I just finished writing a group of ETL scripts in Perl. The longest-running script took eight hours to run, and had been optimized a great deal, with caching and other tricks. At the other end of the scale was a more recent script that ran in six seconds -- so I didn't waste any time trying to speed that one up.
In any case, these are throw-away scripts, since our product was the documentation about how the transformations were achieved, and once our output files matched the standard files, no further work was necessary -- the production transformations are being written in Informatica, based on our documentation.
I've never worked at a scale where I had to stop and look at the efficiency of the algorithm was necessary -- but when that happens, I'll certainly pay attention to it.