`writeTileDBArray()` fails for dense 3D arrays ("Subarray should have num_dims * 2 values")

[kokitsuyuzaki]

Hi,

I'm the developer of the DelayedTensor package, which performs tensor arithmetic on DelayedArray objects. I'm exploring using TileDBArray as a backend (instead of HDF5Array) to take advantage of TileDB's native sparse storage.

I found that writeTileDBArray() works correctly for 2D arrays and sparse 3D arrays, but fails for dense 3D arrays with the error:

Error: Subarray should have num_dims * 2 values: (low, high) for each dimension.

Minimal reproducible example

library(TileDBArray)

# 2D dense: OK
mat <- matrix(runif(100), nrow=10, ncol=10)
writeTileDBArray(mat)
#> 10 x 10 TileDBMatrix

# 3D sparse: OK
sparse3d <- SparseArray::randomSparseArray(c(10, 10, 10), density=0.01)
writeTileDBArray(sparse3d)
#> 10 x 10 x 10 TileDBArray

# 3D dense: FAIL
arr3d <- array(runif(1000), dim=c(10, 10, 10))
writeTileDBArray(arr3d)
#> Error: Subarray should have num_dims * 2 values: (low, high) for each
dimension.

# TileDBRealizationSink creation succeeds for 3D
sink <- TileDBRealizationSink(dim=c(10L, 10L, 10L))
#> OK

# But write_block to that sink fails
grid <- RegularArrayGrid(refdim=c(10L, 10L, 10L), spacings=c(10L, 10L, 10L))
block <- array(runif(1000), dim=c(10, 10, 10))
write_block(sink, grid[[1L, 1L, 1L]], block)
#> Error: Subarray should have num_dims * 2 values: (low, high) for each
dimension.

The error seems to come from the dense path of write_block() in TileDBRealizationSink.R (around line 240), where do.call("[<-", args) calls into the tiledb package's subarray write functionality. The sparse path uses a data.frame-based write and works fine.

It's possible this is actually a tiledb R package issue rather than TileDBArray, but I'm reporting here since the sparse path works. Could the dense path perhaps use selected_ranges() as recommended in TileDB-Inc/TileDB-R#259?

Use case

DelayedTensor needs 3D+ array support for tensor decomposition and arithmetic operations. Being able to use TileDBArray as a backend would be very beneficial for sparse tensor workloads, since HDF5 stores everything as dense internally.

Session info

R version 4.5.3 (2026-03-11)
Platform: x86_64-conda-linux-gnu

TileDBArray 1.20.0
tiledb      0.34.0
DelayedArray 0.36.0
SparseArray  1.10.8
S4Arrays     1.10.1

[LTLA]

Yes, it seems to be a problem with the underlying tiledb package. I've switched to selected_ranges()<- as a workaround in 1.21.1, and it seems to work, but I'd suggest testing it on some larger arrays to check that it does what I think it does. For example, I'm not completely convinced it will do the right thing if your underlying array is not COLUMN_MAJOR.

Ideally I'd use the data.frame-based assignment for both dense and sparse writes, given that (i) [<-,tiledb_array converts my array into a data frame internally anyway, and (ii) the array assignment doesn't let me specify the attribute I want to write to. But [<- doesn't seem to work correctly when the value is a data frame to be written to a dense array.

In general, it's kinda hard to figure out what [<- is even doing, especially when considering interactions with other methods like selected_ranges<-... I wish they'd make it less magical.

[kokitsuyuzaki]

Thanks for the quick fix! I've confirmed that writeTileDBArray() for 3D dense arrays now works correctly in 1.21.1.

During further testing, I found two separate issues and filed them individually:

• write_block fails for 2D dense TileDBRealizationSink with RegularArrayGrid #31: write_block regression for 2D dense RegularArrayGrid
• as.array() on TileDBArray returns list(NULL, ...) dimnames instead of NULL #32: as.array() returns list(NULL, ...) dimnames instead of NULL

I think this issue can be closed now.