Thank you for this great answer, Fabian.
Regarding the yarn JVM heap size, I tried to change
And it somehow looks like working, but the actually memory needed for rocksdb still looks like a blackbox to me. I see there’s already a JIRA ticket talking about this problem, created last year and still open yet. What I can do is just keep enlarging this value until YARN don’t kill my TaskManager because of memory usage:)
By the way, my rough calculation of rocksdb memory on each TM is like
num of slots per task * num of stateful operators(including source and sink?) * (block cache size + write buffer size)
I bet it’s not correct..
This communication may contain privileged or other confidential information of Red. If you have received it in error, please advise the sender by reply e-mail and immediately delete the message and any attachments without copying or disclosing the contents. Thank you.