parallel_setup_cost
No restart — settable per sessioncost units
The planner's assumed fixed cost of starting parallel workers.
What it does
Launching parallel workers isn't free — there's process startup and coordination overhead. parallel_setup_cost (default 1000) is how much the planner charges for that, which is why it only chooses parallelism when the work is big enough to repay it.
How to tune it
Rarely changed. For pure analytical workloads where even mid-size queries should parallelize, lower it (together with parallel_tuple_cost) — values in the 100–500 range are typical. For mixed OLTP/analytics keep it near the default 1000 so short queries stay serial. Don't push it close to 0: the planner will then pick parallelism even where worker startup clearly dominates the work.