Hello,
I am having trouble gracefully draining connections from httpd when we have
KeepAlive turned on. The problem seems specific to idle KeepAlive
connections. It looks like when a process has an established KeepAlive
connection and we send the main httpd process a WINCH or a USR1 signal, the
KeepAlive connection will stay alive until it gets another request, and
then once it is finished with that request it will shut down. However, if
it doesn't get another request on the connection, the process will hang
around fro KeepAliveTimeout since you issued the signal. I would like for
idle KeepAlive connections to shut down immediately, is this not
technically possible for some reason? If it is possible, is it a bug that
the connections stick around?
The problem with them sticking around, is that when I want to fully shut
down the server some of these KeepAlive connections might finally get a
request, and thus if I hard kill a server before waiting KeepAliveTimeout,
browsers can see errors.
When I look at the server-status page, these processes change from K to G
immediately, but will stay in the G mode the whole time they are waiting.
When the mode of operation changes the SS column (seconds since last
requests) doesn't get reset and will keep going up, so this could possibly
go all the way up to KeepAliveTimeout * 2 if you start the restart process
when a connection is near the KeepAliveTimeout.
I've done most of my testing using Apache httpd 2.2.31 with mod_jk 1.2.31,
I also did some testing with Apache httpd 2.4.23 with mod_jk 1.2.42 and saw
the same behavior. I did most of my testing using the prefork module, but
I also saw the same behavior when I twas testing with the event module.
The KeepAlive settings I was using are:
KeepAlive on
KeepAliveTimeout 60
MaxKeepAliveRequests 100
The small python script I was using to test keepAlive connections (the
behavior was the same when I was testing using browsers):
import requests
import time
s = requests.Session()
r = s.get("<server address>")
print("Sleeping")
time.sleep(500)
Thanks,
Alex Kaiser
