On 4.9.2014 01:34, Tomas Vondra wrote:
> On 20.8.2014 20:32, Robert Haas wrote:
>>
>> As I see it, the advantage of Jeff's approach is that it doesn't
>> really matter whether our estimates are accurate or not. We don't
>> have to decide at the beginning how many batches to do, and then
>> possibly end up using too much or too little memory per batch if we're
>> wrong; we can let the amount of memory actually used during execution
>> determine the number of batches. That seems good. Of course, a hash
Also, you don't actually have to decide the number of batches at the
very beginning. You can start start with nbatch=1 and decide how many
batches to use when the work_mem is reached. I.e. at exactly the same
moment / using the same amount of info as with Jeff's approach. No?
Tomas
