I'm not proposing any changes to be done, but this looks like a very interesting topic for thought/hack/learning, so the following are only for thought exercises ....
HBase enforces a single write/read entry point, so you can achieve strong consistency by writing/reading only one node. but just writing to one node exposes you to loss of data if that node fails. so the region server HLog is replicated to 3 HDFS data nodes. the interesting thing here is that each replica sees a complete *prefix* of the HLog: it won't miss a record, if a record sync() to a data node fails, all the existing bytes in the block are replicated to a new data node. if we employ a similar "leader" node among the N replicas of cassandra (coordinator always waits for the reply from leader, but leader does not do further replication like in HBase or counters), the leader sees all writes onto the key range, but the other replicas could miss some writes, as a result, each of the non-leader replicas' write history has some "holes", so when the leader dies, and when we elect a new one, no one is going to have a complete history. so you'd have to do a repair amongst all the replicas to reconstruct the full history, which is slow. it seems possible that we could utilize the FIFO property of the InComingTCPConnection to simplify history reconstruction, just like Zookeeper. if the IncomingTcpConnection of a replica fails, that means that it may have missed some edits, then when it reconnects, we force it to talk to the active leader first, to catch up to date. when the leader dies, the next leader is elected to be the replica with the most recent history. by maintaining the property that each node has a complete prefix of history, we only need to catch up on the tail of history, and avoid doing a complete repair on the entire memtable+SStable. but one issue is that the history at the leader has to be kept really long ----- if a non-leader replica goes off for 2 days, the leader has to keep all the history for 2 days to feed them to the replica when it comes back online. but possibly this could be limited to some max length so that over that length, the woken replica simply does a complete bootstrap. thanks yang On Sun, Jul 3, 2011 at 8:25 PM, AJ <a...@dude.podzone.net> wrote: > We seem to be having a fundamental misunderstanding. Thanks for your > comments. aj > > On 7/3/2011 8:28 PM, William Oberman wrote: > > I'm using cassandra as a tool, like a black box with a certain contract to > the world. Without modifying the "core", C* will send the updates to all > replicas, so your plan would cause the extra write (for the placeholder). I > wasn't assuming a modification to how C* fundamentally works. > Sounds like you are hacking (or at least looking) at the source, so all the > power to you if/when you try these kind of changes. > will > On Sun, Jul 3, 2011 at 8:45 PM, AJ <a...@dude.podzone.net> wrote: >> >> On 7/3/2011 6:32 PM, William Oberman wrote: >> >> Was just going off of: " Send the value to the primary replica and send >> placeholder values to the other replicas". Sounded like you wanted to write >> the value to one, and write the placeholder to N-1 to me. >> >> Yes, that is what I was suggesting. The point of the placeholders is to >> handle the crash case that I talked about... "like" a WAL does. >> >> But, C* will propagate the value to N-1 eventually anyways, 'cause that's >> just what it does anyways :-) >> will >> >> On Sun, Jul 3, 2011 at 7:47 PM, AJ <a...@dude.podzone.net> wrote: >>> >>> On 7/3/2011 3:49 PM, Will Oberman wrote: >>> >>> 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... >>> >>> To send the value to each node is 1.) unnecessary, 2.) will only cause a >>> large burst of network traffic. Think about if it's a large data value, >>> such as a document. Just let C* do it's thing. The extra messages are tiny >>> and doesn't significantly increase latency since they are all sent >>> asynchronously. >>> >>> >>> Of course, I still think this sounds like reimplementing Cassandra >>> internals in a Cassandra client (just guessing, I'm not a cassandra dev) >>> >>> I don't see how. Maybe you should take a peek at the source. >>> >>> >>> 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. >>> >>> >>> >> >> >> >> -- >> Will Oberman >> Civic Science, Inc. >> 3030 Penn Avenue., First Floor >> Pittsburgh, PA 15201 >> (M) 412-480-7835 >> (E) ober...@civicscience.com >> > > > > -- > Will Oberman > Civic Science, Inc. > 3030 Penn Avenue., First Floor > Pittsburgh, PA 15201 > (M) 412-480-7835 > (E) ober...@civicscience.com > >
