Regarding Consistent Hashing....

dormando dormando at rydia.net
Thu Aug 30 22:54:21 UTC 2007


>     My last job was in telecom, so I think ``standard'' is an antonym 
> for ``innovation.''
> 
>     Choice can be confusing, but performance is often more valuable than 
> interoperability here.  If you're in a java-only shop, java's native 
> hash is going to be faster than anything else.  Strings are immutable 
> and the hash is memoized.  If you have to interoperate, you can start 
> making choices.
> 
>     It's way too early to be talking about standardization, anyway.  
> There haven't been enough studies on the performance and appropriateness 
> of different hash algorithms.  When people are interested in things, the 
> last thing you want to do is standardize them.  They'll become 
> standardized on their own once people are no longer interested.
> 

(Just tossing out an idea, I've already discredited myself as useless here!)

Would a "minimal recommended feature set" be along the lines of acceptable?

I can imagine not having at least a common default to be insanity 
inducing. IMHO it's probably okay to request clients at least implement 
the crc32 over such and such details as long as they default to them. 
Then you're still free to implement something with higher performance 
characteristics.

On your last bit ... uhhhh, no strong opinion. I think the idea here is 
to do "something" which makes people feel comfortable to actually *use* 
consistent hashing in some form. Presently larger sites get bitten by 
the "I just added another memcached instance and my shit got all slow!" 
more and more often. Lets help spare sysadmins from having to do 
memcached maintenance at 4am? :)

-Dormando


More information about the memcached mailing list