Throughput For Each Worker Goes Down When Decreasing Parallelism

Question

I have a simple stateless Flink application running on Kubernetes that reads from a Kafka topic and prints out content. The input TPS is consistently high (> 3K) and the number of partitions on that topic is 180. When running with higher parallelism (for example 3), I notice that each worker can process up to 1K records/sec. However, when I have a lower degree of parallelism (2 and 1), the throughput for each worker significantly goes down, sometimes even to 0 records/sec. What could be the reason for such behavior? And would it be possible to mitigate? I am running on Flink 1.12, with unaligned checkpointing enabled. I also checked if there was back pressure on any of the operators, but all show OK.

Thank you!