Correction

Shortly after I published Benchmarking UUIDs, Per Wigren emailed me with a correction. It turns out the approach Jonathan and I used to time how long PostgreSQL takes to generate a million UUIDs is mostly timing how long it takes to generate a million queries:

DO $$
BEGIN
  FOR i IN 0..1000000 LOOP
    PERFORM 1;
  END LOOP;
RETURN;
END;
$$;

They pointed out a better way to test:

SELECT COUNT(*)
FROM (
  SELECT 1 FROM generate_series(1, 1000000)
) AS x;

This results in a roughly order-of-magnitude difference in test times, just in overhead.

When we take this insight and apply it to the two UUID generator functions, we find that PostgreSQL is faster at this task than nodejs:

SELECT COUNT(*)
FROM (
  SELECT uuid_generate_v4() FROM generate_series(1, 1000000)
) AS x;

SELECT COUNT(*)
FROM (
  SELECT gen_random_uuid() FROM generate_series(1, 1000000)
) AS x;

On my machine, I see a big difference between the two functions, more than 5x:

Interestingly, on Per Wigren's machine, running macOS Sierra with PostgreSQL 9.6.2 installed from Homebrew, the two functions were approximately equally fast, with uuid_generate_v4 slightly edging out gen_random_uuid. Both were faster than the nodejs version.

Conclusion

• Writing benchmarks is tricky!
• Using this updated methodology, on my machine, PostgreSQL with pgcrypto is faster at generating UUIDs than nodejs, which in turn is faster than PostgreSQL with uuid-ossp.

Thankfully, I don't think the flaw in my original measurements undermines the conclusion I drew: the difference between these methods is vanishingly small, and the likelihood that generating UUIDs is the bottleneck in your system is low. Better to focus your optimization efforts elsewhere!

Many thanks to Per Wigren for the feedback!

2017-04-14: Updated to credit Per Wigren and clarify the table of new measurements by adding the extension to the title of the columns and including the nodejs measurements from the previous post.