On Thu, 2020-07-30 at 19:16 +0200, Tomas Vondra wrote:
> > Essentially:
> > initHyperLogLog(&hll, 5)
> > for i in 0 .. one billion
> >addHyperLogLog(&hll, hash(i))
> > estimateHyperLogLog
> >
> > The numbers are the same regardless of bwidth.
> >
> > Before my patch, it takes about 15.6s.
On Thu, Jul 30, 2020 at 09:21:23AM -0700, Jeff Davis wrote:
On Wed, 2020-07-29 at 17:32 -0700, Peter Geoghegan wrote:
How did you test this? What kind of difference are we talking about?
Essentially:
initHyperLogLog(&hll, 5)
for i in 0 .. one billion
addHyperLogLog(&hll, hash(i))
estimat
On Wed, 2020-07-29 at 17:32 -0700, Peter Geoghegan wrote:
> How did you test this? What kind of difference are we talking about?
Essentially:
initHyperLogLog(&hll, 5)
for i in 0 .. one billion
addHyperLogLog(&hll, hash(i))
estimateHyperLogLog
The numbers are the same regardless of bwidt
On Wed, Jul 29, 2020 at 10:08 AM Jeff Davis wrote:
> Is there a reason that HyperLogLog doesn't use pg_leftmost_one_pos32()?
Yes: HyperLogLog predates pg_leftmost_one_pos32().
> I tried the following patch and some brief performance tests seem to
> show an improvement.
Makes sense.
How did you
Is there a reason that HyperLogLog doesn't use pg_leftmost_one_pos32()?
I tried the following patch and some brief performance tests seem to
show an improvement.
This came up because my recent commit 9878b643 uses HLL for estimating
the cardinality of spill files, which solves a few annoyances wi