This is the corner stone of why I originally made QueryMetrics single-threaded. 
I agree that the ability to add dimensions and metrics from parallel threads is 
convenient, but allowing to `emit()` from concurrent threads is paving the road 
to hell of literally undiscoverable and undebuggable bugs with metrics.

All exceptions related to QueryMetrics that have been floating up since the 
introduction of this class represent our own misunderstanding of the execution 
flow of some queries - how and when data is transferred between threads, in 
what specific thread pool some code is actually executed, etc. QueryMetrics 
force maintainers to refine the threading and make data transitions more 
explicit, that is a good thing. Making QueryMetrics permissive is easier but 
it's like burying head in the sand.

Where do you need parallel `emit()` and parallel `report()` is not enough?

[ Full content available at: 
https://github.com/apache/incubator-druid/pull/6402 ]
This message was relayed via gitbox.apache.org for [email protected]

Reply via email to