Parallel Read Support
If you enable parallel reads via the enableParallelRead option, the singlestore-spark-connectorattempts to read results directly from aggregator or leaf nodes. This can drastically improve performance in some cases.
enableParallelRead Modes
The enableParallelRead option can have one of the following values:
disabled: Disables parallel reads and performs non-parallel reads.automaticLite: Performs parallel reads if at least one parallel read feature specified inparallelRead.Featuresis supported. Otherwise performs a non-parallel read. InautomaticLite mode, after push down of the outer sorting operation (for example, a nestedSELECTstatement where sorting is done in a top-levelSELECT) into SingleStore is done, a non-parallel read is used.Note
By default,
enableParallelReadis set toautomaticLite.automatic: Performs parallel reads if at least one parallel read feature specified inparallelRead.Featuresis supported. Otherwise performs a non-parallel read. Inautomaticmode, thesinglestore-spark-connectoris unable to push down an outer sorting operation into SingleStore. Final sorting is done at the Spark end of the operation.forced: Performs parallel reads if at least one parallel read feature specified inparallelRead.Featuresis supported. Otherwise it returns an error. Inforcedmode, thesinglestore-spark-connectoris unable to push down an outer sorting operation into SingleStore. Final sorting is done at the Spark end of the operation.
Parallel Read Features
The SingleStore Spark Connector supports the following parallel read features:
readFromAggregatorsreadFromAggregatorsMaterializedreadFromLeaves
Note
Managed Service only supports the readFromAggregators and readFromAggregatorsMaterialized features. To avoid changes in the event of migrating the workload to the cloud, use only these two features for on-premises clusters.
The connector uses the first feature specified in parallelRead.Features which meets all the requirements. The requirements for each feature are specified below. By default, the connector uses the readFromAggregators feature. You can repartition the result set for readFromAggregators and readFromAggregatorsMaterialized features. See Parallel Read Repartitioning for more information.
readFromAggregators
When this feature is used, SingleStore reads the query results in parallel. The number of partitions in the resulting DataFrame will be the same as the number of partitions in the SingleStore database. Each partition is read by a separate Spark task and all the tasks must start reading at the same time. To use this feature, the sum of spark.executor.cores/spark.task.cpus (identified as parallelism level) for all executors in the Spark cluster must be greater than the number of partitions in your database.
Use the parallelRead.tableCreationTimeoutMS option to specify a timeout for result table creation.
Requirements
To use this feature, the following requirements must be met:
SingleStore version 7.5+
SingleStore Spark Connector version 3.2+
Either the
databaseoption is set, or the database name is specified in theloadoptionSingleStore parallel read functionality supports the generated query
The parallelism level on the Spark cluster is enough to start reading from all the database partitions simultaneously
readFromAggregatorsMaterialized
Similar to readFromAggregators, this feature also reads the query results in parallel. The number of partitions in the resulting DataFrame will be the same as the number of partitions in the SingleStore database. The only difference is that readFromAggregatorsMaterialized uses the MATERIALIZED option to create the result table. When this feature is used, the reading tasks do not have to start at the same time. Hence, the parallelism level on the Spark cluster does not affect the reading tasks. Although, using the MATERIALIZED option may cause a query to fail if SingleStore does not have enough memory to materialize the result set.
Use the parallelRead.tableCreationTimeoutMaterializedMS option to specify a timeout for materialized result table creation.
Requirements
To use this feature, the following requirements must be met:
SingleStore version 7.5+
SingleStore Spark Connector version 3.2+
Either the
databaseoption is set, or the database name is specified in theloadoptionSingleStore parallel read functionality supports the generated query
readFromLeaves
When this feature is used, the singlestore-spark-conector skips the transaction layer and reads directly from partitions on the leaf nodes. Hence, each individual read task sees an independent version of the database's distributed state. If some queries (other than read operation) are run on the database, they may affect the current read operation.
This feature supports only those query-shapes that do not perform any operation on the aggregator and can be pushed down to the leaf nodes. To determine if a specific query is being pushed down to leaf nodes, query the DataFrame for the number of partitions using df.rdd.getNumPartitions. If this value is greater than 1, then we are reading in parallel from the leaf nodes.
Requirements
To use this feature, the following requirements must be met:
Either the
databaseoption is set, or the database name is specified in theloadoptionThe username and password provided to the
singlestore-spark-connectormust be uniform across all the nodes in the cluster, because parallel reads require consistent authentication and connectible leaf nodesThe hostnames and ports listed by
SHOW LEAVESmust be directly connectible from SparkThe generated query can be pushed down to the leaf nodes
Parallel Read Repartitioning
You can repartition the result using parallelRead.repartition option for the readFromAggregators and readFromAggregatorsMaterialized features to ensure that each task reads approximately the same amount of data. In queries with top level limit clauses, this option helps distribute the read task across multiple partitions so that all the rows do not belong to a single partition.