Re: Flink CEP Resource Utilisation Optimisation

[simple]

退订



发自我的iPhone


------------------ Original ------------------
From: Abhishek Singla <abhisheksingla...@gmail.com&gt;
Date: Sun,Mar 26,2023 11:58 PM
To: user <user@flink.apache.org&gt;
Subject: Re: Flink CEP Resource Utilisation Optimisation



Hi Team,

Flink Version: 1.15.0
Java Version: 1.8
Standalone Cluster
Task Manager: AWS EC2 of Instance Type&nbsp;c5n.4xlarge (vCPU 16, Memory 42 Gb, 
8 slots per TM)
CEP Scenario: Kafka Event A followed by Kafka Event B within 10 mins
Throughput: 20k events per second for Event A, 0 for Kafka Event B
State Backend: FsStateBackend
Unaligned Checkpoints: Enabled
asynchronousSnapshots: true


While testing this (Kafka Event A followed by Kafka Event B within 10 mins) 
scenario on load environment, it took 20 nodes of TM to achieve this throughput 
otherwise either CPU utilization would reach its peak or backpressure would be 
observed because output buffers are full. The checkpoint size is only&nbsp;6.75 
GB, the state stored within the CEP operator would be much lesser as we do 
unaligned checkpointing.

I am looking for some input on if it takes this many resources to 
archive&nbsp;this throughput, and if not what probably could be the issue here.


There was one more issue that I found If the throughput of Event A goes to 
zero, then also the checkpoint size stays around 2 GB even after hours. Is this 
expected?

Regards,
Abhishek Singla