* Josh Berkus (josh@agliodbs.com) wrote:
> Well, I think that someone needs to actually test doing a sort with,
> say, 100GB of RAM and make sure it doesn't crash. Anyone have a machine
> they can try that on?
It can be valuable to bump up work_mem well beyond the amount of system
memory actually available on the system to get the 'right' plan to be
chosen (which often ends up needing much less actual memory to run).
I've used that trick on a box w/ 512GB of RAM and had near-100G PG
backend processes which were doing hashjoins. Don't think I've ever had
it try doing a sort w/ a really big work_mem.
Thanks,
Stephen