incubator-cassandra-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Jonathan Ellis <jbel...@gmail.com>
Subject thinking about dropping hinted handoff
Date Wed, 27 Jan 2010 16:34:08 GMT
While being able to write (with CL.ZERO or new-in-0.6 ANY) even if all
the real write targets are down is cool, but since your goal in real
life is to keep enough replicas alive that you can actually do reads,
I'm not sure how useful it is.  HH also has a measurable performance
problem in small clusters (that is, where cluster size is not much
larger than replication factor) since having a node go down means you
will increase the write load on the remaining nodes a non-negligible
amount to write the hints, which can be a nasty surprise if you
weren't planning for it.

As for HH's consistency-improving characteristics, remember that HH is
not reliable (it's possible for a node to be down for several seconds
before HH gets turned on for it; it's also possible that the node with
the hints itself goes down before the target node comes back up),
which is why we needed the anti-entropy repair code.  So I think you
could make the case that now that we have anti-entropy, read repair
will be sufficient to handle inconsistency on "hot" keys, with
anti-entropy to handle infrequently accessed ones.  (Remembering of
course that if you wanted strong consistency in the first place, you
need to be doing quorum reads and writes and HH doesn't really matter.
 So we are talking about how to reduce inconsistency, when the client
has explicitly told us they're okay with seeing a little.)

Finally, I note that Cliff Moon, the author of Dynomite (probably the
most advanced pure Dynamo clone), deliberately left HH out for I
believe substantially these reasons.  (CC'd in case he wants to chime
in. :)

-Jonathan

Mime
View raw message