Tomas Vondra <tomas.vondra@2ndquadrant.com> writes:
> I think it's really a matter of underestimate, which convinces the planner
> to hash the larger table. In this case, the table is 42GB, so it's
> possible it actually works as expected. With work_mem = 4MB I've seen 32k
> batches, and that's not that far off, I'd day. Maybe there are more common
> values, but it does not seem like a very contrived data set.
Maybe we just need to account for the per-batch buffers while estimating
the amount of memory used during planning. That would force this case
into a mergejoin instead, given that work_mem is set so small.
regards, tom lane
