In this case, "data" is a bunch of integers. In moving from a hash to an array, the "keys" do have to be integers. I you're just counting, nothing else matters, but if it is about key-value pairs, that move is still valid if just the key is a (positive) integer. The value(s) in that pair do not have to be.

Another thing not yet mentioned is that with datasets this large, not only the data itself may put a limit on the internal available memory footprint, but the overhead in perl structures add to that. Just today I checked what the internal representation of a 1 Mb .csv file was represented as an array(ref) of array(ref)s: it grew to 10Mb! A hash takes slightly more overhead than an array (most overhead goes into converting a single number into a refcounted SV), so when on the verge of swapping, an array might actually be much faster than a hash.