This is all very hypothetical, but I've been bitten by this before.

Does row_loaded happen to be a binary or boolean value?  If so the secondary index generated by Cassandra will have at most 2 rows, and they'll be REALLY wide if you have a lot of entries.  Since Cassandra doesn't distribute columns over rows, those potentially very wide index rows, and their replicas, must live in SSTables in their entirety on the nodes that own them (and their replicas).

Even though you limit 1, I'm not sure what "behind the scenes" things Cassandra does.  I've received advice to avoid the built in secondary indexes in Cassandra for some of these reasons.  Also if row_loaded is meant to implement some kind of queuing behavior, it could be the wrong problem space for Cassandra as a result of all of the above.

On Sat, Mar 31, 2012 at 12:22 PM, aaron morton <> wrote:
Is there anything in the logs when you run the queries ? 

Try turning the logging up to DEBUG on the node that fails to return and see what happens. You will see it send messages to other nodes and do work itself. 

One thing to note, a query that uses secondary indexes runs on a node for each token range. So it will use more than CL number of nodes. 


Aaron Morton
Freelance Developer

On 30/03/2012, at 11:52 AM, Chris Hart wrote:


I have the following cluster:

<ip address>  MountainViewRAC1        Up     Normal  1.86 GB         20.00%  0                                           
<ip address>  MountainViewRAC1        Up     Normal  2.17 GB         33.33%  56713727820156410577229101238628035242      
<ip address>  MountainViewRAC1        Up     Normal  2.41 GB         33.33%  113427455640312821154458202477256070485     
<ip address>     Rackspace   RAC1        Up     Normal  3.9 GB          13.33%  136112946768375385385349842972707284580    

The following query runs quickly on all nodes except 1 MountainView node:

select * from Access_Log where row_loaded = 0 limit 1;

There is a secondary index on row_loaded.  The query usually doesn't complete (but sometimes does) on the bad node and returns very quickly on all other nodes.  I've upping the rpc timeout to a full minute (rpc_timeout_in_ms: 60000) in the yaml, but it still often doesn't complete in a minute.  It seems just as likely to complete and takes about the same amount of time whether the limit is 1, 100 or 1000.

Thanks for any help,