Otherwise I think we'll continue to chase weird corner cases.
Agreed. Here's a little more detail on the case I noticed:
* postgres backend thread managing several libpq connections, with polling is_busy loop
* when client pushed a lot of log messages (say, with 'RAISE NOTICE'), the server would stall for significant periods of time, sometimes minutes -- during that time, there would be no activity in the log (the server doesn't do anything interesting outside of of the polling loop)
* checking backend server log, which relogs the client raised logs via dblink, each pause happened after roughly 16k bytes
* reducing the log verbosity made the problem go away.
