replication - Concurrent writes to cassandra replicas - Is duplication possible? -

i have 2 machine cluster running cassandra 1.2.6. using keyspace has replication factor of 2. application demands me write both replicas in parallel , let cassandra replication , hoping cassandra not duplicate key/value on replica nodes.

for example:

• i have nodes node1 , node2. have keyspace has replication factor 2 configured on , column family push key/value pairs
• i use python client (pycassa) write cluster.
• a key, "keyx", hashes node1 , node2. (i find out key hashes servers through node tool command. (`$nodetool getendpoints keyspacename columnfamilyname keyhexstring`)
• i use client write (keyx, value) concurrently nodes node1 , node2. (in connection pool give specific server name)
• when writing, wait 1 write succeed (to master). (consistency level one)
• now, monitor through `$nodetool status` command amount of disk space cluster uses.
• i write around 100 keys each having 2mb values.

ideally should store around 400mb on disk overhead storing keys should marginal compared value sizes using.

observations:

• if not write nodes key hashes to, cassandra internally handles replication , data size around 400mb. (200mb on each node 100 keys 2mb value)
• if write nodes key hashes to, cassandra writing more expected amount of data disk. high 15% more. in our tests cassandra write ~460mb instead of 400mb.

my question is, behavior (15% overhead) expected? there configuration need tweak cassandra handles concurrent writes replicas.

thanks!

there 2 possible causes of 15% space can think of.

one because replica store 2 copies of column temporarily. if write column twice in cassandra @ different times, 2 copies may go separate memtables end in separate sstables on disk. @ point later, when sstables merged through compaction process, older value discarded, freeing space. in test run nodetool compact force compaction , see if space usage goes down.

another possible cause depends on how did test when didn't write both nodes. if did @ consistency level one, possible of writes dropped other replica, doesn't have keys yet. can sure running nodetool repair. space used in first observation may not keys.

you should aware writing replicas @ consistency level 1 not guarantee each replica holds copy. node receiving data not have store return success write, if replica. may overloaded (in workload, due not enough i/o write data out) , drop write, while succeeding in writing different replica. cause less space used in second observation, isn't happening in test since relatively small amount of data.

if need guarantee have 2 copies should write @ consistency level , write once.