On average, pg_uuidv7 is as fast as the native gen_random_uuid() function in
Postgres. In a worst case scenario it was up to 2% slower, however run-to-run
variations of pgbench were >2%, which likely means that in a real system the
performance impact of pg_uuidv7 is negligible.
Performance benchmarks were evaluated using Postgres 17 and pgbench. The
following functions were benchmarked:
- the native
gen_random_uuid()function which is built in since Postgres 13 - the
uuid_generate_v7()function from this extension (pg_uuidv7) - a pure sql version of
uuid_generate_v7()from kjmph
See the benchmark directory for the implementation.
pgbench --client=8 --jobs=8 --transactions=200000 --file=${TEST}.sql
-- SELECT gen_random_uuid();
scaling factor: 1
query mode: simple
number of clients: 8
number of threads: 8
maximum number of tries: 1
number of transactions per client: 200000
number of transactions actually processed: 1600000/1600000
number of failed transactions: 0 (0.000%)
latency average = 0.073 ms
initial connection time = 4.815 ms
tps = 110325.912399 (without initial connection time)
-- pg_uuidv7 C extension
scaling factor: 1
query mode: simple
number of clients: 8
number of threads: 8
maximum number of tries: 1
number of transactions per client: 200000
number of transactions actually processed: 1600000/1600000
number of failed transactions: 0 (0.000%)
latency average = 0.073 ms
initial connection time = 5.422 ms
tps = 109623.347197 (without initial connection time)
-- sql function r27
scaling factor: 1
query mode: simple
number of clients: 8
number of threads: 8
maximum number of tries: 1
number of transactions per client: 200000
number of transactions actually processed: 1600000/1600000
number of failed transactions: 0 (0.000%)
latency average = 0.088 ms
initial connection time = 5.146 ms
tps = 90448.451336 (without initial connection time)