I created 4 pairs of tables with 5 million rows with same ID-s in both of the tables, that would result in a “complete” join. In addition to UUID I also included int8 and numeric for completeness. In the end it was hard to say anything meaningful as I was just guessing – but now I finally got to testing it out with a simple test setup and am now laying out the numbers for you to judge also. More specifically we talked about bigger joins (full table scans), as on indexed single row operations there shouldn’t be any difference anyways. While it is true that they do occupy some more bytes on the disk (and are a bit rare in the wild as simpler applications just don’t need them) I was still sceptical about the direct translation of those bytes to speed difference. In short he reasoned that one should generally avoid the UUID-s for performance reasons as they take 16 bytes instead of 4/8 bytes for normal integer and thus are slower. Some weeks ago at pgDay Paris, during the evening social event, we got into a small “guestimation” with another Postgres enthusiast, about the precepted speed of using integer based ID-s vs UUID based ones for bigger amounts of data.

Performance test 8 serial key