Conversation

This Tweet was deleted by the Tweet author. Learn more

Peter Geoghegan

@petervgeoghegan

·

Jul 11, 2020

It's not the other way around? HashAggregate becomes a GroupAggregate? The latter is able to spill to disk. (But I could believe that you have it right because HashAggregate is usually much faster.)

This Tweet was deleted by the Tweet author. Learn more

Peter Geoghegan

@petervgeoghegan

·

Jul 11, 2020

You mean the sort didn't spill to disk, causing OOMs that you fixed by increasing work_mem?

This Tweet was deleted by the Tweet author. Learn more

Peter Geoghegan

@petervgeoghegan

Thanks for clearing that up!

4:36 PM · Jul 11, 2020·Twitter Web App

This Tweet was deleted by the Tweet author. Learn more

Peter Geoghegan

@petervgeoghegan

·

Jul 11, 2020

It's weird. I suspect that the planner's grouping estimates are way off here. Postgres 13 probably won't do this, since hash aggs can spill. The planner therefore doesn't try to avoid them entirely. The v13 planner would still "be wrong", but with fewer/no actual consequences.