On Fri, 2022-09-02 at 09:30 +0200, Drouvot, Bertrand wrote:
> Hi,
>
> I'm not sure we are choosing the right victims here (aka the ones
> that are doing the request that will push the total over the limit).
>
> Imagine an extreme case where a single backend consumes say 99% of
> the limit, shouldn't it be the one to be "punished"? (and somehow forced
> to give the memory back).
>
> The problem that i see with the current approach is that a "bad"
> backend could impact all the others and continue to do so.
>
> what about punishing say the highest consumer , what do you think?
> (just speaking about the general idea here, not about the implementation)
Initially, we believe that punishing the detector is reasonable if we
can help administrators avoid the OOM killer/resource starvation. But
we can and should expand on this idea.
Another thought is, rather than just failing the query/transaction we
have the affected backend do a clean exit, freeing all it's resources.
--
Reid Thompson
Senior Software Engineer
Crunchy Data, Inc.
reid.thompson@crunchydata.com
www.crunchydata.com
