Lazy Delayed Arrays with Fused Execution.

delarr

delarr provides a lightweight delayed array type for R with a tidy-friendly API. It keeps the surface area small—one S3 class plus a handful of verbs—while offering fused elementwise transforms, reductions, and streamed materialisation. The package supports ordinary 2D matrices and N-dimensional arrays with length(dim(x)) >= 2. Streamed results can also be written straight to disk via the bundled HDF5 writer.

Installation

You can install the development version from GitHub with:

# install.packages("remotes")
remotes::install_github("bbuchsbaum/delarr")

Once accepted on CRAN:

install.packages("delarr")

Getting started

library(delarr)

mat <- matrix(rnorm(20), 5, 4)
arr <- delarr(mat)

# Lazy pipeline
out <- arr |>
  d_center(dim = "rows", na.rm = TRUE) |>
  d_map(~ .x * 0.5) |>
  d_reduce(mean, dim = "rows")

collect(out)

Multidimensional arrays

delarr is not limited to matrices. In-memory arrays and HDF5 datasets with 3 or more dimensions are supported too.

library(delarr)

x <- array(rnorm(3 * 4 * 5), dim = c(3, 4, 5))

# Slice lazily and operate along an explicit axis
out <- delarr(x) |>
  d_center(axis = 3L) |>
  d_reduce(mean, axis = 3L)

dim(collect(out))
#> [1] 3 4

Streaming straight to disk

# assume `X` lives inside an HDF5 file
lzy <- delarr_hdf5("input.h5", "X")

# Apply a transformation lazily and stream the result into a new dataset
# (dim(lzy)[2] supplies the total column count for the writer)
lzy |>
  d_zscore(dim = "cols") |>
  collect(into = hdf5_writer(
    path = "output.h5",
    dataset = "X_zscore",
    ncol = dim(lzy)[2],
    chunk = c(128L, 4096L)
  ))

Backends

• delarr_mem() wraps any in-memory matrix or array with at least 2 dimensions.
• delarr_hdf5() exposes a dataset through hdf5r, including N-dimensional datasets.
• delarr_mmap() streams 2D matrices from a memory-mapped binary file via the mmap package.
• delarr_backend() lets you create a seed from any (rows, cols) -> matrix pull function.
• hdf5_writer() pairs with collect(into = ...) to stream results back to disk without materialising the full matrix in memory (supply ncol to size the destination dataset up front).

The core package depends only on rlang. The hdf5r and mmap backends are optional: they live in Suggests, and the relevant constructors raise an informative error if the package is not installed. You can also add new backends yourself via delarr_backend() without taking on any extra dependency.

Pipelined verbs

• d_map()/d_map2() for elementwise transformations.
• d_center()/d_scale()/d_zscore()/d_detrend() for common preprocessing, each with optional na.rm handling. For N-d arrays, use axis =.
• d_reduce() for row-wise or column-wise reductions, or explicit axis-based reductions on N-d arrays, with streaming na.rm support for sum/mean/min/max.
• d_where() for masked updates, optionally replacing masked entries via the fill argument.
• collect() to realise the data (streamed in chunks), optionally writing to disk with hdf5_writer(), and block_apply() for chunk-wise computation.
• d_aperm() for lazy dimension permutation on N-d arrays.

All verbs return another delarr, so pipelines stay lazy until collect() materialises the result.

Testing

The test suite exercises the core class, slicing, verb fusion, reductions, chunk-aware execution, and the HDF5 streaming writer. Run it locally with:

pkgload::load_all(".")
testthat::test_dir("tests/testthat")

Roadmap

The core abstraction is stable: the in-memory, HDF5, and memory-mapped backends, the fused verb pipeline, chunk-aware collect(), the streaming HDF5 writer, and lazy matrix products (d_matmul()) are all implemented, documented, and tested. Two vignettes (vignette("delarr-getting-started") and vignette("advanced")) cover the workflow end to end, and benchmark scripts live in notes/.

Possible future directions, none of which are required for current use:

• Optional sparse-matrix adapters, where a backend can return sparse blocks without forcing them dense.
• Writer-style into= targets for N-dimensional collect() (currently supported for 2D output and via custom into = function(...) callbacks).
• Promoting the notes/ benchmarks into a dedicated performance article.