> He has PostgreSQL 10.5. I cannot to understand to too low total cost of Merge > Semi Join because subnode has very high cost 8378397.
The planner seems to be supposing that the merge will stop far short of scanning the entire LHS table, presumably as a result of thinking that the maximum value of follows.users_id_to is much less than the maximum value of tips.users_id. Given the actual rowcounts, that's seemingly not true, which suggests out-of-date stats for one table or the other.
good tip - the table follows was too small for autovacuum, and it was terrible effect. I didn't know about this optimization.