You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Investigate performance of decimal math / aggregation #670
Activity
[-]Investigate performance of decimal math[/-][+]Investigate performance of decimal math / aggregation[/+]andygrove commentedon Jul 17, 2024
microbenchmark results @ sf=1
andygrove commentedon Jul 17, 2024
Benchmark runs @ sf=100 suggest that reading decimal from parquet could potentially be a performance issue.
kazuyukitanimura commentedon Jul 20, 2024
33 iterations of sf=1
Pure scan without doing sum(decimals)
Scan with doing Spark sum(decimals)
So somehow once Sum is applied, Comet scan slows down...
kazuyukitanimura commentedon Jul 20, 2024
I looks like once Comet scan is enabled
GangWokeruses more time.Pure Spark scan + sum
Comet scan + Spark sum
andygrove commentedon Jul 20, 2024
I think these two screen shots are identical?
kazuyukitanimura commentedon Jul 20, 2024
Thanks @andygrove updated to the correct one
fix: dictionary decimal vector optimization (#705)
kazuyukitanimura commentedon Jul 30, 2024
After #741 there are still some issues with
ifandcasekazuyukitanimura commentedon Jul 30, 2024
After merging with latest main
fix: Remove castting on decimals with a small precision to decimal256 (…
fix: optimize some bit functions (#718)
kazuyukitanimura commentedon Aug 16, 2024
We created many fixes. I think decimals are no longer issues. closing for now
fix: dictionary decimal vector optimization (apache#705)
fix: Remove castting on decimals with a small precision to decimal256 (…
fix: optimize some bit functions (apache#718)