it seems that memory watermark did not work

[akaxb]

Describe the bug

server version: 3.6.6
watermark configuration：Memory limit set to 6354MB of 15886MB total(40%).
describe：
The mq process was killed by the Linux OOM, and the system log records that during the kill, the memory usage of mq is as follows:
total-vm:17600196KB,anon-rss:15291812KB

My question is, since I have set the water level of the memory(40%), why does it appear that this control has failed

Under what circumstances will memory threshold fail?

• this is system log:
  
• Before OOM, there were no particularly obvious anomalies observed in Grafana:

- the vhost messages in Grafana:

Reproduction steps

nothing

Expected behavior

MQ uses up to 40% of the host's memory

Additional context

No response

[michaelklishin]

RabbitMQ 3.6 has reached EOL in May 2018. Please upgrade, that's the only recommendation we have.

The watermark is not a hard resource limit. It is used for nodes to decide when it's time to block publishers, not when the node will stop allocating memory.

Start with understanding what consumes memory on the node.

Since 3.6, a lot has changed in the area of memory management: streams, quorum queues and CQv2 classic queues keep significantly less data in memory. Please pick the option that fits your needs.

There's also this curious case of CentOS 9 default. changes that can and will affect RabbitMQ nodes.