A few hours to a day? O, my, what kind of hardware are you running? For 2 ten million record files, it takes me less than 8 minutes to find out which records are in the second file, but not in the third. Just using shell commands (but using Perl to create the large files):

$ # Create two large files to work with.
$ perl -e 'printf "%08d\n", int rand 100_000_000 for 1 .. 10_000_000' 
+> big1
$ perl -e 'printf "%08d\n", int rand 100_000_000 for 1 .. 10_000_000' 
+> big2

# Sort them, make then unique.
$ time sort -u big1 > big1.s

real    4m0.489s
user    2m4.360s
sys     0m7.200s

$ time sort -u big2 > big2.s

real    3m24.848s
user    1m55.430s
sys     0m6.460s 

# Report the number of lines that are in the second file, and not in t
+he first
$ time comm -13 big1.s big2.s | wc -l
8611170

real    0m14.278s
user    0m12.850s
sys     0m0.400s
[download]

Total elapsed time: less than 8 minutes, of which almost half are spend in disk I/O. No point to use a database for such puny sized sets.