Thank you for the patch improving the docs, I think it's a clear improvement from before.
On Thu, 18 Jul 2024, David Rowley wrote:
I considered writing about work_mem, but felt I wanted to keep it as brief as possible and just have some words that might make someone think twice. The details in the work_mem documentation should inform the reader that work_mem is per executor node. It likely wouldn't hurt to have more documentation around which executor node types can use a work_mem, which use work_mem * hash_mem_multiplier and which use neither. We tend to not write too much about executor nodes in the documents, so I'm not proposing that for this patch.
This is the only part I think is missing, since we now know (measurements in [1], reproducible scenario in [2]) that the number of partitions plays an important role in sizing the RAM of the server. It's just too big to not mention that worst case will be n_partitions * work_mem. [1] https://www.postgresql.org/message-id/d26e67d3-74bc-60aa-bf24-2a8fb83efe9c%40gmx.net [2] https://www.postgresql.org/message-id/af6ed790-a5fe-19aa-1141-927595604c01%40gmx.net I would also like to add an entry about this issue with links to the above pages, to the TODO page at [3], as this is the only bugtracker I'm aware of. Am I doing it right bringing it up for approval on this list? [3] https://wiki.postgresql.org/wiki/Todo Thanks, Dimitris
