some questions that weren't FAQs

Anatoly Vorobey
Thu, 27 Nov 2003 02:29:23 +0200

On Wed, Nov 26, 2003 at 07:20:45PM -0500, Justin Matlock wrote:
> A little off topic...
> Brad... in practice, have you found it better to have lots of small 
> memcache's (like a 1gb one on each webserver), or two or three 
> "megacaches" (having multiple caches, using the max 4gb)?  You guys at 
> LJ are the only ones I know that have enough hardware -- and enough data 
> --  to have tried it both ways. :)
> On a different note -- is anyone still seeing that funky problem where a 
> memcached gets in a whacked out state, and starts chewing 100% CPU, or 
> did that get fixed, and I just missed it?

We fixed that problem on the server side about a month ago; it was 
caused by an error in libevent, and you have to use the latest 
version of libevent to be sure you never encounter this problem (it 
happened very rarely and under very stressful conditions). 

On the client side, we've been rewriting the client library (the Perl
one) to use asynchronous IO and select() instead of synchronous IO
and alarm()-based timeouts, like it does now. So far all the get 
commands have been rewritten this way (in CVS), and after this gets some 
testing on the live servers, we'll rewrite the rest of the library. This 
should make it faster and much more resilient to network problems or
servers going down for some reason or other.