Why not send the value itself instead of a placeholder? Now it takes
2x writes on a random node to do a single update (write placeholder,
write update) and N*x writes from the client (write value, write
placeholder to N-1). Where N is replication factor. Seems like extra
network and IO instead of less...
Of course, I still think this sounds like reimplementing Cassandra
internals in a Cassandra client (just guessing, I'm not a cassandra dev)
On Jul 3, 2011, at 5:20 PM, AJ <a...@dude.podzone.net> wrote:
Yang,
How would you deal with the problem when the 1st node responds
success but then crashes before completely forwarding any replicas?
Then, after switching to the next primary, a read would return stale
data.
Here's a quick-n-dirty way: Send the value to the primary replica
and send placeholder values to the other replicas. The placeholder
value is something like, "PENDING_UPDATE". The placeholder values
are sent with timestamps 1 less than the timestamp for the actual
value that went to the primary. Later, when the changes propagate,
the actual values will overwrite the placeholders. In event of a
crash before the placeholder gets overwritten, the next read value
will tell the client so. The client will report to the user that
the key/column is unavailable. The downside is you've overwritten
your data and maybe would like to know what the old data was! But,
maybe there's another way using other columns or with MVCC. The
client would want a success from the primary and the secondary
replicas to be certain of future read consistency in case the
primary goes down immediately as I said above. The ability to set
an "update_pending" flag on any column value would probably make
this work. But, I'll think more on this later.
aj
On 7/2/2011 10:55 AM, Yang wrote:
there is a JIRA completed in 0.7.x that "Prefers" a certain node in
snitch, so this does roughly what you want MOST of the time
but the problem is that it does not GUARANTEE that the same node
will always be read. I recently read into the HBase vs Cassandra
comparison thread that started after Facebook dropped Cassandra for
their messaging system, and understood some of the differences.
what you want is essentially what HBase does. the fundamental
difference there is really due to the gossip protocol: it's a
probablistic, or eventually consistent failure detector while
HBase/Google Bigtable use Zookeeper/Chubby to provide a strong
failure detector (a distributed lock). so in HBase, if a tablet
server goes down, it really goes down, it can not re-grab the
tablet from the new tablet server without going through a start up
protocol (notifying the master, which would notify the clients
etc), in other words it is guaranteed that one tablet is served by
only one tablet server at any given time. in comparison the above
JIRA only TRYIES to serve that key from one particular replica.
HBase can have that guarantee because the group membership is
maintained by the strong failure detector.
just for hacking curiosity, a strong failure detector + Cassandra
replicas is not impossible (actually seems not difficult), although
the performance is not clear. what would such a strong failure
detector bring to Cassandra besides this ONE-ONE strong
consistency ? that is an interesting question I think.
considering that HBase has been deployed on big clusters, it is
probably OK with the performance of the strong Zookeeper failure
detector. then a further question was: why did Dynamo originally
choose to use the probablistic failure detector? yes Dynamo's main
theme is "eventually consistent", so the Phi-detector is
**enough**, but if a strong detector buys us more with little cost,
wouldn't that be great?
On Fri, Jul 1, 2011 at 6:53 PM, AJ <a...@dude.podzone.net> wrote:
Is this possible?
All reads and writes for a given key will always go to the same
node from a client. It seems the only thing needed is to allow the
clients to compute which node is the closes replica for the given
key using the same algorithm C* uses. When the first replica
receives the write request, it will write to itself which should
complete before any of the other replicas and then return. The
loads should still stay balanced if using random partitioner. If
the first replica becomes unavailable (however that is defined),
then the clients can send to the next repilca in the
ring and switch from ONE write/reads to QUORUM write/reads
temporarily until the first replica becomes available again.
QUORUM is required since there could be some replicas that were not
updated after the first replica went down.
Will this work? The goal is to have strong consistency with a read/
write consistency level as low as possible while secondarily a
network performance boost.
