Scott Carey <scott(at)richrelevance(dot)com> writes:
> I consistently see HashJoin plans that hash the large table, and scan
> the small table.
Could we see a self-contained test case? And what cost parameters are
you using, especially work_mem?
> This is especially puzzling in some cases where I have 30M rows in the big table and ~ 100 in the small... shouldn't it hash the small table and scan the big one?
Well, size of the table isn't the only factor; in particular, a highly
nonuniform distribution of the key value will inflate the cost estimate
for using a table on the inner size of the hash. But the example you
show here seems a bit extreme.
regards, tom lane