On Jun 21, 2018, at 1:04 PM, Jeff Davis wrote:
> On Thu, 2018-06-21 at 11:04 -0700, David Gershuni wrote:
>> This approach seems functionally correct, but I don't like the idea
>> of
>> transforming O(N) tuples of disk I/O into O(S*N) tuples of disk I/O
>> (in the
> On Jun 19, 2018, at 10:36 PM, Jeff Davis wrote:
>
> But I am worried that I am missing something, because it appears that
> for AGG_MIXED, we wait until the end of the last phase to emit the
> contents of the hash tables. Wouldn't they be complete after the first
> phase?
You're right. They'
> On Jun 13, 2018, at 12:53 PM, Jeff Davis wrote:
>
>>
>> An adaptive hash agg node would start as a hash agg, and turn into a
>> "partial hash agg + sort + final group agg" when OOM is detected.
>>
>> The benefit over ordinary sort+group agg is that the sort is
>> happening
>> on a potential
As Serge mentioned, we’ve implemented spill-to-disk for SetOps and Aggregates
at Salesforce. We were hitting OOMs often enough that this became a high
priority for us. However, our current spill implementation is based on dynahash
from 9.6, and we’re not happy with its performance (it was primar
