@JozoVilcek I understand and I agree with the importance of detached job 
submission. 
Please know that the beam flink runner (and the other runners) translate beam 
code into native code but still relies on flink engine to run the native code 
(of course). For the metrics it relies on native flink accumulators to deal 
with parallelism and merge. A lot of features of flink are deactivated by flink 
engine itself in this mode, not only accumulators: In `DetachedEnvironment` 
class in Flink code (regular Flink, not Beam Flink runner) an exception is 
thrown when you try to access flink accumulators in detached mode. The message 
states this:
`Job was submitted in detached mode. Results of job execution, such as 
accumulators, runtime, job id etc. are not available. Please make sure your 
program doesn't call an eager execution function [collect, print, printToErr, 
count].`
So, AFAIK, there is nothing to bind to in native flink for the beam metrics in 
detached mode. 
BTW: the current PR was about agnostic metrics **extraction** not collection.

@aljoscha, @StephanEwen do you have any idea about how we could support beam 
metrics in Flink in detached mode (if it is possible in the current state of 
Flink architecture)?



[ Full content available at: https://github.com/apache/beam/pull/4548 ]
This message was relayed via gitbox.apache.org for [email protected]

Reply via email to