area/performancefor/user-attentionhelp wanted
Metriche repository
- Star
- (4748 star)
- Metriche merge PR
- (Merge medio 16g 15h) (25 PR mergiate in 30 g)
Descrizione
Current implementation uses parallel scheduler by default which is in essence a ScheduledThreadPoolExecutor which has the task queue guarded by ReentrantLock, which means two trips through said lock per signal on average
After looking at Netty's https://github.com/netty/netty/blob/4.1/handler/src/main/java/io/netty/handler/timeout/IdleStateHandler.java there appears to be a way to do fewer timeout task reschedules if you let go of the idea illustrated by the current marble diagrams (new signal cancels the previous timeout task)
WDYT? this really prevents Flux.timeout from being useful with high-frequency publishers which may occasionally stall
suggestion inspired by https://gitter.im/reactor/reactor?at=61968a44197fa95a1c7adf05