[rabbitmq-discuss] RabbitMQ broker's death by one cut: robustness problem

[Holger Hoffstätte]

Lev Walkin wrote:
> Speaking of OOM killer, it is not wise to disable OOM on a 
> dedicated-to-AMQ system with 16 Gigs of RAM and 4 Gigs of swap.

Yes.

> Better have erlang killed with OOM than box killed because of kernel 
> panic, or because sshd is killed (when OOM is disabled just for the beam 
> process) and you can never reach your machine again.

echo -17 /proc/<pid of process>/oom_adj will prevent a process from
getting killed. Handle with care. Good for Postgres.

One may (or may not) also want to set vm.oom_kill_allocating_task = 1 in
/etc/sysctl.conf; see
http://www.mail-archive.com/mm-commits@vger.kernel.org/msg25237.html for
what it does.

-h