It could be thrashing I suppose. I have shared_buffers and
8192 (64MB) which is not as large as it could be (although I
intentionally set it a little low since I want to catch performance problems where the db does not fit in memory).
I will bump it up and see if it makes a difference.
Maybe we should drop the container_id index since I don't think it ever has particularly a particularly good key
distribution and so
is hardly ever better than a table scan (it's got 2 distinct
values on openacs.org).
Even on sloanspace the two big container_id's make up 35% of
the rows in group_element_index.
Do we ever really query on container_id w/o also having a group_id?
