> I see an estimate for 1000 rows in your EXPLAIN output too, so you're experiencing the same
> although in your case the estimate of 1000 might be more accurate. The misestimation was causing
> significant performance problems for me.
> My solution was to wrap generate_series() in a custom function that had a ROWS qualifier
In my case, I think the number of rows will vary quite a bit for different time periods/resolutions (and 1000 might not be a bad estimate for some of the workloads). I do wonder whether if the planner had a sense of how big the series result could be for longer periods/finer resolutions (which is a bit of information I could actually trivially generate outside and encode into the query explicitly if need be), it might avoid/minimize the NESTED LOOP at all costs, but I'm not sure how to communicate that information.
Anyway, thank you for sharing! Very helpful to hear what other people have dealt with in similar situations.
