Pavel Stehule <pavel(dot)stehule(at)gmail(dot)com> writes:
> I am trying to fix slow query on PostgreSQL 9.5.4.
> The data are almost in RAM
If it's all in RAM, you'd likely be well-served to lower random_page_cost.
It looks to me like the planner is estimating pretty accurately how many
heap fetches will be eliminated by using the extra index; where it's off
seems to be in the cost of those heap fetches relative to the index work.
regards, tom lane