On 01/23/2014 04:07 AM, Amos Jeffries wrote:
I've seen a lot of messages about closing old connections due to
lifetime timeout, is there any possibility that we're hitting a fd
limit? Or something else that would cause opening a connection to fail?
This is no failure. see client_lifetime directive and its value. Your
client connections are lasting that long. They are probably being
re-used the entire time for keep-alive traffic or CONNECT tunnels until
the timeout is reached which forces it to be closed.
The current default of 1 day is quite old. As the manual notes it is
supposed to be far longer than any browser needs to remain connected.
However these days you might face persistent HTTPS (via long-lived
CONNECT tunnels), long-polling connections, 24x7 mobile device
connectivity, browsers that *never* shutdown, and proxies that support
indefinite streams of keep-alive requests. Any one of which would give
you much more long lived client connections. I extended my proxies
setting to 1 week and spotted facebook connections lasting longer than
even that some time back.
Right, I know those messages aren't a failure. I was just pointing out
that we were receiving a fair number of them. I can check how many fds
are in use next time this happens.
On 01/21/2014 05:53 PM, Will Roberts wrote:
Hi,
I'm having a problem with some of my squids where they'll crash with
one of these two messages:
FATAL: dying from an unhandled exception:
AddOpenedHttpSocket(s->listenConn)
FATAL: dying from an unhandled exception: HttpSockets[NHttpSockets] < 0
I haven't seen anything on the list with that text, nor do I see any
open issues in the bug tracker. What kind of additional information
can I provide to help debug this?
I think these are another bug highlighted by how you have workers with
unique ports that the coordinator does not know about. It is a new bug
definitely but have not the time to investigate properly.
Okay. I'll file a bug with that text and poke around at the code a bit;
see if there's anything that might be worth printing out.
--Will
