Visual guide

The whole model in one sentence: indexing through LazyArray.lazy builds a view, chunk planning partitions its coordinates, and result() materializes the view. This page follows one familiar NumPy selection, source[2:5], through those stages.

The first four sections are for anyone indexing arrays: coordinates, transforms, composition, and result axes. If you are using lazy indexing rather than building a storage backend, you can stop after section four. The last two sections are for integrators: they turn a request into a chunk plan and pair each chunk read with its place in the result.

Throughout, one division of labor holds: the transform answers which values? and is independent of the backend; the reader answers how do I obtain them? and must preserve the transform exactly.

An index selects coordinates

Begin with an ordinary NumPy array. source contains the values 10 through 15. The selection source[2:5] takes source coordinates 2, 3, and 4, containing the values 12, 13, and 14.

source coordinate |  0   1   2   3   4   5
source value      | 10  11  12  13  14  15
selection         |        [12  13  14]
                            source[2:5]

result coordinate |  0   1   2
source coordinate |  2   3   4
result value      | 12  13  14

The result defines its own coordinates: 0, 1, and 2. The aligned rows make the correspondence explicit: those result coordinates receive values 12, 13, and 14 from source coordinates 2, 3, and 4.

The wrapper below gives the same familiar selection a lazy spelling. Indexing through .lazy creates view; the last line asks for its values and checks the observable NumPy result.

source = np.array([10, 11, 12, 13, 14, 15])
lazy = LazyArray.from_numpy(source)
view = lazy.lazy[2:5]

assert view.result().tolist() == [12, 13, 14]

The important first step is simply that an index describes which source values fill a result in a particular order. The next section gives the numbers on both sides of that description a precise meaning.

Coordinates are addresses

How to read a half-open interval

[0, 1) is a half-open interval: start at 0, inclusive, and stop at 1, exclusive. The [ includes the lower boundary, while the ) excludes the upper boundary. For integer coordinates, [0, 1) therefore enumerates the ordered sequence [0].

Half-openness lets adjacent slices and chunks meet without a gap or overlap. Concatenation is ordered: the first interval is followed by the second. When the first interval's exclusive stop matches the second interval's inclusive start, the shared boundary coordinate appears exactly once.

• [0, 1) -> [0] — Start at 0 and stop before 1, so the sequence contains only 0.
• [1, 3) -> [1, 2] — Start at 1 and stop before 3, so the sequence contains 1 and 2.
• concat([0, 1), [1, 3)) = [0, 3) — Append the second interval after the first. Their matching exclusive/inclusive boundary produces one continuous interval without a gap or duplicated coordinate.

Explicit coordinates, including coordinate arrays, are always ordered sequences rather than mathematical sets. Their order is semantic, and repeated coordinates remain repeated in the result.

In the transform algebra, coordinates are just integers. A negative coordinate is a real address in a domain, with the same status as zero or a positive coordinate; it is not automatically shorthand for counting backward from an array's end.

domain [-2, 3)

coordinate |   -2     -1      0      1      2
status     | address address address address address

IndexDomain makes those bounds explicit. In the example below, narrowing the domain at -1 selects the literal address -1; the wrapper at the end treats -1 the way NumPy does — as the last position.

domain = IndexDomain(inclusive_min=(-2,), exclusive_max=(3,))
assert domain.contains((-1,))
assert domain.narrow(-1).inclusive_min == (-1,)

values = np.array([10, 20, 30, 40, 50])
assert LazyArray.from_numpy(values).lazy[-1:].result().tolist() == [50]

Why carry literal coordinates at all? They let independently described regions keep stable addresses — a domain can even grow at its lower end without renumbering what is already there. The design notes work through that prepending example; nothing else in this guide depends on it.

The literal model and NumPy's positional model are both useful, but they answer different questions:

Surface	Meaning of an integer index	Meaning of -1
IndexDomain and IndexTransform	A literal coordinate in the current domain	The actual address -1, if the domain contains it
LazyArray.lazy	A NumPy-style position in the current view	The last position, normalized before it reaches the transform algebra

LazyArray uses positions because it is an array-like wrapper: each derived view starts at position zero and negative indices wrap exactly as they do in NumPy. The lower-level domain and transform types keep literal coordinates.

A transform points from the request to the source

An IndexTransform records how every coordinate in a request finds its source coordinate. For the slice from the first section, request coordinate i maps to source coordinate i + 2. This direction is deliberate: request to source, not source to request.

request coordinate | 0   1   2
                   | |   |   |
             i + 2 | v   v   v
source coordinate  | 2   3   4
source value       | 12  13  14

A transform speaks function vocabulary while this guide speaks array vocabulary. The two line up like this:

the API says	this guide says
input space (domain, input_rank)	request coordinates — the result being built
output space (output, one map per dimension)	source coordinates — where values are read

output names the output side of the coordinate function, not the data: values flow source → request, against the arrow. The neutral names exist because transforms compose — in a chain, an interior transform's output space is just the next transform's input space, neither a request nor a source.

The three map kinds, in NumPy terms

Every output dimension is produced by one of three map forms. Each has a NumPy counterpart, shown executably below. The examples share one helper — and it doubles as the answer to how a bare transform meets data at all: a reader materializes it into a buffer.

source = np.array([10, 11, 12, 13, 14, 15])


def resolve(transform: IndexTransform, values: np.ndarray[Any, Any]) -> np.ndarray[Any, Any]:
    """Materialize `transform` against `values` through the public reader."""
    out = np.empty(transform.domain.shape, dtype=values.dtype)
    numpy_reader.read_into(values, ReadContext(transform), out)
    return out

DimensionMap is an arithmetic rule — the slice above is one, mapping request i to source coordinate i + 2:

# DimensionMap is an affine rule: request i reads source offset + stride * i.
# Its NumPy counterpart is a basic slice.
sliced = IndexTransform(
    domain=IndexDomain.from_shape((3,)),
    output=(DimensionMap(input_dimension=0, offset=2, stride=1),),
)
assert resolve(sliced, source).tolist() == source[2:5].tolist()

# A negative stride walks the source backward, like a negative-step slice.
reversed_view = IndexTransform(
    domain=IndexDomain.from_shape((3,)),
    output=(DimensionMap(input_dimension=0, offset=4, stride=-2),),
)
assert resolve(reversed_view, source).tolist() == source[4::-2].tolist()

ArrayMap stores explicit source coordinates for irregular or fancy indexing; order and repeats survive into the result:

# ArrayMap is an explicit list of source coordinates; order and duplicates
# are semantic. Its NumPy counterpart is fancy indexing.
gather = IndexTransform(
    domain=IndexDomain.from_shape((3,)),
    output=(ArrayMap(index_array=np.array([4, 1, 1])),),
)
assert resolve(gather, source).tolist() == source[[4, 1, 1]].tolist()

ConstantMap fixes one source coordinate for every request cell. Whether an axis appears in the result is decided by the domain, never by the map: image[2, :] compiles to a ConstantMap(2) with no corresponding domain axis (the axis is dropped), while pairing a constant map with a length-n domain axis that no map consumes yields n cells all reading one coordinate — a broadcast, the one arrangement with no NumPy index counterpart:

# ConstantMap reads one source coordinate for every request cell. No NumPy
# selection spells this operation: source[0] drops the axis, and a repeated
# fancy index source[[0, 0, 0, 0]] matches the values but degrades the
# description to a coordinate list. The value-faithful counterpart is a
# broadcast.
repeat = IndexTransform(
    domain=IndexDomain.from_shape((4,)),
    output=(ConstantMap(offset=0),),
)
assert resolve(repeat, source).tolist() == np.broadcast_to(source[0:1], (4,)).tolist()

Together, the request domain and these per-source-dimension maps are the complete reusable description of an index.

Lazy views compose

A lazy view can be indexed again. Each step changes the request-to-source description, but it does not read an intermediate array. The chain is reduced to one direct transform from the newest request to the original source.

source[2:5][::-1][1:]

new request | intermediate view | original source
------------+-------------------+----------------
     0      |         1         |       3
     1      |         2         |       2

direct map: request i -> source (3 - i)

The executable example first selects source[2:5], then reverses that view and trims its first element:

source = np.array([10, 11, 12, 13, 14, 15])
view = LazyArray.from_numpy(source).lazy[2:5]
composed = view.lazy[::-1].lazy[1:]

assert composed.result().tolist() == source[2:5][::-1][1:].tolist()

Immediately after composed is created—and before the final result() call—its metadata is ready to inspect:

Available without reading	Value in this example
composed.shape	(2,)
composed.transform	One transform mapping request i to source 3 - i

Neither property needs source values. Composition works only on the coordinate description; the assertion's call to result() is the first operation in the example that materializes the selected data.

Stop here: the materialization boundary

Indexing through .lazy[...] never reads. These do:

• result()
• eager indexing of the wrapper: view[...]
• numpy.asarray(view), or passing the view to any NumPy function (numpy.add(view, 1) converts, and therefore materializes, the view)

Python arithmetic such as view + 1 raises TypeError instead: this wrapper defers indexing, not a general compute graph.

Nor does it write. There is no __setitem__, so view[...] = values raises TypeError too, and a wrapped source needs no __setitem__ of its own. A consumer that writes plans the selection with plan_chunks and performs its own read-modify-write, keeping chunk atomicity and concurrent-writer policy on the backend's side of the boundary.

An index defines a result array

An index chooses source points and also defines how those points are arranged in the result. In the 3-by-4 image below, image[1, :] and image[1:2, :] choose the same four source points: values 4, 5, 6, and 7.

same selected source cells

source coordinate | (1, 0)  (1, 1)  (1, 2)  (1, 3)
value             |    4       5       6       7

image[1, :]

result coordinate |    0       1       2       3
value             |    4       5       6       7
shape             | (4,); source axis 0 is omitted

image[1:2, :]

result coordinate | (0, 0)  (0, 1)  (0, 2)  (0, 3)
value             |    4       5       6       7
shape             | (1, 4); source axis 0 is retained with length 1

The integer in image[1, :] fixes source axis 0. No result coordinate varies along that axis, so it is omitted and the result shape is (4,). The slice in image[1:2, :] preserves source axis 0 as a length-one result axis, so the result shape is (1, 4).

image = np.arange(12).reshape(3, 4)
lazy = LazyArray.from_numpy(image)

integer_view = lazy.lazy[1, :]
slice_view = lazy.lazy[1:2, :]

INTEGER_RESULT = integer_view.result()
SLICE_RESULT = slice_view.result()

assert INTEGER_RESULT.tolist() == [4, 5, 6, 7]
assert INTEGER_RESULT.shape == (4,)
assert SLICE_RESULT.tolist() == [[4, 5, 6, 7]]
assert SLICE_RESULT.shape == (1, 4)

None inserts a new length-one axis without selecting different source points. Here it produces the shape (4, 1):

inserted_view = lazy.lazy[1, :, None]
INSERTED_RESULT = inserted_view.result()

assert INSERTED_RESULT.tolist() == [[4], [5], [6], [7]]
assert INSERTED_RESULT.shape == (4, 1)

A request becomes a chunk plan

Continue with the 3-by-4 image and image[1, :] introduced above. Giving the image a 2-by-2 chunk shape does not change the four selected values or their order. It changes only how the work is divided: columns 0 and 1 come from chunk (0, 0), while columns 2 and 3 come from chunk (0, 1).

                    column
                0    1  |  2    3
              ----------+----------
row 0            0    1 |   2    3
row 1           [4]  [5]|  [6]  [7]   <- image[1, :]
              ----------+----------
row 2            8    9 |  10   11

                       left part             right part
chunk_coords           (0, 0)                (0, 1)
global chunk_domain    [0,2) x [0,2)         [0,2) x [2,4)
selected global cells  (1,0), (1,1)          (1,2), (1,3)
chunk-local cells      (1,0), (1,1)          (1,0), (1,1)
request coordinates    0, 1                  2, 3

Every planned chunk keeps three coordinate frames distinct:

• chunk_coords identifies a cell in the chunk grid. Chunk coordinates are literal integers, so a grid that grows at its lower end can hold a chunk whose coordinate really is -1 — not an alias for the final chunk (see the design notes).
• chunk_domain gives that chunk's bounds in global source coordinates. Here the two domains are [0, 2) × [0, 2) and [0, 2) × [2, 4).
• Chunk-local positions start from zero inside each chunk. Global column 2 is therefore local column 0 in chunk (0, 1). This zero-origin local frame is separate from both the global chunk_domain and the possibly negative chunk coordinate.

plan_chunks needs only a transform and the chunk layout: one grid object per source dimension. A per-dimension grid answers four questions — which chunk contains a source index, where a chunk starts, how long it is, and the vectorized form of the first (index_to_chunk, chunk_offset, chunk_size, indices_to_chunks). The library builds these from chunk sizes via dimension_grids_from_chunks; the executable example hand-rolls one instead, to show that the whole contract is those four answers. It plans the canonical request over 2-by-2 chunks, and iterates the same plan again to show that planning is reusable. (The two transforms it inspects on each projection are the next section's subject.)

class RegularGrid:
    """A small parameterized implementation of the public grid protocol."""

    def __init__(self, size: int) -> None:
        self.size = size

    def index_to_chunk(self, index: int) -> int:
        return index // self.size

    def chunk_offset(self, chunk: int) -> int:
        return chunk * self.size

    def chunk_size(self, chunk: int) -> int:
        return self.size

    def indices_to_chunks(self, indices: NDArray[np.intp]) -> NDArray[np.intp]:
        return np.floor_divide(indices, self.size).astype(np.intp)


transform = IndexTransform.from_shape((3, 4))[1, 0:4]
grids = cast(
    tuple[DimensionGridLike, DimensionGridLike],
    (RegularGrid(2), RegularGrid(2)),
)
plan = plan_chunks(transform, grids)
PROJECTIONS = tuple(plan)
assert tuple(projection.chunk_coords for projection in plan) == ((0, 0), (0, 1))

PAIRED_DOMAINS = tuple(
    (projection.chunk_transform.domain, projection.cell_transform.domain)
    for projection in PROJECTIONS
)
assert all(chunk_domain == cell_domain for chunk_domain, cell_domain in PAIRED_DOMAINS)

The plan describes work but does not perform it. It contains no array source, storage backend, codec pipeline, buffer, or scheduler. A Zarr reader, a task queue, or a viewport can consume the same logical plan and decide independently how and when to fetch its two chunks.

On the wrapper, this partitioning is called parts: with_parts(shape) gives a LazyArray a grid of uniform boxes to divide its reads along (re-partitioning is a pure setter — it changes how a read is divided, never what result() returns), and a wrapped array advertising its own chunks is partitioned that way automatically.

A zero-length source axis has no chunks. LazyArray accepts a positive uniform part shape for that axis, or explicit per-axis spellings (), (0,), and (0, 0); each produces no parts and the same empty result. Zero-sized parts remain invalid on a nonempty axis.

One cell domain, two projections

A chunk read has to answer two questions at once: which cells belong to this chunk, and where does each of those cells belong in the requested result?

Think of a projection as a small table with one row per selected cell. For each row, chunk_transform gives the cell's zero-origin address inside the chunk, and cell_transform gives the position in the requested result that receives its value. The row numbers of that table are the shared cell domain — a synthetic input space both transforms accept, which is why one input point can be evaluated on both sides.

left chunk (0, 0)

shared cell coordinate |    0       1
cell_transform         |    v       v
request coordinate     |    0       1

shared cell coordinate |    0       1
chunk_transform        |    v       v
chunk-local coordinate | (1, 0)  (1, 1)

right chunk (0, 1)

shared cell coordinate |    0       1
request coordinate     |    2       3
chunk-local coordinate | (1, 0)  (1, 1)

The directions are exact: shared synthetic input cell domain → request via cell_transform, and shared cell domain → chunk-local via chunk_transform. Neither arrow starts at the request or maps one output space into the other.

On the wrapper, view.parts() returns one Partition per planned chunk; each bundles a sub-view of the request (.view), that chunk's projection (.projection), and the NumPy selection placing its values in the result (.out_selection).

Within one Partition, the frames divide: Partition.view.transform is a different, global transform — it maps the part view directly into the raw wrapped source — while only Partition.projection.chunk_transform uses zero-origin chunk-local coordinates. Readers receive both so the global source address and the local planning frame cannot be confused.

Projection field	What its output coordinates mean
cell_transform	Literal coordinates in the original request; its output rank is the request rank
chunk_transform	Zero-origin coordinates in the selected chunk's local frame; its output rank is the source rank

The cell domain enumerates corresponding cells; it is not itself either output coordinate space. The canonical row selection has a one-dimensional request and a two-dimensional source, so its paired projections have request rank one and source rank two.

Order and duplicates need the request-side projection

Orthogonal indexing (.lazy.oindex) applies each axis's indexer independently, like numpy.ix_ — an outer product; the pattern reference develops the dialects. It can visit source cells in an order that does not match chunk order, and it can visit one source cell more than once. In the request below, row 4 comes first and row 1 appears twice.

request position |   0      1      2
source row       |   4      1      1
result row       | row 4  row 1  row 1

A source bounding box cannot reconstruct this result. The box spanning rows 1 through 4 also includes unrequested rows 2 and 3, and its increasing coordinate order does not record that row 4 comes first. Narrowing the read to just rows 1 and 4 still does not record the second use of row 1. For the same reason, a chunk-local selector alone says which cells to read inside a chunk but cannot say which request positions receive them, especially when the chunks are processed in a different order.

The executable example assembles the 3-by-4 request from a 6-by-8 source with 3-by-4 chunks. Each Partition resolves its own sub-view — the global transform addressing the raw source — and out_selection places those values at their request-side positions; the paired projection stays available on part.projection for consumers that read chunks directly. The assertion checks the reordered, duplicated result against direct NumPy indexing.

image = np.arange(48).reshape(6, 8)
advanced = LazyArray.from_numpy(image).with_parts((3, 4)).lazy.oindex[[4, 1, 1], 2:6]

ADVANCED_EXPECTED = image[[4, 1, 1]][:, 2:6]
ADVANCED_RESULT = np.empty_like(ADVANCED_EXPECTED)
for part in advanced.parts():
    ADVANCED_RESULT[part.out_selection] = part.view.result()

np.testing.assert_array_equal(ADVANCED_RESULT, ADVANCED_EXPECTED)

The paired representation preserves information that a bounding box or local selector discards: exact request order, duplicate destinations, and the correspondence between every request position and its chunk-local source cell.

---