Skip to content

Latest commit

 

History

History
29 lines (19 loc) · 2.24 KB

README.md

File metadata and controls

29 lines (19 loc) · 2.24 KB

ScORe by Taboola

ScORe is a java library for effective schema pruning when reading nested data structures. Back in 2018 we've identified that our spark jobs are reading more data than they need to ingest, not leveraging parquet columnar format that support reading partial schema of nested columns. The impact of such reduction was evident also in resource utilization and resulted in much low execution times.

This issue is documented in notorious SPARK-4502, which at that time did not seem close to resolution. As spark issue was finally (partially) resolved, we were already running with ScORe in production for several months, reducing up to 95% of input size for spark sql queries which involves highly nested data structures, with structs, arrays and maps interwoven together.

ScORe traverse the logical plan of the query, determines the schema that is required for running the query, and then we re-create the query using the generated schema. We use ScORe as best effort solution, so in case we fail to generate the schema for a certain query, or if the generated schema is incompatible, it will fallback to the full schema of the datasource.

If you have dozens or hundreds of different queries running over data with nested schema, ScORe will provide you with a tailor-made schema on read, rather than having you define it manually per each query.

For further reading: https://engineering.taboola.com/spark-sql-score/

Example usage

SparkContext sc = new SparkContext("local", "Example"); 
sparkSession = new SparkSession(sc);
Dataset<Row> dataset = sparkSession.read().parquet("/path/to/parquet");
Dataset<Row> query = dataset.select("my.nested.field").filter("condition = true");
query.createOrReplaceTempView("myView");

SchemaOnReadGenerator generator = SchemaOnReadGenerator.generateSchemaOnRead(query);
StructType schema = generator.getSchemaOnReadByAlias("myView"); // or generator.getSchemaOnRead("/path/to/parquet")

Dataset<Row> reducedSchemaDataset = sparkSession.read().schema(schema).parquet("/path/to/parquet");
Dataset<Row> reducedSchemaQuery = dataset.select("my.nested.field").filter("condition = true");
// perform action on reducedSchemaQuery