When adding columns by reference using :=
, we could simply create a new column list vector (one longer) and memcpy over the old vector, with no copy of the column vectors themselves. That requires negligible use of space and time, and is what v1.7.2 did. However, that copy of the list vector of column pointers only (but not the columns themselves), a shallow copy, resulted in inconsistent behaviour in some circumstances. So, as from v1.7.3 data.table over allocates the list vector of column pointers so that columns can be added fully by reference, consistently.
When the allocated column pointer slots are used up, to add a new column data.table
must reallocate that vector. If two or more variables are bound to the same data.table this shallow copy may or may not be desirable, but we don't think this will be a problem very often (more discussion may be required on data.table issue tracker). Setting options(datatable.verbose=TRUE)
includes messages if and when a shallow copy is taken. To avoid shallow copies there are several options: use copy
to make a deep copy first, use alloc.col
to reallocate in advance, or, change the default allocation rule (perhaps in your .Rprofile); e.g., options(datatable.alloccol=10000L)
.
Please note : over allocation of the column pointer vector is not for efficiency per se; it is so that :=
can add columns by reference without a shallow copy.