28. `dplyr::arrange(DT)` uses `vctrs::vec_slice` which retains `data.table`'s class but uses C to bypass `[` method dispatch and does not adjust `data.table`'s attributes containing the index row numbers, [#5042](https://github.com/Rdatatable/data.table/issues/5042). `data.table`'s long-standing `.internal.selfref` mechanism to detect such operations by other packages was not being checked by `data.table` when using indexes, causing `data.table` filters and joins to use invalid indexes and return incorrect results after a `dplyr::arrange(DT)`. Thanks to @Waldi73 for reporting; @avimallu, @tlapak, @MichaelChirico, @jangorecki and @hadley for investigating and suggestions; and @mattdowle for the PR. The intended way to use `data.table` is `data.table::setkey(DT, col1, col2, ...)` which reorders `DT` by reference in parallel, sets the primary key for automatic use by subsequent `data.table` queries, and permits rowname-like usage such as `DT["foo",]` which returns the now-contiguous-in-memory block of rows where the first column of `DT`'s key contains `"foo"`. Multi-column-rownames (i.e. a primary key of more than one column) can be looked up using `DT[.("foo",20210728L), ]`. Using `==` in `i` is also optimized to use the key or indices, if you prefer using column names explicitly and `==`. An alternative to `setkey(DT)` is returning a new ordered result using `DT[order(col1, col2, ...), ]`.
0 commit comments