Re: OOM kill during planning - pg_terminate_backend is ignored

Hi,

On 2026-09-01 19:53:05 +0000, Floris Van Nee wrote:
> I am not sure what the right fix is.  CHECK_FOR_INTERRUPTS() in
> expression_tree_walker_impl()/expression_tree_mutator_impl() and
> copyObjectImpl() would cover it, but those are relatively hot; an
> amortised check every N nodes would be cheaper.

I think we'd need a separation between the function to start initiating a tree
walk/mutation and the function called for every recursion level. Then we can
do the CFI() in the "start a tree walk" function, without having to do it the
expression_tree_walker_impl(), once per processed node.

> Separately, nothing bounds planner memory in aggregate, so there is no
> ceiling short of the OOM killer -- but that seems like its own thread.

That's a ginormous topic, with lots of work needed to get anywhere.

Greetings,

Andres Freund