Hello, At CERN we are developing a Big Data system called NXCALS that uses Spark as Extraction API. We have implemented a custom datasource that was wrapping 2 existing ones (parquet and Hbase) in order to hide the implementation details (location of the parquet files, hbase tables, etc) and to provide an abstraction layer to our users. We have entered a stage where we execute some performance tests on our data and we have noticed that this approach did not provide the expected performance observed using pure Spark. In other words reading a parquet file with some simple predicates behaves 15 times slower if the same code is executed from within a custom datasource (that just uses Spark to read parquet). After some investigation we've learnt that Spark did not apply the same optimisations for both. We could see that in Spark 2.3.0 there was a new V2 version that abstracts from SparkSession and focuses on low level Row API. Could you give us some suggestions of how to correctly implement our datasource using the V2 API? Is this a correct way of doing it at all?
What we want to achieve is to join existing datasources with some level of additional abstraction on top. At the same time we want to profit from all catalyst & parquet optimisations that exist for the original ones. We also don't want to reimplement access to parquet files or Hbase at the low level (like Row) but just profit from the Dataset API. We could have achieved the same by providing an external library on top of Spark but the datasource approach looked like a more elegant solution. Only the performance is still far from the desired one. Any help or direction in that matter would be greatly appreciated as we have only started to build our Spark expertise yet. Best regards, Jakub Wozniak Software Engineer CERN -- Sent from: http://apache-spark-developers-list.1001551.n3.nabble.com/ --------------------------------------------------------------------- To unsubscribe e-mail: dev-unsubscr...@spark.apache.org
