Hi,
I've got a problem with PostgreSQL 9.6.5: backend gets killed by OOM killer, and it shuts the DB down.
Of course, the OOM case is to be investigated (MemoryContextStatsDetail, etc), however I wonder if DB can be more robust.
The sad thing is a single backend crash results in the DB shutdown, so it interrupts lots of transactions.
I wonder if a GUC can be implemented, so it could fail just a single backend by limiting its memory use.
For instance: max_mememory_per_backend=100MiB.
The idea is to increase stability by limiting each process. Of course it would result in "out of memory" in case a single query requires 100500MiB (e.g. it misunderestimates the hash join). As far as I understand, it should be safer to terminate just one bad backend rather than kill all the processes.
I did some research, and I have not found the discussion of this idea.
Vladimir Rusinov> FWIW, lack of per-connection and/or global memory limit for work_mem is major PITA
Vladimir
