_and_scale=True (xarray's default), masked cells become NaN (floats) / NaT (datetimes) and _FillValue is moved into var.encoding. So those cells are recorded as null (counted in null_count, excluded from min/max stats), arrays default to a sentinel fill on write: NaN for floats, NaT for datetime64[ns], and "" for strings (integers have none). Missing string cells (None/NaN) are substituted with the string fill and a warning is emitted, since a string can't be stored as null directly.

Override per write with fill_value — a bare scalar (numeric arrays) or a {var: scalar} dict (None disables the default for that var); an explicit CF _FillValue attribute still takes precedence over the default:

store.add_xarray_dataset(ds, "jan_2024", fill_value={"temperature": -9999.0})

See the xarray guide for the full resolution order.

Supported dtypes

numpy dtype	atlas dtype
int8/16/32/64, uint8/16/32/64, float32/64	matching numeric
datetime64[ns]	timestamp_nanoseconds (aliases: timestamp_ns, datetime64[ns])
object (str/bytes), |S<n>, |U<n>	string (variable-length; reads return Python str)

• 0-D scalar arrays (shape=[]) are supported for every dtype above.
• bool is available as an attribute type but not as an array dtype.
• binary, list[...], fixed_size_list[...,N] are reserved for a later release.

Cross-dataset pruning

ds.array_stats(name) answers "what's the range of temperature in this dataset?" one at a time. To ask "which of my 10 000 datasets could contain a value above 25?" as a single vectorised scan, use the pruning index — a flattened, columnar view of the per-dataset statistics, one row per dataset:

import numpy as np

store = atlas.Atlas.open("/tmp/my_store")

# Footer only — every column's collection-wide min/max, no column data fetched.
summ = store.column_summaries()
if summ["temperature"]["max"] < 25.0:
    ...  # nothing can match — skip the column entirely

# Fetch just the columns you need; each is a dict of numpy arrays over the row space.
idx = store.pruning_index(arrays=["temperature"])
col = idx["columns"]["temperature"]

ok   = col["present"] & col["stats_valid"] & idx["live"]   # exclude gaps + deleted rows
hits = np.where(ok & (col["max"] > 25.0))[0]
candidates = [idx["datasets"][i] for i in hits]            # row -> dataset name

The return value is self-describing: idx["datasets"][i] names row i, idx["live"] is the delete mask (always & it in), and datasets that don't declare a column are explicit gaps (present is False, with row_count 0). Only the columns you ask for are fetched from storage, so the cost is independent of how many other columns the collection has. min/max keep their source type (strings come back as bytes, timestamps as int64 nanoseconds). See the stats & scans guide for the full API, including global_attrs= / array_attrs= columns.

Cloud / object storage

With the cloud extra, Atlas.open / Atlas.create accept an obstore-constructed S3 / GCS / Azure / HTTP store handle instead of a local path. The path-based local-filesystem API works without it. See the cloud storage guide.

API reference

atlas.Atlas

Method	Description
Atlas.create(path, codec="zstd")	Create a new store at path.
Atlas.open(path)	Open an existing store.
create_dataset(name) -> DatasetView	New dataset (in-memory until flush).
open_dataset(name) -> DatasetView	Existing dataset.
delete_dataset(name)	Remove a dataset (persisted on next flush).
list_datasets() -> list[str]	All dataset names.
list_arrays() -> list[str]	Distinct array names across datasets.
dataset_exists(name) -> bool	Existence check.
add_xarray_dataset(ds, name, chunks=None, fill_value=None)	Append an xarray.Dataset (does not flush). fill_value overrides the per-array fill (scalar or {var: scalar}); see Missing data.
open_as_xarray_dataset(name) -> xr.Dataset	Read a dataset back (chunked vars come back dask-backed).
open_as_many_xarray_dataset(names, concat_dim="dataset") -> xr.Dataset	Open many datasets stacked along concat_dim (eager numpy).
pruning_index(arrays=None, global_attrs=None, array_attrs=None) -> dict	Flattened statistics for only the requested columns; see Cross-dataset pruning.
column_summaries() -> dict	Every column's collection-wide min/max, read from the index footer alone (no column data fetched).
dataset_row(name) -> int | None	The dataset's fixed row ordinal in the pruning index.
row_slots() -> int	Total row slots (live + tombstoned) — the pruning index's height.
flush()	The single durability boundary — persist everything.
close()	Alias for flush(); also the with-block exit.
compact()	Reclaim tombstoned space across cached array files, and renumber row ordinals.
__enter__ / __exit__	Context-manager support (__exit__ calls close()).

atlas.DatasetView

Method	Description
name (property)	Dataset name.
list_arrays() -> list[str]	Array names in this dataset.
define_array(name, dtype, dims, shape, chunk_shape=None, fill_value=None)	Declare a new array. fill_value is a Python scalar matching the dtype; unwritten cells read back as it, and written cells equal to it count as nulls in array_stats. Dtype is enforced (TypeError on mismatch, OverflowError for out-of-range ints).
write_array(name, start, data)	Write a numpy ndarray (matching the stored dtype).
read_array(name, start=None, shape=None) -> np.ndarray | None	Read full or partial; None if the array isn't in this dataset.
delete_array(name)	Tombstone the array within this dataset.
array_meta(name) -> dict | None	{"dtype", "shape", "chunk_shape", "dimension_names"}.
array_stats(name) -> dict | None	{"row_count", "null_count", "min", "max"} — populated after flush().
set_attribute(key, value, dtype=None)	Set a dataset-level (global) attribute. Type inferred from the Python value; pass dtype to override (e.g. "int8", "float32", "timestamp_nanoseconds"). Values are stored in the _global .af file.
get_attribute(key) / attributes()	Single global attribute or dict of all.
set_array_attribute(array, key, value, dtype=None)	Set a per-variable attribute on array (e.g. units); KeyError if the array isn't defined.
get_array_attribute(array, key) / array_attributes(array)	Single per-variable attribute or dict of all, for array.

DatasetView does not expose its own flush / compact — both go through the parent Atlas.

Examples

Runnable, self-contained scripts (each writes to a temp directory):

• 01_basics.py — create a store, define arrays, set attributes, reopen, read back.
• 02_xarray.py — round-trip an xr.Dataset via both store.add_xarray_dataset(...) and the ds.atlas.write(...) accessor.
• 03_dask_streaming.py — stream a dask-chunked xr.Dataset in one chunk at a time.
• 09_missing_data.py — masked/missing cells with default NaN / NaT / "" fills, null_count, and fill_value= overrides.

Performance

ATLAS is tuned for collections of many similarly-shaped datasets. On a "1000 datasets" benchmark against netCDF4 and Zarr v3, the bulk read paths (Atlas.open_as_many_xarray_dataset / Atlas.read_array_across_stacked) beat Zarr by ~2.8× on large chunked slice reads, and on small per-dataset workloads ATLAS leads on both reads and writes. See the benchmarks for the full methodology, numbers, and an API picker for the fastest read path per workload.

Links

• Source & issues: https://github.com/maris-development/atlas
• License: Apache-2.0