Hi team,
We are exploring flink stateful function<https://nightlies.apache.org/flink/flink-statefun-docs-release-3.2/> for one of our use case. As part of feasibility test, we are doing load testing to determine the time spent by flink app orchestrating the functions (E2E request latency - time spent by each stateful function). We have deployed sample flink app on AWS KDA, with 10 embedded stateful functions connected sequentially each doing dummy work (10ms thread sleep). We found some interesting insight where time spent from receiving the input to invoking of first embedded stateful function is very high compared to e2e time spent across invoking stateful functions. With 100TPS load, e2e time spent in orchestrating functions is around P99 2044ms and e2e orchestration time without time spent in invoking first function is P99 14ms, and there was no FLINK back pressure and CPU usage was less than 100% overall during the experiment. Can someone help with the following queries? * Is it expected for flink app to take considerable time to invoke first stateful function after receiving an request from input stream (kinesis in our case)? * Is there any way to fasten the invoking of first stateful function? * Any pointer on how to control request to be processed by only one host irrespective of no. of stateful functions the requests have to processed? Thanks Rakesh
