Re: PostgreSQL 8.0.6 crash

Greg Stark <gsstark(at)mit(dot)edu> writes:
> I was just trying to clarify the situation since someone made some comment
> about it having to do with memory being swapped out and then finding nowhere
> to swap in when needed. That's not exactly what's happening.

No. I believe the case that is actually hard for the kernel to predict
comes from copy-on-write: when a process forks, you could potentially
need twice its current memory image, but in reality you probably won't
ever need that much since many of the shared pages won't ever be written
by either process. However, a non-overcommitting kernel must assume
that worst case, and hence fail the fork() if it doesn't have enough
swap space to cover both processes. If it does not, then the crunch
comes when one process does touch a shared page. If there is no
available swap space at that time, kill -9 is the only recourse, because
there is no way in the Unix API to fail a write to valid memory.

The reason for having a lot more swap space than you really need is just
to cover the potential demand from copy-on-write of pages that are
currently shared. But given the price of disk these days, it's pretty
cheap insurance.

regards, tom lane