Apache Spark provides an important feature to cache intermediate data and provide significant performance improvement while running multiple queries on the same data. There are two ways to cache a Dataframe or a DataSet i.e. call
cache() is the same as calling
persist(MEMORY_AND_DISK). There are many articles online that can be read about caching and its benefits as well as when to cache but as a rule of thumb we should identify the Dataframe that is being reused in a Spark Application and cache it. Even if the system memory isn’t big enough, Spark will utilize disk space to spill over. To read more about what storage levels are available look at
StorageLevel.scala in Spark.
Starting in Spark 3.1.1 users can add their own cache serializer, if they desire, by setting the
spark.sql.cache.serializer configuration. This is a static configuration that is set once for the duration of a Spark application which means that you can only set the conf before starting a Spark application and cannot be changed for that application’s Spark session.
RAPIDS Accelerator for Apache Spark version 0.4+ has the
ParquetCachedBatchSerializer that is optimized to run on the GPU and uses Parquet to compress data before caching it. ParquetCachedBatchSerializer can be used independent of what the value of
spark.rapids.sql.enabled is. If it is set to true then the Parquet compression will run on the GPU if possible, and importantly
spark.sql.inMemoryColumnarStorage.enableVectorizedReader will not be honored as the GPU data is always read in as columnar. If
spark.rapids.sql.enabled is set to false the cached objects will still be compressed on the CPU as a part of the caching process.
To use this serializer please run Spark with the following conf.
spark-shell --conf spark.sql.cache.serializer=com.nvidia.spark.ParquetCachedBatchSerializer
All types are supported on the CPU. On the GPU, MapType and BinaryType are not supported. If an unsupported type is encountered the Rapids Accelerator for Apache Spark will fall back to using the CPU for caching.