Performance

pledge is benchmarked against pgledger, a double-entry ledger written in PL/pgSQL, with pgbench. Both run in the same PostgreSQL 18 container on the same laptop, with the same accounts and the same workload. Higher is better.

Transfers per second #

One client, synchronous_commit off so the numbers show the work each ledger does, not the disk:

Eight clients at once, same settings:

Bigger batches #

With durable commits on, larger batches pay off most: create_transfers writes a whole batch in one statement. Transfers per second at 8 clients, median of three runs:

Storage #

Each transfer and its two entries take about 16% less space:

Where the speed comes from #

• One statement per call. Locking the accounts, checking balances and writing the transfer and both entries happen in one SQL statement, prepared once per connection.
• No replanning. The write function always uses its cached plan; before, PostgreSQL replanned every call, which cost about 30% of the server time.
• 16-byte ids. Ids are stored as UUIDv7 values and shown as text like pla_01K9... only when you read them.

Tested for correctness #

Speed only counts if the books stay straight. Besides the unit tests, three validation suites run against a real database (just validate):

After each suite pledge.verify() passes, every currency sums to zero and every balance matches the expected value.

Run it yourself #

just docker-bench PROFILE=quick   # about a minute
just docker-bench PROFILE=cpu     # the numbers above
just validate                     # the three validation suites

Results are written to bench/results/. See bench/README.md for every option.