Apache Kylin™ Technical Blog

Fast Cubing on Spark in Apache Kylin


In order to make POC phase as simple as possible, a standalone spark cluster is the best choice.
So the environment setup is as below:

  1. hadoop sandbox (hortonworks hdp 2.2.0)

    (8 cores, 16G) * 1

  2. spark (1.4.1)

    master:(4 cores, 8G)

    worker:(4 cores, 8G) * 2

The hadoop conf should also be in the SPARK_HOME/conf

Fast Cubing Implementation on Spark

Spark as a computation framework has provided much richer operators than map-reduce. And some of them are quite suitable for the cubing algorithm, for instance aggregate.

As the Fast cubing algorithm, it contains several steps:

  1. build dictionary
  2. calculate region split for hbase
  3. build & output cuboid data

build dictionary

In order to build dictionary, distinct values of the column are needed, which new API DataFrame has already provided(since spark 1.3.0).

So after got the data from the hive through SparkSQL, it is quite natural to directly use the api to build dictionary.

calculate region split

In order to calculate the distribution of all cuboids, Kylin use a HyperLogLog implementation. And each record will have a counter, whose size is by default 16KB each. So if the counter shuffles across the cluster, that will be very expensive.

Spark has provided an operator aggregate to reduce shuffle size. It first does a map-reduce phase locally, and then another round of reduce to merge the data from each node.

build & output cuboid data

In order to build cube, Kylin requires a small batch which can fit into memory in the same time.

Previously in map-reduce implementation, Kylin leverage the life-cycle callback cleanup to gather all the input together as a batch. This cannot be directly applied in the map reduce operator in spark which we don’t have such life-cycle callback.

However spark has provided an operator glom which coalescing all elements within each partition into an array which is exactly Kylin want to build a small batch.

Once the batch data is ready, we can just apply the Fast Cubing algorithm.

Then spark api saveAsNewAPIHadoopFile allow us to write hfile to hdfs and bulk load to HBase.


We use the sample data Kylin provided to build cube, total record count is 10000.

Below are results(system environments are mentioned above)

Spark MR
Duration 5.5 min 10+ min


Since hdp 2.2+ requires Hive 0.14.0 while spark 1.3.0 only supports Hive 0.13.0. There are several compatibility problems in hive-site.xml we need to fix.

  1. some time-related settings

    There are several settings, whose default value in hive 0.14.0 cannot be parsed in 0.13.0. Such as hive.metastore.client.connect.retry.delay, its default value is 5s. And in hive 0.13.0, this value can only be in the format of Long value. So you have to manually change to from 5s to 5.

  2. hive.security.authorization.manager

    If you have enabled this configuration, its default value is org.apache.hadoop.hive.ql.security.authorization.plugin.sqlstd.SQLStdConfOnlyAuthorizerFactory which is newly introduced in hive 0.14.0, it means you have to use the another implementation, such as org.apache.hadoop.hive.ql.security.authorization.DefaultHiveAuthorizationProvider

  3. hive.execution.engine

    In hive 0.14.0, the default value of hive.execution.engine is tez, change it to mr in the Spark classpath, otherwise there will be NoClassDefFoundError.

NOTE: Spark 1.4.0 has a bug which will lead to ClassNotFoundException. And it has been fixed in Spark 1.4.1. So if you are planning to run on Spark 1.4.0, you may need to upgrade to 1.4.1

Last but not least, when you trying to run Spark application on YARN, make sure that you have hive-site.xml and hbase-site.xml in the HADDOP_CONF_DIR or YARN_CONF_DIR. Since by default HDP lays these conf in separate directories.

Next move

Clearly above is not a fair competition. The environment is not the same, test data size is too small, etc.

However it showed that it is practical to migrate from MR to Spark, while some useful operators in Spark will save us quite a few codes.

So the next move for us is to setup a cluster, do the benchmark on real data set for both MR and Spark.

We will update the benchmark once we finished, please stay tuned.