I've digged into the source code a little bit and found that chain:
PLPGSQL_STMT_ASSIGN -> exec_stmt_assign() -> exec_assign_expr() -> exec_eval_expr() -> exec_run_select() -> SPI_execute_plan_with_paramlist() -> _SPI_execute_plan() which finnaly calls PushActiveSnapshot() and PopActiveSnapshot() wich just do memory context allocations and use malloc() to copy snaphot.
Andrey Zhidenkov wrote: > When I run this test in 2 threads I expect that running time will be the > same, because PostgreSQL will fork process for the second connection and > this process will be served by a separate CPU core because I have more than > 2 cores. > Yes, IMMUTABLE flag helps, but I think It's just because Postgres actually > executes procedure only once.
Just a guess without actually looking at the WaitEvents (which you should do) is that this is blocking on snapshot acquisition or something like that. -- Álvaro Herrera https://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services