Thanks much, verifies what I thought it is doing when connecting to a random node. Will play with RackAware and DCQUORUM.
Wanted to see if anybody else has a case where they want to connect to local Data Center always. A case where the Nodes are geographically apart like A (NY) and D (London). Wanted to make sure client A1 will get the key[100] from A and not D. (minimum latency and network traversal between Data centers) Can somebody comment if I am on the right track To achieve this (keep the read/write in the same Data Center) use RackAware, Read/Write CL=DCQUORUM Might need RF=4 is this required and what effects RF has ? Will manual Tokens help keep the Data as per keys to a particular DC - maybe not practical ? On Tue, Oct 26, 2010 at 8:38 PM, Aaron Morton <aa...@thelastpickle.com> wrote: > Lets start with the simple case, all nodes have the same proximity to each > other. > The client connects to a random node, called the coordinator. When a request > is made the coordinator asynchronously sends it to all nodes that are a > replica for the requested key. It waits for the response, and in the best > case of a read is able to return the data to the client as soon as CL nodes > respond. > The client does not have any knowledge of where the data is located in the > cluster. Thats the job of the coordinator and it takes only 1 hop to get to > each replica. > The consistency check is between the full data returned from one node and a > digest returned from the others. If it fails then RR will kick in (under CL >> ONE, if CL ONE is probabilistic). > > Read about the RackAwareStrategy now NetworkTopologyStrategy discussed > here http://wiki.apache.org/cassandra/Operations?highlight=(network)|(strategy) and > the Snitch. > Also http://wiki.apache.org/cassandra/StorageConfiguration?highlight=(strategy)|(replication) These > features let you tell cassandra about your topology. > You will then want to use something like DCQUORUM or DCQUORUMSYNC (0.7+ > AFAIK) for your requests. > Hope that helps. > Aaron > On 27 Oct, 2010,at 11:58 AM, Joe Alex <joe.m.a...@gmail.com> wrote: > > Hi, > > I have Cassandra 0.6.6 running on 4 nodes with RF=2. > > Let say nodes A, B, C, D > > If I have clients A1, B1, C1, D1 connected to respective nodes what > happens when A1 requests A for a key "100" for which D is responsible > as per the Token. C has the second copy. > As per the logs A1 requests A which requests D and gets the data. D > also checks a consistency check in the background on C. > If I have RF=3 I assume D will do 2 consistency checks. > > If I need to get the data from A itself with minimum latency and > network traversal between Data centers is this what I need to do ? > > 1. maybe RF=4 or at least >= 3 > 2. Adjust Read Consistency (ONE, QUORUM, DCQUORUM...) > 3. Use RackAware strategy with DCQUORUM > 3. Adjust Write Consistency > > Is there a way to get/write the data from the closest node - example A > is in NY, D in London etc. > For above example key=100. A1 calls A and A gets the data all the way from > D. > Also when A1 writes key=100 data needs to be written in D and C by A > > Probably need RF=4 for this in combination with DCQUORUM or ANY/ONE ? > Want to know how everybody is approaching this cases ? > > > A > DEBUG [pool-1-thread-21] 2010-10-26 18:29:25,231 CassandraServer.java > (line 216) get_slice > DEBUG [pool-1-thread-21] 2010-10-26 18:29:25,231 StorageProxy.java > (line 386) weakread reading SliceFromReadCommand(table='Keyspace1', > key='100', column_parent='QueryPath(columnFamilyName='Standard2', > superColumnName='null', columnName='null')', start='', finish='', > reversed=true, count=1000000) from 1311748@/10.210.32.92 > DEBUG [RESPONSE-STAGE:2] 2010-10-26 18:29:25,234 > ResponseVerbHandler.java (line 52) Processing response on an async > result from 1311748@/10.210.32.92 > DEBUG [Timer-1] 2010-10-26 18:29:26,511 LoadDisseminator.java (line > 36) Disseminating load info ... > > D > DEBUG [ROW-READ-STAGE:5] 2010-10-26 18:29:19,415 SliceQueryFilter.java > (line 116) collecting middle:false:1...@1288128381467000 > DEBUG [ROW-READ-STAGE:5] 2010-10-26 18:29:19,415 SliceQueryFilter.java > (line 116) collecting last:false:3...@1288128369639000 > DEBUG [ROW-READ-STAGE:5] 2010-10-26 18:29:19,415 SliceQueryFilter.java > (line 116) collecting first:false:4...@1288128358062000 > DEBUG [ROW-READ-STAGE:5] 2010-10-26 18:29:19,415 ReadVerbHandler.java > (line 93) Read key 100; sending response to 1311748@/10.210.32.74 > DEBUG [CONSISTENCY-MANAGER:4] 2010-10-26 18:29:19,416 > ConsistencyChecker.java (line 73) Reading consistency digest for 100 > from 1081...@[/10.210.32.92, /10.210.32.93] > DEBUG [RESPONSE-STAGE:1] 2010-10-26 18:29:19,418 > ResponseVerbHandler.java (line 42) Processing response on a callback > from 1081388@/10.210.32.93 > > C > DEBUG [ROW-READ-STAGE:4] 2010-10-26 18:29:25,237 SliceQueryFilter.java > (line 116) collecting middle:false:1...@1288128381467000 > DEBUG [ROW-READ-STAGE:4] 2010-10-26 18:29:25,238 SliceQueryFilter.java > (line 116) collecting last:false:3...@1288128369639000 > DEBUG [ROW-READ-STAGE:4] 2010-10-26 18:29:25,238 SliceQueryFilter.java > (line 116) collecting first:false:4...@1288128358062000 > DEBUG [ROW-READ-STAGE:4] 2010-10-26 18:29:25,238 ReadVerbHandler.java > (line 75) digest is c1ba97c56693d7fe4cbb9ac0544034b3 > DEBUG [ROW-READ-STAGE:4] 2010-10-26 18:29:25,238 ReadVerbHandler.java > (line 93) Read key 100; sending response to 1081388@/10.210.32.92 >
