ClusterStatusMonitor is a bottleneck in BestPossibleStateCalcStage.process

apache / helix

Mirror of Apache Helix

Apache License 2.0

454 stars 218 forks source link

Closed richardstartin closed 2 years ago

richardstartin commented 2 years ago

This profile was taken from an Apache Pinot controller:

About 18% of the samples were in ClusterStatusMonitor.setPerInstanceResourceStatus (zoomed in)

The root causes are:

inefficient hashCode implementation in PerInstanceResourceMonitor$BeanName
doing twice the work in ClusterStatusMonitor.setPerInstanceResourceStatus because containsKey is called before get, when the same effect could be achieved efficiently by computeIfAbsent

desaikomal commented 2 years ago

Richard, great finds. If you don't mind asking me, what profiling tool are you using?

richardstartin commented 2 years ago

Richard, great finds. If you don't mind asking me, what profiling tool are you using?

I'm using JFR, visualised with JMC