On Fri, Jan 14, 2011 at 7:12 PM, Sean Cribbs <[email protected]> wrote:

>
>
> Crap, the second after I hit "send" the lightbulb goes on!  Why is that?
>
> The quorum _was_ met (all vnodes just migrated to the one machine) but
> since some of them were fail-overs they didn't have the value yet (or the
> wrong value)?  In this case a read repair happened and subsequent gets
> worked.
>
>
> Your understanding is correct. However, when I say "quorum was met" I
> usually mean that "it had R successful replies". Minor semantic quibble.
>
> You are correct in saying that the wiki is misleading -- read repair
> happens when any successful reply reaches the FSM, even if "not found" was
> returned to the client, that is, if quorum was not met. We'll get that
> fixed.
>
> I'm still dark on the second question.
>
>
>> 2) Why doesn't r=1 work?
>>
>> In the IRC session, you claimed that r=1 would not have helped this
>> problem.  Just like the OP, this confused me.  You then went on to say it
>> was because of some optimization and then mentioned a "basic quorum."
>>
>> I took a few minutes to think about this and the only conclusion I came to
>> is that when r=1 you will treat the first response as the final response,
>> and in this case the notfound response will always come back first?  I'm not
>> sure if what I just said makes sense but I would have expected r=1 to work,
>> just like the OP.  I'll admit that I still haven't read all the wiki docs
>> yet (but I've read Read Repair 3 times now), so I'd be happy to hear RTFM.
>>
>
> A number of months ago, we ran into some issues with a cluster where "not
> found" responses were not returning in a reasonable amount of time,
> especially when R=1. That is, the requests took MUCH longer than a
> SUCCESSFUL read. We determined that this occurred because one of the
> partitions was too busy to reply, causing the request timeout to expire.  So
> we added a special case called "basic quorum" (n_val/2 + 1) that is invoked
> only when receiving a "not found" response from a replica.  The idea is that
> if a simple majority of the replica partitions report "not found", it's
> probably not there.  This way, you don't sit around waiting for the last
> lonely partition to reply when R=1 (and your successful reads are still fast
> because you only wait for one replica).  It's a tradeoff of availability:
> returning a potentially incorrect response vs. appearing unavailable (timing
> out). We chose the former.
>
> Hope that helps,
>


Reading your explanation made me realize it's because I'm mucking up
the semantics of "quorum."  It was previously my understanding that if R=1
then you only need a quorum of 1 vnode, where a quorum is simply defined as
a response.  Which would mean that the first reply (whether notfound or a
value) would be considered the cluster value.  However, as you subtly hinted
to above, quorum does not mean that, i.e. it's more than just a response.
 It's that R vnodes found _a_ value and agreed on it's contents.  Going back
to the case of R=1, N=3, and the value is missing on 2 of it's preferred
vnodes it means that the request will take as long as the longest vnode to
respond, even if 2 vnodes reply immediately with no value.  This sucks, so
as an optimization you decided that if a majority of the cluster thinks it's
X, well then it must be X!  I'm not sure I explained that well, but I'm sure
I understand it now :)


-Ryan
_______________________________________________
riak-users mailing list
[email protected]
http://lists.basho.com/mailman/listinfo/riak-users_lists.basho.com

Reply via email to