[SPARK-29372] Codegen grows beyond 64 KB for more columns in case of SupportsScanColumnarBatch

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Incomplete
• Affects Version/s: 2.3.2
• Component/s: Spark Core, SQL
• Labels: bulk-closed
• Assignee: Unassigned
• Reporter: Shubham Chaurasia
• Created: 07/Oct/19 15:50
• Updated: 25/May/21 01:51
• Resolved: 25/May/21 01:40

Description

In case of vectorized DSv2 readers i.e. if it implements SupportsScanColumnarBatch and number of columns is around(or greater than) 1000 then it throws

Caused by: org.codehaus.janino.InternalCompilerException: Code of method "processNext()V" of class "org.apache.spark.sql.catalyst.expressions.GeneratedClass$GeneratedIteratorForCodegenStage0" grows beyond 64 KB
	at org.codehaus.janino.CodeContext.makeSpace(CodeContext.java:990)
	at org.codehaus.janino.CodeContext.write(CodeContext.java:899)
	at org.codehaus.janino.CodeContext.writeBranch(CodeContext.java:1016)
	at org.codehaus.janino.UnitCompiler.writeBranch(UnitCompiler.java:11911)
	at org.codehaus.janino.UnitCompiler.compileBoolean2(UnitCompiler.java:3675)
	at org.codehaus.janino.UnitCompiler.access$5500(UnitCompiler.java:212)

I can see from logs that it tries to disable Whole-stage codegen but it's failing even after that on each retry.

19/10/07 20:49:35 WARN WholeStageCodegenExec: Whole-stage codegen disabled for plan (id=0):
 *(0) DataSourceV2Scan [column_0#3558, column_1#3559, column_2#3560, column_3#3561, column_4#3562, column_5#3563, column_6#3564, column_7#3565, column_8#3566, column_9#3567, column_10#3568, column_11#3569, column_12#3570, column_13#3571, column_14#3572, column_15#3573, column_16#3574, column_17#3575, column_18#3576, column_19#3577, column_20#3578, column_21#3579, column_22#3580, column_23#3581, ... 976 more fields], com.shubham.reader.MyDataSourceReader@5c7673b8

Repro code for a simple reader can be:

public class MyDataSourceReader implements DataSourceReader, SupportsScanColumnarBatch {

  private StructType schema;

  private int numCols = 10;
  private int numRows = 10;
  private int numReaders = 1;

  public MyDataSourceReader(Map<String, String> options) {
    initOptions(options);
    System.out.println("MyDataSourceReader.MyDataSourceReader: Instantiated...." + this);
  }

  private void initOptions(Map<String, String> options) {
    String numColumns = options.get("num_columns");
    if (numColumns != null) {
      numCols = Integer.parseInt(numColumns);
    }

    String numRowsOption = options.get("num_rows_per_reader");
    if (numRowsOption != null) {
      numRows = Integer.parseInt(numRowsOption);
    }

    String readersOption = options.get("num_readers");
    if (readersOption != null) {
      numReaders = Integer.parseInt(readersOption);
    }
  }

  @Override public StructType readSchema() {
    final String colPrefix = "column_";
    StructField[] fields = new StructField[numCols];
    for (int i = 0; i < numCols; i++) {
      fields[i] = new StructField(colPrefix + i, DataTypes.IntegerType, true, Metadata.empty());
    }

    schema = new StructType(fields);
    return schema;
  }

  @Override public List<DataReaderFactory<ColumnarBatch>> createBatchDataReaderFactories() {
    System.out.println("MyDataSourceReader.createDataReaderFactories: " + numReaders);
    return new ArrayList<>();
  }
}

If I pass num_columns 1000 or greater, the issue appears.

spark.read.format("com.shubham.MyDataSource").option("num_columns", "1000").option("num_rows_per_reader", 2).option("num_readers", 1).load.show

Any fixes/workarounds for this?

SPARK-16845 and SPARK-17092 are resolved but looks like they don't deal with the vectorized part.