@nkeywal Primary point: with 3x replication on 100 nodes, all replicas will be on the other side of the partition for some shards.
Ah, @nkeywal just wrote the blog I had wanted to write for several years on the CAP theorem in sharded data systems. http://blog.thislongrun.com/2015/04/cap-availability-high-availability-and_16.html …
-
-
-
@nkeywal And hence unavailable from the point of the client. -
@nkeywal CAP theorem still holds but this violates intuition about what is meant by "availability". -
@jaykreps exactly, also because high availability seems to lack an generally accepted definition, while CAP defines availability clealy -
@nkeywal Yeah I submit 100% of misconceptions about CAP come from conflating CAP-availability with empirical high availability ("5 nines") -
@nkeywal Empirical availability is well defined (did your client get an answer in the SLA) but only can be measured on an operational system -
@nkeywal Related: people ignore overwhelming evidence that most real failures are correlated and hence replication /=> emp. availability. -
@jaykreps Stonebraker could have said this :-)
End of conversation
New conversation -
-
-
@jaykreps@nkeywal Quoted in#hadoopweekly, greatThanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.