Repository navigation
More consistency in handling matrices in j #6055
Description
Outside grouping, matrices in j are auto-converted by as.data.table.list() so that entries are OK as long as the dimensions look "right" in the same way that as.data.table() works:
DT=data.table(a=rep(1:2, 3), b=1:6)
DT[, .(cbind(b))] # OK, 1 column output
DT[, .(cbind(b, b))] # OK, 2 columns output
DT[, .(b, cbind(b, b))] # OK, 3 columns output
DT[, .(b[1L], cbind(b, b))] # OK, 3 columns output, length-1 recycling rules
# OK-ish, 3 columns output with warning about uneven recycling, but could be error?
DT[, .(b[-1L], cbind(b, b))]Note that the case without .(...) in grouping is not interesting, since the return value is always "plain" like what you'd get with with(), i.e. the output need not be a data.table.
Inside grouping, as of #6054, we have this behavior:
DT[, cbind(b), by=a] # OK, 2 columns output
# !! inconsistent with ungrouped case: 2 columns, equivalent to `DT[, .(c(b,b)), by=a]`
# IMO consistency would say this should be DT[, .(b, b), by=a]
DT[, cbind(b, b), by=a]
DT[, .(cbind(b)), by=a] # OK, 2 columns output
# error as of #6054, but for consistency could be DT[, .(b,b), by=a]
DT[, .(cbind(b, b)), by=a]
# error, IMO this implies the last un-grouped case should also be an error -- only singletons are recycled.
DT[, .(1, 1:2, 1:3), by=a]
Activity