Details
-
Sub-task
-
Status: Closed
-
Major
-
Resolution: Won't Fix
-
1.7.0
-
None
-
None
Description
Latency metrics are currently organized as a histogram, which is significantly more expensive than other metric types.
Depending on the user's metric infrastructure a histogram in Flink is overkill; several metric backends explicitly discourage measuring histograms in the application itself. Instead only the raw latency data should be exposed.
We should add an option to expose latency not as a histogram but as a simple gauge that returns the latency based on the last received latency marker.