Re: BUG #3504: Some listening sessions never return from writing, problems ensue

Peter Koczan wrote:
> There is a problem where connections listening for asynchronous notifies
> occasionally block for writing on ther server side and never finish,
> resulting in connections that always have the status "notify interrupt".
> Apparently, this causes no ill effects for the application that's listening
> (i.e. it still updates somehow), but no data is written to the connection
> when it should be. It also becomes a problem for restarting the server since
> postgres cannot kill the connections when they're in that state. I either
> have to explicitly kill the connections, kill the client apps, or reboot the
> server. If I try to restart postgres, it kills all but the "notify
> interrupt" connections, but it doesn't shut down so I can't restart it short
> of rebooting.

Does the client read the async notifies? The write in the server will
block if the client doesn't keep up with reading the data.

> ...
>
> (This stack trace is of the last 3 updates, and is farily representative of
> what I've seen).
> [root(at)vero ~]# strace -p 30728

That's actually not a stack trace, but a system call trace.

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com