"Robert Haas" <robertmhaas@gmail.com> writes:
>> What's strange about it? A probe into an in-memory hashtable is a lot
>> cheaper than a probe into an index, so this type of plan makes plenty
>> of sense if the hashtable will fit in RAM and there are going to be a
>> lot of probes. (Where "a lot" means "enough to amortize the cost of
>> building the hashtable", of course.)
> Hmm... it didn't occur to me that the index probe itself might be
> more expensive than a hash probe. Is that due to concurrency control,
> or are you talking about the need to possibly read index pages in from
> disk?
Both, plus the loss of sequentiality of access to the table itself.
>>> Experimentation shows this is actually about 25% faster.
>>
>> Well, that just says your cost parameters need a bit of adjustment
>> if you'd like the planner to get the crossover point exactly right.
> Any sense of which ones might be worth fiddling with?
random_page_cost, effective_cache_size, maybe the cpu_xxx parameters.
regards, tom lane