For the nodes we have calibrated so far, we designed workloads which would return rather small result sets (i.e., < 1000 documents). That's because we noticed that for larger result sets the startup cost would become 0, as the incremental costs would dominate.

However we believe the cost model should be more accurate for queries which process a large number of documents. Especially when the startup cost is low (e.g., smaller than 100 incremental costs), we should just skip it because it's of the same order as the cardinality estimation error.

Concretely, we should revisit all calibrated nodes so far to examine whether the startup cost is negligible and we can thus examine more documents to get a more accurate value for the incremental cost.