allspaw at gmail.com
Mon Nov 29 21:32:22 PST 2004
yeah, it might not be 100% clear what I'm asking about, without going
into a huge amount of detail about the specific situation, and I'm
pretty sure that I can't give that amount of detail, but yes, it does
include somewhat insane amounts of traffic, and we're using load
balancers not only in the front but almost everywhere.
I've gotten a lot of excellent and definitely helpful comments
here...if I can post any more info later, I will for sure.
thanks mucho, all!
On Mon, 29 Nov 2004 21:27:01 -0800 (PST), Brad Fitzpatrick
<brad at danga.com> wrote:
> Local port exhaustion is an annoying problem, but easy to solve: keep
> your connections to memcached servers open a long time.
> (sorry, not totally following this thread...)
> On Mon, 29 Nov 2004, John Allspaw wrote:
> > "which seems unlikely short of truly mammoth
> > levels of traffic."
> > exactly! :)
> > suffice to say that it's socket exhaustion that is driving my question.
> > --john
> > On Mon, 29 Nov 2004 23:22:53 -0500, Perrin Harkins <perrin at elem.com> wrote:
> > > On Mon, 2004-11-29 at 17:17 -0800, christopher at baus.net wrote:
> > > > >
> > > > > you might be better off using libhttpd or libwebserver (or whatever) and
> > > > > having memcached just reading HTTP requests, instead of adding a
> > > > > apache-module to proxy requests through to it.
> > > >
> > > > To me it just sounds like you are going to end up reimplementing something
> > > > like squid.
> > >
> > > I agree. The most interesting feature of memcached is the fancy non-
> > > blocking I/O that allows it to scale well. HTTP would have to be
> > > implemented to work in the same aysnchronous way to make it worth doing.
> > >
> > > This feature request seems to be driven by a desire to use an HTTP load
> > > balancer in front of memcached, but memcached can already be balanced
> > > across multiple machines with the hashing in the client libraries. The
> > > only reason to do HTTP-style load balancing (i.e. balancing across a
> > > bank of identical machines) would be to handle memcached servers maxing
> > > out on CPU or connections, which seems unlikely short of truly mammoth
> > > levels of traffic.
> > >
> > > - Perrin
> > >
> > >
More information about the memcached