On Tue, May 8, 2012 at 11:01 AM, Mladen Turk <mt...@apache.org> wrote:
> On 05/08/2012 07:09 PM, Costin Manolache wrote: > >> IMHO neither 'graceful shutdown' nor 'deploy' are best served by not >> accepting connections: >> - for 'graceful' - a number of connections will get to backlog and >> timeout, >> > > Not sure if ServerSocket.close() would cause cascade close of all > accepted sockets that are still running. If not, even pause could > just close the socket and then init if needed. > In that case clients will not fill the backlog. You may still want to accept requests for existing sessions. Both 'graceful shutdown' and app deploy are important cases - it's just that current 'pause()' is not that good for either of them, if you really want to cleanup... Maybe an option to return 503 or custom status/headers - so people can adjust to various LBs. Costin > > > which is bad for the user. A better solution would be to support an option >> to accept and return immediately ( maybe with a way to do this only for >> requests not matching any existing session ). I never worked with a LB >> that >> detects status based only on not accepting connections and timeout. >> >> > Hmm, right, a valid option if LB won't get confused. > > > >> Besides that - is there any other use for pause() ? Maybe that's what >> should be removed/replaced :-). The behavior ( delay/timeout all TCP >> connections until backlog is full, then reject ) doesn't seem ideal. >> >> > Agreed. A 'pause' or invalidating backed node should be done in LB. > We are doing that in mod_jk, mod_proxy, mod_cluster, etc. > If LB wrongly instructs the clients to connect to a node that is going to > get shutdown we can't do much about it. Client will experience one > or other type of connection failures. > > > > > Regards > -- > ^TM > > > ------------------------------**------------------------------**--------- > To unsubscribe, e-mail: > dev-unsubscribe@tomcat.apache.**org<dev-unsubscr...@tomcat.apache.org> > For additional commands, e-mail: dev-h...@tomcat.apache.org > >
