org.apache.spark.sql.columnar

Interface CachedBatchSerializer

All Superinterfaces:

java.io.Serializable

All Known Implementing Classes:

SimpleMetricsCachedBatchSerializer

public interface CachedBatchSerializer
extends scala.Serializable

Provides APIs that handle transformations of SQL data associated with the cache/persist APIs.

Method Summary

Modifier and Type	Method and Description
scala.Function2<Object,scala.collection.Iterator<CachedBatch>,scala.collection.Iterator<CachedBatch>>	buildFilter(scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Expression> predicates, scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Attribute> cachedAttributes) Builds a function that can be used to filter batches prior to being decompressed.
RDD<ColumnarBatch>	convertCachedBatchToColumnarBatch(RDD<CachedBatch> input, scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Attribute> cacheAttributes, scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Attribute> selectedAttributes, org.apache.spark.sql.internal.SQLConf conf) Convert the cached data into a ColumnarBatch.
RDD<org.apache.spark.sql.catalyst.InternalRow>	convertCachedBatchToInternalRow(RDD<CachedBatch> input, scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Attribute> cacheAttributes, scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Attribute> selectedAttributes, org.apache.spark.sql.internal.SQLConf conf) Convert the cached batch into InternalRows.
RDD<CachedBatch>	convertColumnarBatchToCachedBatch(RDD<ColumnarBatch> input, scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Attribute> schema, StorageLevel storageLevel, org.apache.spark.sql.internal.SQLConf conf) Convert an RDD[ColumnarBatch] into an RDD[CachedBatch] in preparation for caching the data.
RDD<CachedBatch>	convertInternalRowToCachedBatch(RDD<org.apache.spark.sql.catalyst.InternalRow> input, scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Attribute> schema, StorageLevel storageLevel, org.apache.spark.sql.internal.SQLConf conf) Convert an RDD[InternalRow] into an RDD[CachedBatch] in preparation for caching the data.
boolean	supportsColumnarInput(scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Attribute> schema) Can convertColumnarBatchToCachedBatch() be called instead of convertInternalRowToCachedBatch() for this given schema? True if it can and false if it cannot.
boolean	supportsColumnarOutput(StructType schema) Can convertCachedBatchToColumnarBatch() be called instead of convertCachedBatchToInternalRow() for this given schema? True if it can and false if it cannot.
scala.Option<scala.collection.Seq<String>>	vectorTypes(scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Attribute> attributes, org.apache.spark.sql.internal.SQLConf conf) The exact java types of the columns that are output in columnar processing mode.

Method Detail

buildFilter

scala.Function2<Object,scala.collection.Iterator<CachedBatch>,scala.collection.Iterator<CachedBatch>> buildFilter(scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Expression> predicates,
                                                                                                                  scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Attribute> cachedAttributes)

Builds a function that can be used to filter batches prior to being decompressed. In most cases extending SimpleMetricsCachedBatchSerializer will provide the filter logic necessary. You will need to provide metrics for this to work. SimpleMetricsCachedBatch provides the APIs to hold those metrics and explains the metrics used, really just min and max. Note that this is intended to skip batches that are not needed, and the actual filtering of individual rows is handled later.

Parameters:

predicates - the set of expressions to use for filtering.

cachedAttributes - the schema/attributes of the data that is cached. This can be helpful if you don't store it with the data.

Returns:

a function that takes the partition id and the iterator of batches in the partition. It returns an iterator of batches that should be decompressed.

convertCachedBatchToColumnarBatch

RDD<ColumnarBatch> convertCachedBatchToColumnarBatch(RDD<CachedBatch> input,
                                                     scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Attribute> cacheAttributes,
                                                     scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Attribute> selectedAttributes,
                                                     org.apache.spark.sql.internal.SQLConf conf)

Convert the cached data into a ColumnarBatch. This currently is only used if supportsColumnarOutput() returns true for the associated schema, but there are other checks that can force row based output. One of the main advantages of doing columnar output over row based output is that the code generation is more standard and can be combined with code generation for downstream operations.

Parameters:

input - the cached batches that should be converted.

cacheAttributes - the attributes of the data in the batch.

selectedAttributes - the fields that should be loaded from the data and the order they should appear in the output batch.

conf - the configuration for the job.

Returns:

an RDD of the input cached batches transformed into the ColumnarBatch format.

convertCachedBatchToInternalRow

RDD<org.apache.spark.sql.catalyst.InternalRow> convertCachedBatchToInternalRow(RDD<CachedBatch> input,
                                                                               scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Attribute> cacheAttributes,
                                                                               scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Attribute> selectedAttributes,
                                                                               org.apache.spark.sql.internal.SQLConf conf)

Convert the cached batch into InternalRows. If you want this to be performant, code generation is advised.

Parameters:

input - the cached batches that should be converted.

cacheAttributes - the attributes of the data in the batch.

selectedAttributes - the field that should be loaded from the data and the order they should appear in the output rows.

conf - the configuration for the job.

Returns:

RDD of the rows that were stored in the cached batches.

convertColumnarBatchToCachedBatch

RDD<CachedBatch> convertColumnarBatchToCachedBatch(RDD<ColumnarBatch> input,
                                                   scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Attribute> schema,
                                                   StorageLevel storageLevel,
                                                   org.apache.spark.sql.internal.SQLConf conf)

Convert an RDD[ColumnarBatch] into an RDD[CachedBatch] in preparation for caching the data. This will only be called if supportsColumnarInput() returned true for the given schema and the plan up to this point would could produce columnar output without modifying it.

Parameters:

input - the input RDD to be converted.

schema - the schema of the data being stored.

storageLevel - where the data will be stored.

conf - the config for the query.

Returns:

The data converted into a format more suitable for caching.

convertInternalRowToCachedBatch

RDD<CachedBatch> convertInternalRowToCachedBatch(RDD<org.apache.spark.sql.catalyst.InternalRow> input,
                                                 scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Attribute> schema,
                                                 StorageLevel storageLevel,
                                                 org.apache.spark.sql.internal.SQLConf conf)

Convert an RDD[InternalRow] into an RDD[CachedBatch] in preparation for caching the data.

Parameters:

input - the input RDD to be converted.

schema - the schema of the data being stored.

storageLevel - where the data will be stored.

conf - the config for the query.

Returns:

The data converted into a format more suitable for caching.

supportsColumnarInput

boolean supportsColumnarInput(scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Attribute> schema)

Can convertColumnarBatchToCachedBatch() be called instead of convertInternalRowToCachedBatch() for this given schema? True if it can and false if it cannot. Columnar input is only supported if the plan could produce columnar output. Currently this is mostly supported by input formats like parquet and orc, but more operations are likely to be supported soon.

Parameters:

schema - the schema of the data being stored.

Returns:

True if columnar input can be supported, else false.

supportsColumnarOutput

boolean supportsColumnarOutput(StructType schema)

Can convertCachedBatchToColumnarBatch() be called instead of convertCachedBatchToInternalRow() for this given schema? True if it can and false if it cannot. Columnar output is typically preferred because it is more efficient. Note that convertCachedBatchToInternalRow() must always be supported as there are other checks that can force row based output.

Parameters:

schema - the schema of the data being checked.

Returns:

true if columnar output should be used for this schema, else false.

vectorTypes

scala.Option<scala.collection.Seq<String>> vectorTypes(scala.collection.Seq<org.apache.spark.sql.catalyst.expressions.Attribute> attributes,
                                                       org.apache.spark.sql.internal.SQLConf conf)

The exact java types of the columns that are output in columnar processing mode. This is a performance optimization for code generation and is optional.

Parameters:

attributes - the attributes to be output.

conf - the config for the query that will read the data.

Returns:

(undocumented)