rcxl

CRAN checks	Linux	macOS	Windows
results

Read tabular data from xlsx files with a native C parser.

rcxl scans worksheet XML without building an XML document tree, using intermediate column buffers before constructing R vectors. The miniz and libdeflate decompressors are bundled, so the package has no R dependencies beyond base R.

Installation

From CRAN:

install.packages("rcxl")

The development version from GitHub:

# install.packages("remotes")
remotes::install_github("vlshields/rcxl")

Usage

read_xlsx() returns a plain data.frame, taking the first row as the header:

library(rcxl)

path <- system.file("extdata", "flights.xlsx", package = "rcxl")
df <- read_xlsx(path)
dim(df)
#> [1] 100  10

Column types are guessed from the cells in the read area; dates come back as POSIXct in UTC.

xlsx_sheets() lists sheets, and sheet takes a name or an index:

multi <- system.file("extdata", "multisheet.xlsx", package = "rcxl")
xlsx_sheets(multi)
#> [1] "Alpha" "Beta"  "Gamma"

read_xlsx(multi, sheet = "Beta")
read_xlsx(multi, sheet = 2)

read_xlsx_all() reads every sheet from a single workbook parse and returns a named list of data.frames:

sheets <- read_xlsx_all(multi)
names(sheets)
#> [1] "Alpha" "Beta"  "Gamma"

Reads can be windowed with an A1 range (column-only and row-only forms work too), or with skip / n_max:

read_xlsx(path, range = "B3:D10")
read_xlsx(path, range = "B:C")
read_xlsx(path, skip = 5, n_max = 10)

col_names accepts TRUE, FALSE (auto names, header row becomes data), or a character vector sized to the read area. col_types takes "text", "numeric", "date", "logical", "guess", "skip", or "list", recycled if length 1:

read_xlsx(path, col_names = FALSE)
read_xlsx(path, col_types = "text")                    # everything as character
read_xlsx(path, col_types = c("skip", rep("guess", 9)))  # drop the first column

Other arguments: na blanks matching cells, trim_ws trims surrounding whitespace, and name_repair is one of "unique" (default), "minimal", or "check_unique". See ?read_xlsx for details.

By default, worksheets with less than 4 MB of uncompressed XML are read serially and larger ones on several threads. Set the RCXL_THREADS environment variable to an explicit worker count; it overrides both the default count and the size threshold.

Performance

Median wall-clock time to read each file, warm cache, measured with the bench package. The timing runs do not compare outputs. Parity with readxl is checked separately: tools/parity_big.R requires an exact match, cell for cell, on the on-time file and the other large workbook, and the test suite compares the fixtures with a tolerance of 1e-9 on numeric values.

rcxl 0.1.1, readxl 1.5.0, R 4.5.0, bench 1.1.4, on an AMD Ryzen 3 7320U (8 cores) under Linux, 2026-10-15. Medians of 15 iterations (5 for the on-time file). Across three runs the rcxl medians agreed within 3% and the readxl medians within 8%, the widest spread being readxl on the on-time file (15.6 s to 16.7 s). The raw results are in tools/bench_results. The fixtures come from tools/make_fixtures.R; the on-time file is the public Bureau of Transportation Statistics on-time reporting table converted to xlsx. The script is tools/bench_mark.R.

Coming from readxl

rcxl has no guess_max or progress argument. The arguments sheet, skip, n_max, trim_ws, col_names, and col_types take the same values as in readxl, and the test suite checks skip, n_max, range, col_names, and col_types (including forced conversions) against readxl 1.5.0 on the shipped fixtures. Dates come back as POSIXct in UTC. na and range each differ in one case, and name_repair offers fewer options (see below).

Masking

Both packages export read_xlsx(). If both are attached, the one attached last masks the other:

library(readxl)
library(rcxl)
#> Attaching package: 'rcxl'
#>
#> The following object is masked from 'package:readxl':
#>
#>     read_xlsx

To be explicit about which one runs, call it with the namespace:

rcxl::read_xlsx(path)
readxl::read_xlsx(path)

License

MIT. Bundled miniz and libdeflate retain their own copyright and license terms; see inst/COPYRIGHTS.