I have the need to executing a map/reduce job to identity data stored in Cassandra before indexing this data to Elastic Search.

I have already used ColumnFamilyInputFormat (before start using CQL) to write hadoop jobs to do that, but I use to have a lot of troubles to perform tunning, as hadoop depends on how map tasks are split in order to successfull execute things in parallel, for IO/bound processes. 

First question is: Am I the only one having problems with that? Is anyone else using hadoop jobs that reads from Cassandra in production?

Second question is about the alternatives. I saw new version spark will have Cassandra support, but using CqlPagingInputFormat, from hadoop. I tried to use HIVE with Cassandra community, but it seems it only works with Cassandra Enterprise and doesn't do more than FB presto (http://prestodb.io/), which we have been using reading from Cassandra and so far it has been great for SQL-like queries. For custom map reduce jobs, however, it is not enough.

Does anyone know some other tool that performs MR on Cassandra? My impression is most tools were created to work on top of HDFS and reading from a nosql db is some kind of "workaround".

Third question is about how these tools work. Most of them writtes mapped data on a intermediate storage, then data is shuffled and sorted, then it is reduced. Even when using CqlPagingInputFormat, if you are using hadoop it will write files to HDFS after the mapping phase, shuffle and sort this data, and then reduce it. 

I wonder if a tool supporting Cassandra out of the box wouldn't be smarter. Is it faster to write all your data to a file and then sorting it, or batch inserting data and already indexing it, as it happens when you store data in a Cassandra CF? I didn't do the calculations to check the complexity of each one, what should consider no index in Cassandra would be really large, as the maximum index size will always depend on the maximum capacity of a single host, but my guess is that a map / reduce tool written specifically to Cassandra, from the beggining, could perform much better than a tool written to HDFS and adapted. I hear people saying Map/Reduce on Cassandra/HBase is usually 30% slower than M/R in HDFS. Does it really make sense? Should we expect a result like this?

Final question: Do you think writting a new M/R tool like described would be reinventing the wheel? Or it makes sense?

Thanks in advance. Any opinions about this subject will be very appreciated.

Best regards,
Marcelo Valle.