So, if I'm reading this correctly, you've created a hash for each catalog and an anonymous hash for the part, then inserted a reference to the anon hash into each catalog under the corresponding part number, yes? Slick... I like it and it seems like it should handle the piecemeal nature of the source data nicely as well. Thanks!

As far as the memory/disk matters, it just needs to suck in source files of the form $lester\t$wai, correlate the numbers in all the files, and spew $lester\t$wai\t$pic\t$manu to stdout. From there, I'll capture it to a text file, send it to someone else, and forget about it until they either correct an input file or it comes time to update another catalog. I anticipate that the resulting data set will run around 12,000 records, so memory's not a concern.