What will happen if the client's DNS caches the domain name to an IP, which is then dead ? If I understand it corrently, the current system can work only if there is something like NAT in front of the machines, which dynaimcally forward each request. Won't it ?
Peter Bi ----- Original Message ----- From: "Medi Montaseri" <[EMAIL PROTECTED]> To: "Andrew Ho" <[EMAIL PROTECTED]> Cc: "Benjamin Elbirt" <[EMAIL PROTECTED]>; "modperl list" <[EMAIL PROTECTED]> Sent: Wednesday, April 17, 2002 9:08 PM Subject: Re: Sharing Variable Across Apache Children > You had us going for a while....I thought you are talking about some > distributed > session management (accross different boxes).... > > Another suggestion is to use lbnamed. lbnamed is a DNS server and Load > Balancing > server that listens to port 53 and resolves IPs, but on the other side of its > personality > it talks to bunch of agents who are running on workers. You get to set what > the > parameters or criteria is and assign a cost factor for a worker. lbnamed then > > distributes the work on the lightest/least cost worker. > > In this scenario, whence a box is out (or its critical piece like Oracle, or > HTTP server) > then no further work is routed to it . > > Also, be aware that using CNAME in DNS does not provide a uniform > distribution > of load. Imagine a web page having 20 images and another 5 images. You'll not > know > with good certainty that if your heavy work like database access is really > being > distributed. With CNAME you do have a chance of developing harmonics. CNAME > (aka round robin) is totally unaware of the load on the worker. Maybe that's > why > your boxes are bulking.... > > See http://www.stanford.edu/~schemers/docs/lbnamed/lbnamed.html > > > > Andrew Ho wrote: > > > Hello, > > > > BE>Let me explain in more detail what I'm doing. > > > > So if the situation you explain is the only reason you want a variable > > shared load balanced machines, I'd suggest a totally different way of > > doing this altogether. Best would be to use an already shared persistent > > storage mechanism (NFS or Oracle) but it looks like Oracle warnings are > > precisely what you want distinct alerts on (why are you getting so many > > Oracle errors anyway? that might be the first thing you want to check). > > > > I'm assuming the number of warning e-mails you get is less than 450,000 / > > 5 == 90,000 each day, so that if each warning e-mail were a web request, a > > single box could handle them (if more than 1/5 of your requests result in > > errors, you REALLY want to just fix the problem first). So put up a single > > webserver box to serve as an error reporter and logger. You could either > > use distributed logging (like mod_log_spread) or simpler, just set up > > another webserver that takes requests like > > /record_error.pl?error_msg=foo&remote_addr=bar or whatever. > > > > Your error handlers on your five load-balanced boxes send an HTTP request > > to this error handling box. All error e-mails can originate from this box, > > and the box can internally keep a lookup table (using any of the fine > > techniques discussed by the folks here) to avoid sending duplicate errors. > > > > In this scenario error handling is offloaded to another box, and as a > > bonus you can track the aggregate number of errors each day in an > > automated way and run reports and such (without having to count e-mails in > > your inbox). > > > > Humbly, > > > > Andrew > > > > ---------------------------------------------------------------------- > > Andrew Ho http://www.tellme.com/ [EMAIL PROTECTED] > > Engineer [EMAIL PROTECTED] Voice 650-930-9062 > > Tellme Networks, Inc. 1-800-555-TELL Fax 650-930-9101 > > ---------------------------------------------------------------------- > > -- > ------------------------------------------------------------------------- > Medi Montaseri [EMAIL PROTECTED] > Unix Distributed Systems Engineer HTTP://www.CyberShell.com > CyberShell Engineering > ------------------------------------------------------------------------- > > >