I got an answer on SO on this question, basically confirming that the CQLSSTableWrite cannot be used in Spark (at least in the form shown in the code snippet). DataStax filed a bug on that and might get solved on a future version.
"As you have observed, a single writer can only be used in serial (ConcurrentModificationExceptions will happen if you do not), and creating multiple writers in the JVM fails due to static schema construction within the Cassandra code that the SSTableWriter uses. I'm not aware of any workaround other than to spawn multiple JVMs, each writing to a separate directory. We have filed a Cassandra JIRA ticket to address this issue. https://issues.apache.org/jira/browse/CASSANDRA-7463"; - Tupshin Harper <http://stackoverflow.com/users/881195/tupshin-harper> S.O. question: http://stackoverflow.com/questions/24396902/using-cqlsstablewriter-concurrently/24455785#24455785 On Thu, Jun 26, 2014 at 11:03 AM, Rohit Rai <ro...@tuplejump.com> wrote: > Hi Gerard, > > What is the version of Spark, Hadoop, Cassandra and Calliope are you > using. We never built Calliope to Hadoop2 as we/or our clients don't use > Hadoop in their deployments or use it only as the Infra component for Spark > in which case H1/H2 doesn't make a difference for them. > > I know atleast of one case where the user had built Calliope against 2.0 > and was using it happily. If you need assistance with it we are here to > help. Feel free to reach out to me directly and we can work out a solution > for you. > > Regards, > Rohit > > > *Founder & CEO, **Tuplejump, Inc.* > ____________________________ > www.tuplejump.com > *The Data Engineering Platform* > > > On Thu, Jun 26, 2014 at 12:44 AM, Gerard Maas <gerard.m...@gmail.com> > wrote: > >> Thanks Nick. >> >> We used the CassandraOutputFormat through Calliope. The Calliope API >> makes the CassandraOutputFormat quite accessible and is cool to work with. >> It worked fine at prototype level, but we had Hadoop version conflicts >> when we put it in our Spark environment (Using our Spark assembly compiled >> with CDH4.4). The conflict seems to be at the Cassandra-all lib level, >> which is compiled against a different hadoop version (v1). >> >> We could not get round that issue. (Any pointers in that direction?) >> >> That's why I'm trying the direct CQLSSTableWriter way but it looks >> blocked as well. >> >> -kr, Gerard. >> >> >> >> >> On Wed, Jun 25, 2014 at 8:57 PM, Nick Pentreath <nick.pentre...@gmail.com >> > wrote: >> >>> can you not use a Cassandra OutputFormat? Seems they have >>> BulkOutputFormat. An example of using it with Hadoop is here: >>> http://shareitexploreit.blogspot.com/2012/03/bulkloadto-cassandra-with-hadoop.html >>> >>> Using it with Spark will be similar to the examples: >>> https://github.com/apache/spark/blob/master/examples/src/main/scala/org/apache/spark/examples/CassandraTest.scala >>> and >>> https://github.com/apache/spark/blob/master/examples/src/main/scala/org/apache/spark/examples/CassandraCQLTest.scala >>> >>> >>> On Wed, Jun 25, 2014 at 8:44 PM, Gerard Maas <gerard.m...@gmail.com> >>> wrote: >>> >>>> Hi, >>>> >>>> (My excuses for the cross-post from SO) >>>> >>>> I'm trying to create Cassandra SSTables from the results of a batch >>>> computation in Spark. Ideally, each partition should create the SSTable for >>>> the data it holds in order to parallelize the process as much as possible >>>> (and probably even stream it to the Cassandra ring as well) >>>> >>>> After the initial hurdles with the CQLSSTableWriter (like requiring >>>> the yaml file), I'm confronted now with this issue: >>>> >>>> >>>> >>>> >>>> java.lang.RuntimeException: Attempting to load already loaded column >>>> family customer.rawts >>>> at org.apache.cassandra.config.Schema.load(Schema.java:347) >>>> at org.apache.cassandra.config.Schema.load(Schema.java:112) >>>> at >>>> org.apache.cassandra.io.sstable.CQLSSTableWriter$Builder.forTable(CQLSSTableWriter.java:336) >>>> >>>> I'm creating a writer on each parallel partition like this: >>>> >>>> >>>> >>>> >>>> def store(rdd:RDD[Message]) = { >>>> rdd.foreachPartition( msgIterator => { >>>> val writer = CQLSSTableWriter.builder() >>>> .inDirectory("/tmp/cass") >>>> .forTable(schema) >>>> .using(insertSttmt).build() >>>> msgIterator.foreach(msg => {...}) >>>> })} >>>> >>>> And if I'm reading the exception correctly, I can only create one >>>> writer per table in one JVM. Digging a bit further in the code, it looks >>>> like the Schema.load(...) singleton enforces that limitation. >>>> >>>> I guess writings to the writer will not be thread-safe and even if >>>> they were the contention that multiple threads will create by having all >>>> parallel tasks trying to dump few GB of data to disk at the same time will >>>> defeat the purpose of using the SSTables for bulk upload anyway. >>>> >>>> So, are there ways to use the CQLSSTableWriter concurrently? >>>> >>>> If not, what is the next best option to load batch data at high >>>> throughput in Cassandra? >>>> >>>> Will the upcoming Spark-Cassandra integration help with this? (ie. >>>> should I just sit back, relax and the problem will solve itself?) >>>> >>>> Thanks, >>>> >>>> Gerard. >>>> >>> >>> >> >
