Memcached modules behaviour

[Pedro Mata-Mouros Fonseca]

Thanks for the explanation. In the meantime I kind of went around the  
problem, using the nginx consistent hash module for memcached upstream  
gets and also that same algorithm in PHP's sets. Working flawlessly.

Thanks again!

Pedro.

On Nov 12, 2009, at 1:51 AM, agentzh wrote:

> On Tue, Nov 10, 2009 at 2:23 AM, Pedro Mata-Mouros Fonseca
> <pedro.matamouros at sapo.pt> wrote:
>> I have a pool of two memcached servers that I use in nginx for  
>> querying
>> (using Valery's eval module) and, in case of cache miss, for  
>> setting with
>> the key/value pair in a PHP script. How does nginx/eval module  
>> query this
>> memcached pool?
>
> The nginx "memcached" module will first try the first upstream
> memcached server, if hit by any of the error conditions specified by
> "memcached_next_upstream", then try the next. (The actual logic is
> actually implemented in the "upstream" module which "memcached" is
> relied on.)
>
>> Any kind of hashing is done? Round-robin?
>
> Nope, no magic will happen, completely different from libmemcached ;)
>
>> While we're at it,
>> does eval module reuse the default nginx memcache module  
>> implementation?
>>
>
> Yup, behind the scene, the "eval" module creates a temporary
> "location" for the memcached settings, and directly sends a subrequest
> to that "location", and in its "post_subrequest" handler, "eval" just
> reads the buffered response generated by the "memcached" module (well,
> it requires the actual upstream module supports "in-meomory
> subrequests" and "memcached" fortunately meets the requirement here
> but "fastcgi" does not IIRC.)
>
>> The original problem is that I'm having memcached key/value sets in  
>> PHP
>> being done to a specific server (some key hashing algorithm is used  
>> with the
>> PECL memcache extension), while later on nginx appears to be  
>> querying both
>> servers (causing a cache miss when contacting the wrong server).
>>
>
> This is the expected behavior from the perspective of the actual
> implementation :)
>
>> Any suggestions on how to solve this? Right now, the options I'm  
>> seeing are:
>> 1. extend either nginx or PECL's memcache (or memcached)
>
> I'd vote for this option. And I'm going to work on this myself because
> it's now part of my $work.
>
> A note to the implementation details: be careful to use exactly the
> same escaping function as the nginx memcached module, or you'll have
> mysterious cache miss even if you get the key-hashing-to-host process
> right. Please ensure you have grokked the source of the "memcached"
> module (and possibly the "eval" module). They're *very* concise after
> all ;)
>
> Hope this helps :)
>
> Cheers,
> -agentzh
>
>