Thomas Munro <thomas.munro@enterprisedb.com> writes:
> On Wed, Nov 22, 2017 at 7:04 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> Now, there's definitely something busted here; it should not have gone as
>> far as 2 million batches before giving up on splitting.
> I had been meaning to discuss this. We only give up when we reach the
> point when a batch is entirely entirely kept or sent to a new batch
> (ie splitting the batch resulted in one batch with the whole contents
> and another empty batch). If you have about 2 million evenly
> distributed keys and an ideal hash function, and then you also have 42
> billion keys that are the same (and exceed work_mem), we won't detect
> extreme skew until the 2 million well behaved keys have been spread so
> thin that the 42 billion keys are isolated in a batch on their own,
> which we should expect to happen somewhere around 2 million batches.
Yeah, I suspected it was something like that, but hadn't dug into the
code yet.
> I have wondered if our extreme skew detector needs to go off sooner.
> I don't have a specific suggestion, but it could just be something
> like 'you threw out or kept more than X% of the tuples'.
Doing this, with some threshold like 95% or 99%, sounds plausible to me.
I'd like to reproduce Cory's disk-space issue before we monkey with
related logic, though; fixing the part we understand might obscure
the part we still don't.
regards, tom lane