gistfile1.txt

7:54 <tsal> where did the `ring` notion come from 

07:54 <tsal> is that related to the consistent hashing 

07:55 <bingeldac> yes 

07:55 <bingeldac> http://wiki.basho.com/display/RIAK/Replication 
07:56 <tsal> thanks, think someone may have linked that when I asked a similar question 
yesterday, but I missed to bookmark it 

07:57 <tsal> what is the significance (if I may be so naieve) of the R value… it seems that setting 
n=3 and r=3 makes sense, but if you lose a node, and r=3 -- the read would fail… why 
not read from one of the other two ? 

07:58 <tsal> … or… is that somehow to ensure integrity ? 

07:59 <tsal> although, I recognise you can reduce the R value on reads - it doesn't 
*really* make sense (at least day #1 reading about Riak) to have a read fail because of this 
condition - or is it expected to make reads with R=1 ? 

07:59 <seancribbs> tsal: R is simply how many replicas will be waited for before replying 

08:00 <seancribbs> R=N is generally not a good idea because it will return "not found" 
responses when you have node failure 

08:00 <johnae> so, basically a low R-value means faster response? why not just go with 1 then? 

08:00 <seancribbs> not necessarily faster 

08:01 <seancribbs> depending on the churn, R=1 may give you stale data more 
than a higher R value 

08:01 <bingeldac> depends on the tradeoffs your app has 

08:01 <seancribbs> right 

08:01 <johnae> yeah ok, if an updated value hasnt replicated yet? 

08:02 <seancribbs> johnae: right, or that replica just came back after a partition 

08:02 <johnae> ok good to know, I think we might actually have use for R=1 since some of 
our data will just be put in ONCE and never updated 

08:02 <seancribbs> for most applications, N/2+1 (aka "quorum") is a good value for R 

08:03 <seancribbs> johnae: it's also good to know about "basic quorum" 

08:03 <seancribbs> which is a special case for "not found" responses 

08:03 <johnae> ok... 

08:03 <seancribbs> basically it returns not-found sooner, rather than waiting for the last 
replica to reply 

08:04 <seancribbs> i.e. if you get N/2+1 not-found responses, not-found will be 
returned to the client 

08:04 <seancribbs> however, you don't generally hit that unless R=1 

08:05 <seancribbs> or R =< N - (N/2+1) 

08:05 <seancribbs> so, for N=5, R=2 would invoke basic quorum for not-founds 

08:06 <seancribbs> etc 

08:06 <johnae> ok 

08:08 <tsal> thanks dudes ^ and great to see I inspired a bit of a chat about it 

08:09 <tsal> so for example if you have n=3 and r=3 – you effectively wait for all 
three to agree what is most recent, which might involve waiting for replication of 
that key to finish 

08:09 <seancribbs> tsal: no, it only waits for the _type_ of response 

08:09 <tsal> ah, ok 

08:09 <seancribbs> if all come back "ok, here's the object" 

08:10 <seancribbs> then R is satisfied 

08:10 <seancribbs> once it has all the responses, the object is resolved to the latest 
via the vector clock 

08:10 dizzyd joined 

08:10 <tsal> I just had a wtf moment with that last thing you just said, maybe 
I need more manual :-D 

08:11 <seancribbs> tsal: think of the vector-clock like a "family tree" for the object 

08:11 <seancribbs> you can compare them and see which version is the latest 

08:11 <tsal> ok… so … ah 

08:11 <tsal> ok, interesting 

08:12 <seancribbs> or you could think of them like a git commit log, if that helps 

08:12 <tsal> somehow :) 

08:13 <tsal> thanks again seancribbs 

08:14 <seancribbs> np. a lot of this is on the wiki, btw. 

08:15 <tsal> yeah, it's a little much to digest, when I have no clue about 
this sort of thing 

08:16 uberjar joined 

08:17 nanoyak joined 

08:18 <uberjar> congrats on the release! 

08:19 <seancribbs> tsal: most of the time you need not worry about it 

08:19 <seancribbs> but it's nice to understand 

08:20 <tsal> :) yeah, naturally - I tried to watch the talk from nosql last year - 

08:21 <tsal> lots of stuff in the first 5 mins I didn't understand… but the PAC -- "pick two" 
approach you guys 
took, which is a strength relates directly to this model of hashing and replication