Linux – uninterruptible sleep

You can tune the OOM killer, to prefer certain processes. Each process has a score that indicates the likelihood of the process being killed in case the system runs in OOM situation. You can see the score of a process in /proc/${PID}/oom_score. You can bias the decision of the OOM killer using /proc/${PID}/oom_adj: a high value increases the probability that OOM killer is going to kill ${PID}. The value of oom_adj is inherited by children so you just set it in the master apache process at startup (in this case you should hava a watchdog that restarts apache); an alternative is to leave the master process at its default level and tune up each child so that the master stays alive and re-spawns the children as needed (in this case you need an external daemon that checks for new children and tunes oom_adj). Of course this assumes that you are using the multi process worker ;-)

Please note that the normal range of oom_adj values (-16,15) is only a bias, the heuristic of the OOM killer may still choose another process if its score is high enough. The special value -17 makes the process unkillable by OOM killer, but it's dangerous, because if the unkillable process (e.g. you DB) goes berserk the kernel may be unable to recover from the memory shortage.

I would say it depends on your use case and the rest of the answers have covered this pretty well. 4G of swap are after all a cheap way to buy some safety. And I feel that this cheapness is what is making people not want to turn it off.

But let me answer with a rhetorical question. If money is not an issue, and you have a choice between two systems - one with 12G of RAM and 4G of swap, and another with 16G of RAM and no swap - which one would you choose? Unfortunately most people would still answer that they'd choose 16G of RAM and still add 4G of swap, which is missing my point.

And on another note, I personally find a swappy system worse than a crashed system. A crashed system would trigger a standby backup server to take over much sooner. And in an active-active (or load balanced setup) a crashed system would be taken out of rotation much sooner. A win for the no-swap system again.