> This is almost certainly unrelated. It looks like that query did a seq scan > and accessed a large number of tuples (and pages from "shared_buffers"), which > the OS then shows as part of that processes memory, even though *shared* > buffers are not specific to that one process.
Yeah. This server looks highly overprovisioned, I'm in particularly suspicious of the high max_connections setting. To fetch this out I'd be tracking connections in the database, both idle and not idle, continuously. The solution is most likely to install a connection pooler such as pgbouncer.
